Upload
others
View
28
Download
0
Embed Size (px)
Citation preview
Research ArticleWeighed Nonlinear Hybrid Neural Networks inUnderground Rescue Mission
Hongxing Yao12 Mary Opokua Ansong13 and Jun Steed Huang4
1 Institute of System Engineering Faculty of Science Jiangsu University 301 Xuefu Zhenjiang 212013 China2 College of Finance and Economics Jiangsu University 301 Xuefu Zhenjiang 212013 China3Department of Computer Science School of Applied Science Kumasi Polytechnic PO Box 854 Kumasi Ghana4Computer Science and Technology School of Computer Science amp Telecommunication Jiangsu University 301 XuefuZhenjiang 212013 China
Correspondence should be addressed to Hongxing Yao hxyaoujseducn
Received 25 September 2013 Accepted 11 November 2013 Published 22 January 2014
Academic Editors O Castillo K W Chau D Chen and P Kokol
Copyright copy 2014 Hongxing Yao et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited
In our previous work a novel model called compact radial basis function (CRBF) in a routing topology control has been modelledThe computational burden of Zhang and Gaussian transfer functions was modified by removing the power parameters on themodelsThe results showedoutstanding performance over theZhang andGaussianmodelsThis study researched on several hybridsforms of themodel where cosine (cos) and sine (sin) nonlinear weights were imposed on the two transfer functions such that119884(out) =
logsig(119877) + [exp(minusabs(119877))] lowast (plusmn cos or plusmn sin(119877)) The purpose was to identify the best hybrid that optimized all of its parameterswith a minimum error The results of the nonlinear weighted hybrids were compared with a hybrid of Gaussian model Simulationrevealed that the negative nonlinear weights hybrids optimized all the parameters and it is substantially superior to the previousapproaches presented in the literature with minimized errors of 00098 00121 00135 and 00129 for the negative cosine (HSCR-BFminuscos) positive cosine (HSCR-BF
+cos) negative sine (HSCR-BFminussin) and positive sine (HSCR-BF
+sin) hybrids respectively whilesigmoid andGaussian radial basis functions (HSGR-BF
+cos) were 00117The proposed hybrid could serve as an alternative approachto underground rescue operation
1 Introduction
11 Background In our earlier work we demonstrated how arouting path was generated and how the compact radial basisfunction could be improved by reducing the computationalburden of Gaussian by removing the power parameter fromthemodelWe had discussed the robustness and fault tolerantnature of the compact radial basis function for an emergencyunderground rescue operation and had discussed the perfor-mance of the sigmoid basis function and the compact radialbasis function of which the latter optimised its parametersbetter than that of the former [1] In this paper we look at thehybrid formof this novel algorithm by introducing nonlinearweights of positive and negative cosine and sine
12 Sigmoid Basis Function (SBF) and Radial Basis Func-tion (RBF) Sigmoid basis function (SBF) and radial basis
function (RBF) are the most commonly used algorithmsin neural training The output of the network is a linearcombination of radial basis function of the inputs and neuralparameters Radial basis function networks have many usesincluding function approximation time series prediction [23] classification and system control The structure supportsthe academic school of connectionist and the idea wasfirst formulated in 1988 by Broomhead and Lowe [4] TheSBF a mathematical function having an ldquoSrdquo shape (sigmoidcurve) and is related to brain reasoning and the structurefavors the computational believers The sigmoid functionrefers to the special case of the logistic function Anotherexample is the Gompertz curve which is used in modelingsystems that saturate at large values of input for examplethe ogee curve used in spillway of dams A wide variety ofsigmoid functions have been used as activation functions ofneurons including the logistic and hyperbolic tangent weight
Hindawi Publishing CorporationISRN Artificial IntelligenceVolume 2014 Article ID 864020 13 pageshttpdxdoiorg1011552014864020
2 ISRN Artificial Intelligence
functions Sigmoid curves are also common in statistics suchas integrals and logistic distribution normal distribution andStudentrsquos probability density functions In our opinion SBFoffers nonlinear effects for large input value and RBF providenonlinear effect at small input value A nonlinear hybrid ofcosine and sine will result in more nonlinear blending acrossthe entire region
13 Wireless Sensor Networks (WSN) Wireless sensor net-works (WSN) gather and process data from the environmentand make possible applications in the areas of environmentmonitoring logistics support health care and emergencyresponse systems as well as military operations Transmittingdata wirelessly impacts significant benefits to those inves-tigating buildings allowing them to deploy sensors whichmonitor from a remote location Multihop transmissionin wireless sensor networks conforms to the undergroundtunnel structure and provides more scalability for communi-cation system construction in rescue situations A significantdiscovery in the field of complex networks has shown that alarge number of complex networks including the internet arescale-free and their connectivity distribution is described bythe power-law of the form 120588(119896) sim 119896
minus120601 to allow few nodesof very large degree to exist making it difficult for randomattack A scale-free wireless network topology was thereforeused
This paper proposes a nonlinearHybridNeural Networksusing radial and sigmoid transfer functions in undergroundcommunication based on particle swarm optimisation Analternative to this model is without hybrid either RBF orSBF The SBF is known to be fast while RBF is accurate [5]and therefore blending the two will provide both speed andaccuracy In addition the two models have been examinedin our previous work [1] To this end we model the incidentlocation as a pure random event and calculate the probabilitythat communication chain through particular rock layersto the ground is not broken and allows neural network tomemorize the complicated relationship such that when realaccident happens the neural network resident in the robot isused to predict the probability based on the rock layer it seesinstantly If the result is positive the robot waits to receive therescue signal otherwise it moves deeper to the next layer andrepeats the procedure
Section 2 explains the preliminaries to the study andgenerates the routing path that has the highest survival proba-bility for the neural training Section 3 discusses related workto the study Section 4 discusses the network optimizationmodel based on the nonlinear weight on the compact radialbasis function Section 5 highlights the method used andSection 6 shows the simulation results of the various hybridsand compare with the Gaussian model whiles Section 7 givesa summary of the findings and future work
2 Preliminaries
21 Sensor Deployment Some assumptions such as 20 ofsoftware failure 2 safe exits assumed to be available afteraccidents and additional errors committed after the accidents
and the failure rate of radio frequency identification (RFID)weremade in various sectionsThese assumptions are alreadyincluded in the model and they are there to ensure that thesystem remains reliable
Topological deployment of sensor nodes affects the per-formance of the routing protocol [6 7] The ratio of com-munication range to sensing range as well as the distancebetween sensor nodes can affect the network topology
Let Ω be the sensor sequence for the deployment of totalsensors 119879 = 119909119910119911 = 119871 119877 119862 such that
Ω =
For 119905 = 119905 + 1
node (119905 2) = (
10038171003817100381710038171003817minus (119877 + 1) lowast (1 + (minus1)
tog119869)10038171003817100381710038171003817
2+ 119895)
node (119905 3) = (
10038171003817100381710038171003817minus (119862 + 1) lowast (1 + (minus1)
tog119870)10038171003817100381710038171003817
2+ 119896)
for 119894 = 1 119871 119895 = 1 119877 119896 = 1 119862 and node (119905 1) = 119894
(1)
tog119869 = ceil(119905119862119877) and tog119870 = ceil(119905119862) check source anddestination node respectively
Ω(119894 119895 119896) = 1 1 1 1 2 1 119894th 119895th 119896th for [level 1row 1 column 1] [level 1 row 2 column 1] [ ] and [119894th level119895th row and 119896th column] respectively Therefore 119879 = 119879 times 119879
matrix in an underground mine with dimensions of 119871 = 3119877 = 2 and 119862 = 1 for depth (level) row (length) and width(column) respectively with ldquopmrdquo a sensor apart implies thata minimum of 6 sensors will have to be deployed
22 Communication The Through-The-Earth (TTE) Com-munication system transmits voice and data through solidearth rock and concrete and is suitable for challengingunderground environments such as mines tunnels and sub-waysThere were stationary sensor nodes monitoring carbonmonoxide temperature and so forth as well as mobilesensors (humans and vehicles) distributed uniformly Bothstationary and mobile sensor nodes were connected to eitherthe Access Point (AP) andor Access Point Heads (APHeads)based on transmission range requirements [6]TheAPHeadsserve as cluster leaders and are located in areas where therock is relatively soft or signal penetration is better to ensurethat nodes are able to transmit the information they receivefrom APs and sensor nodes The APs are connected to otherAPs orThrough-The-Earth (TTE) which is dropped througha drilled hole down 300 meters apart based on the rock typeThe depth and rock type determine the required numberof TTEs needed Next the data-mule is discharged to carryitems such as food water and equipments to the minersunderground and return with underground information torescue team
23 SignalTransmission Reach Major challenges of sen-sor networks include battery constraints energy efficiencynetwork lifetime harsh underground characteristics bettertransmission range and topology design among othersSeveral routing approaches for safety evacuation have been
ISRN Artificial Intelligence 3
Table 1 Common rocks found in typical mines in relation to hardness or softness
Nonlinear mapping Mica Coal Granite Feldspar Quartz MineralSoftness 070 080 083 086 0875 090Hardness 2 3 5 6 7 9Distance 750m 470m 390m 315m 278 78m
proposed [8ndash14]Thesewere developed depending on specificemergency situations and management requirements Trans-mitting data wirelessly impacts significant benefits to thoseinvestigating buildings allowing them to deploy sensors andmonitor froma remote location [15 16] To effectively gain theneeded results researchers have come out with a number oftechniques to address the problem of topology control (TC)These include localization of nodes and time error and pathloss transmission range and total load eachnode experiencesas well as energy conservation which is very crucial inoptimizing efficiency and minimizing cost in wireless sensornetworks [17ndash19] Minimizing transmission range of wirelesssensor networks is vital to the efficient routing of the networkbecause the amount of communication energy that eachsensor consumes is highly related to its transmission range[6]The nodes signal reach119873120575was defined as the integrationof the change of the minimum and maximum signal reachtaking into consideration the number of cases (120591) of the rockstructure120573 120573 is the rock hardness119873120575 is the signal reach for anode and 119873120575min 119873120575max are minimum and maximum signalreach respectively The node signal reach is calculated as
119873120575 = 119873120575min +1003817100381710038171003817119873120575min minus 119873120575max
1003817100381710038171003817lowastrandom (
1015840
1205731015840
rock 1 1)
119873120575min = min (119871min (RowCol))
119873120575max = max (119871max (RowCol)) (2)
andwhere120573 is a geometric figure in the range of 07 to 09 Fora connection to be made the absolute difference between 119894 119895should be less than the node signal reach-119873120575The connectionmatrix was given as
120593119896 (119894 119895) = 1 if 1003817100381710038171003817119894 minus 119895
1003817100381710038171003817 le 119873120575
0 otherwise(3)
The relationship between rock hardness and the signal reachis a complicated nonlinear function which is related to theskin depth of the rock with alternating currents concentratedin the outer region of a conductor (skin depth) by opposinginternal magnetic fields as follows
Skin depth = radic2
(120588 lowast 120596 lowast 120590) (4)
120588 is material conductivity 120596 is frequency and 120590 is magneticpermeability
The signal (119861-field) is attenuated by cube of distance (119889)
119861 = (119896)1198893
Signal reach (distance) is equal to 3lowast skin depth
Table 1 identifies 6 common rocks found inmines in rela-tion to hardness or softness of each rock
A routing path was modeled using a number of 119879times119879 sizematrices namely the connection matrix (120593119896) routing matrix(120593119903) explosion matrix (120593119909) failed matrix (120593119891) hope matrix(120593ℎ) optimized matrix (120593119900) and the exit matrix (120593119890) Thehardware survival rate vector (120593ℎ) the survival rate vector(120593V) of each miner and the final average survival rate vector(119877) were also generated A sensor node is named by its 3Dinteger (119909 119910 119911) coordinates where 1 le 119909 le 119877 1 le 119910 le 1198621 le 119911 le 119871 for 119879 = 119877 lowast 119862 lowast 119871 being total number ofnodes If the node (119886 119887 119888) is connected with node (119889 119890 119891)
then the element on ((119886 minus 1) lowast 119862 lowast 119871 + (119887 minus 1) lowast 119871 + 119888)throw and ((119889 minus 1) lowast 119862 lowast 119871 + (119890 minus 1) lowast 119871 + 119891)th columnis 1 otherwise 0 and routing are limited to total number ofmultiple points connections that can be made In arriving atthe final optimized vector for transmission each matrix wasgenerated 120591 times
The 120593119903 sub 120593119896 119872120588
le 120572 119872120588
is even 119872120588
representing themaximum point-to-multipoint connection and 119872
120588
is evenallowing bidirectional communication with 119894119895 checkingsource and destination nodes respectively
120593119903 =
1 if 1003817100381710038171003817119894 minus 1198951003817100381710038171003817 ge (
119872120588
2)
0 otherwise
for 119894 119895 = 1 119879
(5)
24 Hardware Software and Network Fault Tolerant Con-siderations Network security is a critical issue in wirelesssensor networks as it significantly affects the efficiency ofthe communication andmany keymanagement schemes hadbeen proposed to mitigate this constraint [18 20] In an eventof accidents (120595) occurring the routing pathwould be affectedby (1minus120595) where120595 is any randomvaluewithin120573 whichwouldcause explosion on 119877120593 matrix and result in 120593119909 such that theresulting matrix would be the failed matrix (120593119891)
120593119909 = (1 minus 120595) 120593119903
120593119891 (119894 119895) = 1 if 120593119909 (119894 119895) lt 120582119871
0 if 120593119909 (119894 119895) ge 120582119871
(6)
120593119891(119894 119895) = 0 if 120593119909(119894 119895) ge 120582119867 else 120593119891 = 120593119909(119894 119895)120582119867 for (120582119871or 120582119867) representing the lower and higher accident impactthresholds respectively
A new set of routing paths (120593ℎ) and exit matrices (120593119890) fortransmission was calculated as
120593ℎ = 120593119891 lowast 120593119903
120593119890 = 119873119890 minus 120595
(7)
4 ISRN Artificial Intelligence
The mathematical objective here is to find an optimizedrouting matrix (120593119900) that has the maximum survivability Thesurvivability is defined as the entropy of a number of parallelconnections between every node to all the sink(s) The exitmatrix (120593119890) described the success rate from each node to thesink(s) 120593119890 assumes that the number of exits (119873119890) is availablewith an errormargin (120595) Inmost practical applicationsmorethan one sink is used and sink node is either through the fiberorThrough-The-Earth (TTE) link It is important to note thatin real rescue situations the software (relational databasemanagement system (RDBMS)) and hardware includingradio frequency identification (RFID) may fail as a result ofthe effect from the explosion and attack The matrix (120593119904)
was used to describe software survival rate including bugs orattacks
120593119904 = 1 minus ((1
119879 + random) (ldquoGeometricrdquo fail 119879 119879)) (8)
To obtain the final survival vector (120593119877) it was assumed thateach miner will have an RFID a vector 120593119868 was used todescribe its failure rate including risks of running out ofbattery a vector 120593119867 for the hardware failure rate for119879 = totalnumber of nodes
120593119868 = 1 minus ((1
119879 + 119903) lowast (ldquoGeometricrdquo fail 1 119879)) (9)
119903 is a random number generated from the vector 119879 and 119903
0 rarr infin 119879 + random 119879 rarr infin (1(119879 + 119903)) 1119879 rarr 01 minus (1(119879 + 119903)) 1 minus 1119879 rarr 1 is the minimum thereforefor 119879 = 119873 nodes we have 1 minus 1119873 = (119873 minus 1)119873 rarr 1 for(119873 minus 1)119873 rarr node is dead and 1 rarr node is alive
120593119867 = min (1 120593119868 [119883Ψ]) for [119883Ψ]) =120593119890
119872119901
(10)
The survival rate of each miner was (120593V) All these assump-tions happen in real life and must be considered The finalsurvival rate vector (120593119877) was calculated
120593V = 120593119904 lowast 120593119890 (11)
120593119877 = 120593V lowast 120593119904 (12)
119877 = 120593119877 (13)
3 Related Work
31 Artificial Neural Networks (ANN) Having found theoptimum set of routing table that has the highest survivalprobability of communicating with and rescuing miners it isimportant to train the neurons such that the initial error willbe minimized and more importantly have a reliable system[21]The topology of a neural network can be recurrent (withfeedback contained in the network from the output back tothe input) the feedforward (where data flow from the input tothe output units)The data processing can extend over severallayers of units but no feedback connections are presentManyresearchers have come out with neural network predictivemodels in both sigmoid and radial basis functions [22 23]
with applications such as nonlinear transformation [23]extreme learning machine predicting accuracy in gene clas-sification [24] crisp distributed support vectors regression(CDSVR) model was monthly streamflow prediction wasproposed by Valdez et al [25] and other applications includefuzzy inference systems (FIS) which have been successfullyapplied in fields such as automatic control data classificationdecision analysis and expert systems [26] among othersArtificial neural networks (ANN) are learning algorithmusedto minimize the error between the neural network outputand desired output This is important where relationshipsexist between weights within the hidden and output layersand among weights of more hidden layers In addition otherparameters including mean iteration standard variationstandard deviation and convergent time (in sections) wereevaluated The architecture of the learning algorithm andthe activation functions were included in neural networksNeurons are trained to process store recognize and retrievepatterns or database entries to solve combinatorial optimiza-tion problems Assuming that the input layer has 4 neuronsoutput layer 3 neurons and the hidden layer 6 neurons wecan evolve other parameters in the feedforward network toevolve the weight So the particles would be in a group ofweights and there would be 4 lowast 6 + 6 lowast 3 = 42 weightsThis implies that the particle consists of 42 real numbersThe range of weight can be set to [minus100 100] or a fittingrange After encoding the particles the fitness function wasthen determined The goodness of the fit was diagnosedusing mean squared error (MSE) as against mean cubicerror (MCE) and the mean absolute error (MAE) The meancubic error will allow for fast convergence and that willgross over accuracy making the process unstable while themean absolute error is stable but converges slowly A midwaybetween the two is the MSE and is given as
MSE =1
119899119904
119899
sum
119895=1
119904
sum
119894=1
(119884119895119894
minus 119910119895119894
)2
(14)
where 119899 is number of samples 119904 is the number of neurons atoutput layer 119884
119894119895
is the ideal value of 119894th sample at 119895th outputand 119884
119894119895
is the actual value of 119894th sample at 119895th output
32 The Gaussian and Zhang Models The standard or directapproach to interpolation at locations 119909
1
119909119873
sub 119877119889
without the first term using Gaussian kernel is given as
120601 (119903) = 119890minus(120576119903)
2
(15)
Zhang has [27] tried tomodify theGaussianmodel as followsa function 120595 [0infin) rarr R such that 119896(119909 1199091015840) = 120593(119909 minus 119909
1015840
)where 119909 1199091015840 isin 120594 and 119909 sdot 119909
1015840 denotes the Euclidean norm with119896(119909 119909
1015840
) = exp(minus119909 minus 1199091015840
2
1205752
) as an example of the RBFkernels The global support for RBF radials or kernels hasresulted in dense Grammatrices that can affect large datasetsand therefore construct the following two equations119896119862V(119909 119909
1015840
) = 120601119862V(119909 minus 119909
1015840
)119896(119909 1199091015840
) and 120601119862V(119909 minus 119909
1015840
) =
[(1 minus 119909 minus 1199091015840
119862)(sdot)+
]V where119862 gt 0 V ge (119889+1)2 and (sdot)
+
isthe positive part The function 120601C(sdot) is a sparsifying operator
ISRN Artificial Intelligence 5
which thresholds all the entries satisfying (119909 minus 1199091015840
) ge 119862
to zeros in the Gram matrix The new kernel resulting fromthis construction preserves positive definiteness This meansthat given any pair of inputs 119909 and 119909
1015840 where 119909 = 1199091015840 the
shrinkage (the smaller 119862) is imposed on the function value119896(119909 119909
1015840
) the result is that the Gram matrices 119870 and 119870119862V can
be either very similar or quite different depending on thechoice of 119862 However Zhang also ended up with an extrapower parameter
4 The Proposed Hybrid
The previous work looked at the Gaussian model and par-alyzed the computational power parameter The result wascompact radial basis function (CRBF) This was used to runboth SBF and RBF and the results compared The scalabilityand processing efficiency were also analyzed
The novelty of this algorithm the weighed nonlinearhybrid was to find several hybrids and the best for rescueoperationsThe cosine and sine functionswere used to reducehigh level nonlinear and increase small level nonlinearBoth the previous and the current algorithm used the samepreliminary considerations but the results is slightly differentbecause data are random
The sigmoid basis function was given as
log sig (119877) 119877 = 119882 sdot 119875 + 119861
log sig (119882 sdot 119875 + 119861) =1
1 + 119890minus(119882sdot119875+119861)
(16)
Neuron function 119878 (sigmoid) is log sig 119882 is weight matrix119875 is input vector and 119861 is threshold We therefore proposeda compact radial basis function based on the Gaussian radialbasis function and Helensrsquo [28] definition expressed as
[exp (minus 119877)] 119877 = 119882 sdot 119875 + 119861
(exp (minusabs ((119882 sdot 119875 + 119861)))
120601 (out) = [exp (minus 119877)]
(17)
119882 is weight matrix 119875 is input vector and 119861 is threshold Thefocus was to improve on the radial basis function for themineapplication
An optimized vector 119877 was generated as the optimumset of transmission routing table that has the highest survivalprobability for data transmission (13)
119875119873 =
119873
sum
119894=1
119877119894
119878119894
+
2
sum
119894=1
119878119894
+
2
sum
119894=1
119878119894
119878119894
+ 1
for 1 le 119894 le 119878 1 le 119896 le 119877
(18)
119877 1198781
1198782
are number of neurons at input hidden and outputlayers respectively
119875119873 is the position of the 119899th particle
119882119894
= 119878119894
1198771198822
= 119878119894
119877 119882119898
= 119878119898
119877
119875119894
= [119877 (119894 minus 119895) + 119870] = 119882119894
(119894 119896)
(19)
There are two thresholds (1198781
rArr 1198611
)Hidden 119861
1
(119894 119895) = 119875(1198771198781
+ 1198781
1198782
) rArr S1
Output 1198612
(119894 119895) = 119875(1198771198781
+ 1198781
1198782
) rArr 1198782
From (16) (17) and (18) the nonlinear weight of [plusmn cos(119877)] or[plusmn sin(119877)] was imposed on the CRBF before being combinedwith the SBF as (HSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sinHSCR-BF
+ sin)The cosine weight was used to keep high levelnonlinear for small input value and reduce the nonlinear forlarge input values while sine weight was used to keep highlevel nonlinear for large input value and reduce the nonlinearfor small input values
HSCR-BFminus cos
119884 (out) = log sig (119877) + [exp (minus 119877)] sdotlowast
[minus cos (119877)]
HSCR-BF+ cos
119884 (out) = log sig (119877) + [exp (minus 119877)] sdotlowast
[+ cos (119877)]
HSCR-BFminus sin
119884 (out) = log sig (119877) + [exp (minus 119877)] sdotlowast
[minus sin (119877)]
HSCR-BF+ sin
119884 (out) = log sig (119877) + [exp (minus 119877)] sdotlowast
[+ sin (119877)]
(20)
Thenonlinear weight of [+ cos(119877)]was imposed on theGRBFbefore being combined with the SBF
HSGR-YBF+ cos
119884 (out) = log sig (119877) + [exp (minus1198772
)] sdotlowast
[+ cos (119877)]
(21)
5 Particle Swarm Optimization
Particle swarm optimization (PSO) an evolutionary algo-rithm is a population based stochastic optimization tech-nique The idea was conceived by an American researcherand social psychologist James Kennedy in the 1950s Thetheory is inspired by social behavior of bird flocking or fishschooling The method falls within the category of swarmintelligence methods for solving global optimization prob-lems Literature has shown that the PSO is an effectivealternative to established evolutionary algorithms (GA) It isalso established that PSO is easily applicable to real worldcomplex problems with discrete continuous and nonlineardesign parameters and retains the conceptual simplicity ofGA [29 30] Each particle within the swarm is given an initialrandom position and an initial speed of propagation Theposition of the particle represents a solution to the problemas described in a matrix 120591 where 119872 and 119873 represent thenumber of particles in the simulation and the number ofdimensions of the problem respectively [31 32] A randomposition representing a possible solution to the problem withan initial associated velocity representing a function of thedistance from the particlersquos current position to the previous
6 ISRN Artificial Intelligence
position of good fitness value was given A velocity matrix119881el with the same dimensions as matrix 120591
119909
described this
120591119909
= (
12059111
12059112
1205911119873
12059121
12059122
1205912119873
1205911198981
1205911198982
120591119872119873
)
119881el = (
V11 V12 V1119873
V21 V22 V2119873
V1198981 V1198982 V119872119873
)
(22)
The best feasible alternative is the genetic algorithmHoweverthis is generally slow as compared to the PSO and thereforecomputationally expensive hence the PSO
While moving in the search space particles commit tomemorize the position of the best solution they have foundAt each iteration of the algorithm each particle moves witha speed that is a weighed sum of three components the oldspeed a speed component that drives the particles towardsthe location in the search space where it previously foundthe best solution so far and a speed component that drivesthe particle towards the location in the search space wherethe neighbor particles found the best solution so far [7] Thepersonal best position can be represented by an119873times119873matrix120588best and the global best position is an 119873-dimensional vector119866best
120588best = (
12058811
12058812
1205881119873
12058821
12058822
1205882119873
1205881198981
1205881198982
120588119872119873
)
119866best = (119892best1
119892best12
119892best119873
)
(23)
All particles move towards the personal and the globalbest with 120591 120588best 119881el and 119892best containing all the requiredinformation by the particle swarm algorithmThese matricesare updated on each successive iterations
119881119898119899
= 119881119898119899
+ 1205741198881
1205781199031
(119901best119898119899
minus 119883119898119898
)
+ 1205741198882
1205781199032
(119892best119899
minus 119883119898119899
)
119883119898119899
= 119881119898119899
(24)
1205741198881
and 1205741198882
are constants set to 13 and 2 respectively and1205781199031
1205781199032
are random numbers
51 AdaptiveMutation according toThreshold Particle swarmoptimization has been effective in training neural networkssuch as a Parallel Particle Swarm optimization (PPSO)methodwith dynamic parameter adaptation used to optimizecomplex mathematical functions and improved evolutionarymethod in fuzzy logic [30 31 33] To prevent particlesfrom not converging or converging at local minimum an
Input nodes64748
64748
64748
Hidden nodesOutput nodes
S11R
S12R
SR
120573 ge 07
S1mR
120573 ge 07 le 09
120573 ge 07 le 09
Figure 1 The structure AMPSO for CRBF GRBF and SBF transferfunctions
1 2 3 4 5 6 7 8 9 100
02040608
11214
The impact of the accident
Link
avai
labi
lity
The curve of transmission link after accident
The higher the impact the lower the link availability
Region 3Links are completely dead
Region 1Links are not
Region 2Indicate probability of link being able to transmit data
affected
Figure 2 Impact of explosionaccident on transmission link
adaptivemutation according to thresholdwas introduced Analternative to this is the use of the nonadaptive mutation andthis could converge to the local minimum or fail to convergeat all The advantage of the nonadaptation is simple and fastbut may not provide the needed results
Particles positions were updated with new value onlywhen the new value is greater than the previous value 20 ofparticles of those obtaining lower values weremade tomutatefor faster convergence and the structure of adaptive mutationPSO (AMPSO) with threshold can be found in [1] The inputlayer takes the final survival vector (13) with a number ofhidden layers and an output layer The feedforward neuralnetwork was used The structure of Adaptive Mutation PSO(AMPSO) with threshold Neural Network was used Figure 1[32]
6 Results and Discussion
61 Generating the Routing Path Elements 0 1 and 2 in 120593119909
imply that the link(s) were not affected and elements 3 45 and 6 represent a probability for the links being able totransmit data while figures from 7 means the link is totallydead Region 1 that indicates that links are not affected region
ISRN Artificial Intelligence 7
Table 2 Optimum set of routing table redundancy with the highest survival probability (6 cases)
Vectors for each case Rock hardness cases1198771 05409 04281 08834 0585 08134 05163 071198772 04526 05841 04547 08997 08436 02696 081198773 02931 05899 05903 08309 06971 02898 0831198774 04851 06085 06913 04002 05026 0 0861198775 04917 05857 07805 06037 05182 06472 08751198776 02902 07253 0458 05762 01397 0 09
2 gives the probability of link available and the last regionindicates the link is completely down (Figure 2)
Thematrices120593119894 120593119896 120593119903 120593119909 120593ℎ for 120591 = 119899 with dimensionsof 119871 = 3 119877 = 2 and119862 = 1 for depth (level) row (length) andwidth (column) respectively were generated as follows
119878eq = (
(
1 1 1
1 2 1
2 2 1
2 1 1
3 1 1
3 2 1
)
)
119894120593 =
[[[[[[[
[
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
]]]]]]]
]
120593119896 = (
(
1 1 1 1 0 0
1 1 1 0 0 0
0 1 1 1 0 0
0 0 1 1 1 0
0 0 0 1 1 1
0 0 0 1 1 1
)
)
1 = signal points
120593119903 = (
(
1 1 1 0 0 0
1 1 1 0 0 0
0 1 1 1 0 0
0 0 1 1 1 0
0 0 0 1 1 1
0 0 0 1 1 1
)
)
1 = transmission path
120593119909 = (
(
0 1 2 2 0 0
1 5 1 2 0 0
0 8 1 5 9 1
1 1 5 1 1 17
1 1 4 11 0 2
4 4 19 5 3 0
)
)
explosion sizes
120593119891 = (
(
1 1 1 1 1 1
1 6 1 1 1 1
1 0 1 6 0 1
1 1 6 1 1 0
1 1 75 0 1 1
75 75 0 6 1 1
)
)
(25)
120593ℎ = (
(
1 1 1 0 0 0
1 6 1 0 0 0
0 0 1 6 0 0
0 0 6 1 1 0
0 0 0 0 1 1
0 0 0 6 1 1
)
)
120593119900 = (
(
8 8 8 0 0 0
8 48 8 0 0 0
0 0 8 48 0 0
0 0 48 8 8 0
0 0 0 0 8 8
0 0 0 48 8 8
)
)
120593119890 = (
(
144 144 144 0 0 0
144 864 144 0 0 0
0 0 144 864 0 0
0 0 864 144 144 0
0 0 0 0 144 144
0 0 0 864 144 144
)
)
1 or fractional figures are transmission path
(26)
Element ldquo0rdquo on 120593119891 depicts a connection lost while ldquo1rdquo meansthat the connection will never go down and represent theconnection to the fixed sink node(s) along the edge or theemergency connection to the mobile data mule(s)
120593s = (
(
8333 8889 9091 8333 8571 8333
8889 8333 8750 8571 8750 8333
8571 9167 8750 9091 8889 8333
8750 8571 8571 8571 8571 8333
8333 8333 8571 8333 8889 9000
9167 8333 9091 8333 8889 8571
)
)
(27)
The survival rate of RFID hardware and each minervectors was displayed as
120593119868 = 8571 8750 8571 8889 8571 8333
120593119867 = 6236 4976 1 6851 9286 6086
120593V = 5409 4281 8834 5850 8134 5163
(28)
Optimization was done numerically usingMatlab simulationtool to find the optimum set of routing tables throughparticle swarm search for rescue operation as discussed inthe preliminaries From (13) the final survival vector fordimensions of 119871 = 3 119877 = 2 and 119862 = 1 will be an average ofthe 6 vectors as in Table 2 Total cases 119877119899 = 120591 = 119879times119879 (which
8 ISRN Artificial Intelligence
Table 3 Scalability of model to survival probability range robot location and rock type
Location Mica Coal Granite Feldspar Quartz MineralCompact radial basis function with negative nonlinear weight (HSCR-BF
minus cos)(10 6 5) 08769ndash10000 09070ndash09927 08402ndash09943 08023ndash09778 07386ndash09376 07345ndash08740(10 5 4) 09024ndash09696 09032ndash09505 08836ndash09430 08633ndash09527 07944ndash09465 07263ndash08720(6 5 4) 08906ndash09877 08827ndash09845 08384ndash09675 08627ndash09820 07911ndash09142 06392ndash07840(3 1 10) 07302ndash08481 06588ndash07816 06341ndash07473 06503ndash07249 05930ndash07156 04445ndash05621
Gaussian radial basis function with positive nonlinear weight (HSGR-BF+ cos)
(10 6 5) 08732ndash09963 08526ndash09987 08492ndash09950 07867ndash09568 07404ndash09025 06606ndash08315(10 5 4) 08843ndash09593 08843ndash09853 08444ndash09753 07746ndash09828 07304ndash09595 06678ndash08817(6 5 4) 08785ndash09983 08902ndash09996 08200ndash09997 07927ndash09446 07065ndash08565 06572ndash08063(3 1 10) 07608ndash08552 06600ndash08325 06090ndash07371 05482ndash06249 05026ndash06056 04856ndash05453
10 20 30 40 50 60 70 80 90 1000
02
04
06
X 94Y 0009904
Number of iterations
Fitn
ess v
alue
PSO optimized error for hybrid SBF-CRBF with negative nonlinear weight
15 2 25 3 35 4 45 5 55 604
06
08
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(HSCRminuscos )
(a) HSCR-BFminus cos
50 100 150 200 2500
01
02
03
X 249Y 001151
PSO optimized error with nonlinear weight [minussine(R)]
Number of iterations
Fitn
ess v
alue
1 2 3 4 5 6
07
08
09
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(b) HSCR-BFminus sin
Figure 3 Hybrid of SBF and CRBF (HSCR-BF) with (a) HSCR-BFminus cos and (b) HSCR-BF
minus sin nonlinear weight
represent the total nodes deployed) for this scenario Rocksoftness or hardness takes the values of 120573 and each matrix isgenerated 120591 timecases and the final routing vector will be anaverage of the 119899 vectors 120593119877 = Average (1198771 1198776) [Table 2]which becomes the input for the PSO training
The final survival rate is 119877 = (05409 04281 08834
0585 08130 05163) from (13)
62 The Final Survival Vector The 119877 = (05409 04281
08834 0585 08130 05163) from (13) is the maximumsurvival probability for a total of 6 nodes deployed Itdescribes the success rate from each node to the sink(s) Inmost practical applications more than one sink is used andsink node is either through the fiber or TTE connection Thesize of the vector depends on the dimensions of the fieldThe elements 119877 = (05409 04281 08834 0585 08130
05163) represent the probability of 54 43 88 5981 and 52 success of each node transmitting data to andfrom its source or destination It assists decision makers as towhether to send data through one ormore nodes or send each
message twice The total nodes used for the simulation were300 with undergroundmine dimensions of 119871 = 10119877 = 6 and119862 = 5 for depth (level) row (length) and width (column)respectively with ldquopmrdquo a sensor apart pm = 100119872
120588
= 4 and119873119890 = 2 The PSO training used swarm size of 20 maximumposition was set to 100 max velocity = 1 and maximumnumber of iteration = 250 The thresholds 120582119871 and 120582119867 were3 and 6 respectively 120591 = 6 cases (each case representa rock hardness case and also represent the total neuronsused) thus each matrix and vectors were run 6 times beforeneural training The survival probability (bottom) indicatesthat the model survived between 90 and 100 where rockcases were relatively soft (ge07) The survival probabilitydeclines as the rock becomes harder and approaches 08 Atthe hardest rock of 09 the survival probability fell between72 and 84 for the entire hybrids In view of this the APheads had to be deployed at a location where the rock isrelatively soft for maximum signal strength Each computersimulation incorporated all the 6 different cases of rockhardnesssoftness to produce the matrices Detailed locationanalysis or the scalability of the model in relation to survival
ISRN Artificial Intelligence 9
Table 4 Hybrid of SBF and CRBF with negative nonlinear weight on CRBF (HSCR-BFminus cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 231020 343498 0382 91650 00985AVG 23102 34350 0038 9165 00098MinMax 29540 45314 0024 80000 00043
Table 5 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BFminus sin)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 307320 488381 0376 114300 0135AVG 30732 48838 0038 11430 0013MinMax 30740 56926 0007 61500 0021
probability range robot location and different rock types isrecorded in Table 3 Figures 4ndash6 represent different scenariosof SBF and CRBF hybrids with the [plusmn cos(119877)] and [plusmn sin(119877)]
nonlinear weights while Figure 7 represents the SBF andGaussian hybrids for [+ cos(119877)] nonlinear weight The tophalf of each figure indicates the optimized error or the finalerror after the neural training and the bottom half revealsmodel survival probability
Figure 3 discusses the SBF and CRBF hybrids (HSCR-BF) with nonlinear weight Training in negative nonlinearweights (HSCR-BF
minus cos HSCR-BFminus sin) responded well The
HSCR-BFminus cos had a steady and compact routing path which
was consistent through all rock layers with initial survivalprobability between 877 and 100 in soft layers decliningto 735ndash874 at harder rock layers The HSCR-BF
minus sinperformed quite well but was more dispersed (879 to 985at the soft rock and 665 to 818 at the hard rock)
In addition the various parameters mean iterationstandard variance standard deviation and the convergenttime of HSCR-BF
minus cos were optimised This is demonstratedby finding the relationship between the maximum and theminimum values of the parameters and comparing with theaverage figures and the closer the difference is to the averagethe more consistent the data are in the dataset (Table 4) thusHSCR-BF
minus cos provided the best results among the proposedhybrids The HSCR-BF
minus sin on the RBF yielded strength inmean iteration and standard variance with error of 0013(Table 5)
In Figure 4 training with nonlinear weight of positivecosine and sine of HSCR-BF (HSCR-BF
+ cos and HSCR-BF+ sin) had some similarities The two models streamed
well at the initial stages from 896 to 992 and 908to 990 at soft layers of the rock respectively Howeverat the last stages the probability of transmitting effectivelybecame marginal from 795 to 852 and 821 to 854respectively Much as this hybrid could perform well in areaswhere routing conditions aremuch better in less dense rescuesituations this could hamper rescue mission in both casesdue to battery drain or collision from traffic conjunctionThedata formean iteration and convergent time forHSCR-BF
+ cos(Table 6) and mean Iteration standard variance the final
error HSCR-BF+ sin (Table 7) were consistent with the nega-
tive sine having better results than the positive cosineThe survival probability of the positive nonlinear weight
Gaussian hybrids (HSGR-BF+ cos) in Figure 5 is more com-
pact both at the initial (ie 873ndash996) and later (661ndash832) stages declining to harder rock layers with an errorof 001173 (Table 8)
At the initial stage particles are sensitive to inputs asthey moved quickly in the search space towards the targetand while particles peaked closer to the target it became lesssensitive to the input and began to align (Figure 6) At thelater stage more RBF are used to keep the error at minimumfor accuracy It was also noticed that instead of acceler-ating higher and becoming more sensitive to the inputswhile particles were far from the target the hybrids withlowast
[minus sin(119877)] and lowast[ + sin(119877)] were less sensitive to the inputsas lowast[minus sin(119877)] descended from [119909 = minus1 119910 = 05785] to[119909 = 1 119910 = 04215] while lowast[+ sin(119877)] descended from [119909 =
minus3 119910 = 00404] to [119909 = minus1 119910 = minus004062] before becomingconscious of the inputs In addition the hybrids with thelowast
[+ cos(119877)] lagged slightly between [119909 = minus4 119910 = 0006014]
and [119909 = minus2 119910 = 006288] before becoming sensitive toinputs However they all lined up for accuracy in terms of theminimised error
The nonlinearity of the hybrids is presented (Figure 7)the negative cosinesine weight is used to reduce nonlinearfor small input values and this lies between minus1 lt 119909 lt +1 andthe positive cosinesine weight is used to keep high nonlinearfor large inputs for the remaining region
63 CPU Time Efficiency The relationship between varioushybrids with respect to the central processing time (CPU)was profiled for different runs (Table 9) and expressed in asixth-order polynomial given as 119884
119903
= 1205737
1198836
1
+ 1205736
1198835
1
+ 1205735
1198834
1
+
1205734
1198833
1
+ 1205733
1198832
1
+ 1205732
1198831
+ 1205730
where 1198831
is time (seconds) and 120573
is the coefficient of the polynomial (Figure 8) The proposedhybrid has better usage of CPU time with 119877
2
= 09160followed byHSCR-BF
minus sin andHSCR-BF+ cos with119877
2 of 07551and 07244 respectively (Table 10) Applying the proposedalgorithm into Gaussian the CPU usage had 119877
2 of 07345
10 ISRN Artificial Intelligence
0 50 100 150 200
002
004
006
008
X 249Y 001103
PSO optimized error for hybrid SBF-CRBF with positive nonlinear weight
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 607
08
09
1
11 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(HSCR+cos )
(a) HSCR-BF+ cos
0 50 100 150 200
005
01
015
02
X 249Y 001247
PSO optimized error with nonlinear weight [+sine(R)]
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 6
070809
111 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(b) HSCR-BF+ sin
Figure 4 Hybrid of SBF and CRBF (HSCR-BF) with (a) HSCR-BF+ cos and (b) HSCR-BF
+ sin nonlinear weight
Table 6 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BF+ cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 350780 522257 0311 115650 0121AVG 35078 52226 0031 11565 0012MaxMin 54840 63413 0009 93500 0006
and indeedmarginally outperformedHSCR-BF+ cos Detailed
work on SBF CRBF and GRBF with regard to scalabilitymemory usage and the central processing time has beencarried by the authors [1]
7 Conclusion
In summary we made the following contributions First weused the mix of SBF and CRBF to present several hybridswith different nonlinear weights of cosine and sine func-tions on compact radial basis function Next we showedthe performance of the proposed nonlinear weight hybridsHSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sin and HSCR-BF+ sin optimised all the parameters with minimised error
of 00098 0012 0013 and 0013 respectively compared to00117 for Gaussian HSCR-BF
+ cos The analyzed CPU usagewith corresponding 119877
2 values of 09160 075 072440 06731and 07345 for HSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sinHSCR-BF
+ sin and HSCR-BF+ cos respectively demonstrated
that the algorithm is scalable There exist some evacuationmodels that is Goh and Mandic [11] which offered a choicefor travelers and several schemes for the decisionmakersmaynot be applicable in emergency undergroundmine situationsThe proposed nonlinear hybrid algorithm with particle swarmoptimisation has better capability of approximation to underly-ing functions with a fast learning speed and high robustness and
is competitive and more computationally efficient to Gaussianwith the same nonlinear weight The algorithm is new andthat makes it difficult to identify limitations and we intend toinvestigate other hybrids and comparewith genetic algorithm(GA) in the future
Nomenclature
119873120575min 119873120575max Minimum and maximum signalreach
HSCR-BFHSGR-BF Hybrid of SBF with CRBFHybrid ofSBF with GRBF
HSCR-BFminus cos Hybrid of SBF and CRBF minuscos
nonlinear weightHSCR-BF
+ cos Hybrid of SBF and CRBF +cosnonlinear weight
HSCR-BFminus sin Hybrid of SBF and CRBF minussine
nonlinear weightHSCR-BF
+ sin Hybrid of SBF and CRBF +sinenonlinear weight
HSGR-BF+ cos Hybrid of SBF and GRBF +cos
nonlinear weight
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
ISRN Artificial Intelligence 11
Table 7 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BF+ sin)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 260360 407109 03457 99800 0129AVG 26036 40711 00346 9980 0013MinMax 27840 56040 0008 106500 0014
Table 8 Hybrid of SBF and GRBF with nonlinear weight on GRBF (HSGR-BF+ cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 219840 349010 0439 112100 01173AVG 21984 34901 0044 11210 00117MinMax 27000 46580 0008 80000 00165
0 50 100 1500
005
01
015
02
X 144Y 0009595
GRBF-PSO optimized error with nonlinear weight [+cos(R)]
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 604
06
08
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
Figure 5 Hybrid of SBF and GRBF with nonlinear weight on GRBF(HSCR-BF
+ cos)
0 2 4 6 8 10
002040608
1121416
Weighed position of particles
Min
imise
d er
ror
Optimised error in nonlinear weighed SBF-CRBF hybridsParticles
Less sensitive to inputsmuch closer to target
more accurate
Less sensitive to inputsmuch closer more accurate
Neurons are more sensitive to inputsfast and far from target
minus02minus2minus4minus6minus8minus10
[x = minus1 y = 05785]
[x = 1 y = 04215]
[minus1 004062]
[x = minus2 y = 006288]
[x = minus4 y = 0006014]
[minus3 00404]
Hybrid+cosHybrid+sin
HybridminuscosHybridminussin
Figure 6 Particles weighed position
Acknowledgments
The authors would like to appreciate the immense contribu-tion of the mining companies where the study was under-taken The authors are grateful to these individuals for
0 1 2 3 4 5
0
05
1
15
2
X 37Y 0989
Particles position
Opt
imise
d er
ror
Nonlinear weight hybrids
Keep large inputvalues more
nonlinear
Keep small inputvalues lessnonlinear
Large input values
Small input values
Using positive cosine and sine
Using negative cosine and sineminus05
minus5 minus4 minus3 minus2 minus1
HSCR-BF+cosHSGR-BF+cos
HSCR-BF+sin
HSCR-BFminuscos
HSCR-BFminuscos
HSCR-BFminussin
HSCR-BF+cos HSCR-BF+sin HSCR-BFminussinand HSGR-BF+cos
Figure 7 Nonlinear weighed curves for optimized error withoutGaussian negative cosine curve
HSGR-BF+cos
HSGR-BF+cos
HSCR-BF+sinHSCR-BFminuscosHSCR-BFminussin
00020000400006000080000
100000120000
0 2 4 6 8 10 12
CPU
tim
e
Time (s)
Nonlinear hybrids
Figure 8 The trend of hybrids CPU time curves using sixth-OrderPolynomial
their immeasurable contributions to this work Fred Atta-kuma Patrick Addai Isaac Owusu Kwankye Thomas KwawAnnan Willet Agongo Nathaniel Quansah Francis OwusuMensah Clement Owusu-Asamoah Joseph Adu-MensahShadrack Aidoo Martin Anokye F T Oduro Ernest Ekow
12 ISRN Artificial Intelligence
Table 9 CPU time usage efficiency for the hybrids
Runs HSCR-BFminus cos HSCR-BF
minus sin HSCR-BF+ cos HSCR-BF
+ sin HSGR-BF+ cos
Total 255889 254350 422344 643962 438133Average 25589 25435 42234 64396 43813
Table 10 Analysis of CPU time on the various hybrids
1205736
1205735
1205734
1205733
1205732
1205731
1205730
1198772
HSCR-BFminus cos 00159119909 minus02575119909 minus0759119909 33494119909 minus19901119909 40345119909 25322 09160
HSCR-BF+ cos 00903119909 minus23158119909 19932119909 minus58027119909 minus31337119909 32808119909 14747 07244
HSCR-BFminus sin minus0033119909 11354119909 minus15811119909 11368119909 minus4337119909 78008119909 minus20794 07551
HSCR-BF+ sin 02571119909 minus14263119909 17173119909 minus10069119909 2993119909 minus42374119909 27879 06731
HSGR-BF+ cos minus00437119909 10911119909 minus82923119909 66218119909 15934119909 minus53939119909 8191 07345
Abano E Omari-Siaw and Qiang Jia This work was sup-ported by the National Natural Science Foundation of China(no 71271103) and by the Six Talents Peak Foundation ofJiangsu Province
References
[1] M O Ansong H X Yao and J S Huang ldquoRadial and sigmoidbasis functions neural networks in wireless sensor routingtopology control in underground mine rescue operation basedon particle swarm optimizationrdquo International Journal of Dis-tributed Sensor Networks-Hindawi vol 2013 Article ID 37693114 pages 2013
[2] R Neruda and P Kudova ldquoLearning methods for radial basisfunction networksrdquo Future Generation Computer Systems vol21 no 7 pp 1131ndash1142 2005
[3] C T Cheng J Y Lin and Y G C Sun ldquoLong-term predictionof discharges in Manwan hydropower using adaptive-network-based fuzzy inference systems modelsrdquo in Advances in NaturalComputation vol 3498 of Lecture Notes in Computer Sciencepp 1152ndash1161 2005
[4] D Broomhead and D Lowe ldquoMultivariate functional interpo-lation and adaptive networksrdquo Complex System vol 2 pp 321ndash355 1988
[5] Y Chen C Chuah and Q Zhao ldquoNetwork configuration foroptimal utilization efficiency of wireless sensor networksrdquo AdHoc Networks vol 6 no 1 pp 92ndash107 2008
[6] P S Rajpal K S Shishodia and G S Sekhon ldquoAn artificialneural network for modeling reliability availability and main-tainability of a repairable systemrdquo Reliability Engineering andSystem Safety vol 91 no 7 pp 809ndash819 2006
[7] W Jang W M Healy and S J Mirosław ldquoWireless sensornetworks as part of a web-based building environmental mon-itoring systemrdquo Automation in Construction vol 17 no 6 pp729ndash736 2008
[8] M S Pan C H Tsai and Y C Tseng ldquoEmergengy guiding andmonitoring application in indoor 3D environment by WirelessSensor Networkrdquo Internation Journal Os Sensor Netwourk vol1 pp 2ndash10 2006
[9] S Zarifzadeh A Nayyeri and N Yazdani ldquoEfficient construc-tion of network topology to conserve energy in wireless ad hocnetworksrdquo Computer Communications vol 31 no 1 pp 160ndash173 2008
[10] R Riaz A Naureen A Akram A H Akbar K Kim and HFarooq Ahmed ldquoA unified security framework with three keymanagement schemes for wireless sensor networksrdquo ComputerCommunications vol 31 no 18 pp 4269ndash4280 2008
[11] S L Goh and D P Mandic ldquoAn augmented CRTRL forcomplex-valued recurrent neural networksrdquo Neural Networksvol 20 no 10 pp 1061ndash1066 2007
[12] L A B Munoz and J J V Ramosy ldquoSimilarity-based heteroge-neous neural networksrdquo Engineering Letters vol 14 no 2 pp103ndash116 2007
[13] R Taormina K Chau and R Sethi ldquoArtificial neural networksimulation of hourly groundwater levels in a coastal aquifer sys-tem of the Venice lagoonrdquo Engineering Applications of ArtificialIntelligence vol 25 no 8 pp 1670ndash1676 2012
[14] K Leblebicioglu and U Halici ldquoInfinite dimensional radialbasis function neural networks for nonlinear transformationson function spacesrdquo Nonlinear Analysis Theory Methods andApplications vol 30 no 3 pp 1649ndash1654 1997
[15] F Fernandez-Navarro C Hervas-Martınez J Sanchez-Monedero and P A Gutierrez ldquoMELM-GRBF a modifiedversion of the extreme learning machine for generalized radialbasis function neural networksrdquo Neurocomputing vol 74 no16 pp 2502ndash2510 2011
[16] C L Wu KW Chau and Y S Li ldquoPredicting monthly stream-flow using data-drivenmodels coupledwith data-preprocessingtechniquesrdquoWater Resources Research vol 45 no 8 Article IDW08432 2009
[17] C EACheng ldquoLong-termprediction of discharges inManwanReservoir using artificial neural network modelsrdquo in Advancesin Neural NetworksmdashISNN 2005 vol 3498 of Lecture Notes inComputer Science pp 1040ndash1045 2005
[18] H H Zhang M Genton and P Liu ldquoCompactly supportedradial basis function kernelsrdquo 2004
[19] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquo inProceedings of the 1995 IEEE International Conference on NeuralNetworks pp 1942ndash1945 December 1995
[20] J Eberhart R Eberhart and Y Shi ldquoParticle swarm optimiza-tion developments applications and resourcesrdquo IEEE vol 1 pp81ndash86 2001
[21] R Malhotra and A Negi ldquoReliability modeling using ParticleSwarm optimisationrdquo International Journal of System AssuranceEngineering and Management vol 4 no 3 pp 275ndash283 2013
[22] D Gies and Y Rahmat-Samii ldquoParticle swarm optimization forreconfigurable phase-differentiated array designrdquo Microwaveand Optical Technology Letters vol 38 no 3 pp 168ndash175 2003
ISRN Artificial Intelligence 13
[23] F Valdez P Melin and C Oscar ldquoAn improved evolution-ary method with fuzzy logic for combining Particle Swarmoptimization and genetic algorithmsrdquo Applied Soft ComputingJournal vol 11 no 2 pp 2625ndash2632 2011
[24] J Zhang and K Chau ldquoMultilayer ensemble pruning vianovel multi-sub-swarm particle swarm optimizationrdquo Journalof Universal Computer Science vol 15 no 4 pp 840ndash858 2009
[25] F Valdez P Melin and O Castillo ldquoParallel Particle Swarmoptimization with parameters adaptation using fuzzy logicrdquoin Proceedings of the 11th Mexican International Conference onArtificial Intelligence (MICAI rsquo12) vol 2 pp 374ndash385 San LuisPotosi Mexico October 2012
[26] J Peng and C P Y Pan ldquoParticle swarm optimization RBFfor gas emission predictionrdquo Journal of Safety Science andTechnology vol 7 pp 77ndash85 2011
[27] G Ren Z Huang Y Cheng X Zhao and Y Zhang ldquoAnintegrated model for evacuation routing and traffic signaloptimization with background demand uncertaintyrdquo Journal ofAdvanced Transportation vol 47 pp 4ndash27 2013
[28] X Feng Z Xiao and X Cui ldquoImproved RSSI algorithm forwireless sensor networks in 3Drdquo Journal of ComputationalInformation Systems vol 7 no 16 pp 5866ndash5873 2011
[29] F Valdez and P Melin ldquoNeural network optimization witha hybrid evolutionary method that combines particle swarmand genetic algorithms with fuzzy rulesrdquo in Proceedings ofthe Annual Meeting of the North American Fuzzy InformationProcessing Society (NAFIPS rsquo08) NewYork NYUSAMay 2008
[30] F Fernandez-Navarro C Hervas-Martınez P A Gutierrez JM Pena-Barragan and F Lopez-Granados ldquoParameter estima-tion of q-Gaussian Radial Basis Functions Neural Networkswith a Hybrid Algorithm for binary classificationrdquo Neurocom-puting vol 75 pp 123ndash134 2012
[31] H Soh S Lim T Zhang et al ldquoWeighted complex networkanalysis of travel routes on the Singapore public transportationsystemrdquo Physica A vol 389 no 24 pp 5852ndash5863 2010
[32] C K S Kumar R Sukumar and M Nageswari ldquoSensorslifetime enhancement techniques in wireless sensor networksmdasha critical reviewrdquo International Journal of Computer Science andInformation Technology amp Security vol 3 pp 159ndash164 2013
[33] S Li and F Qin ldquoA dynamic neural network approach forsolving nonlinear inequalities defined on a graph and itsapplication to distributed routing-free range-free localizationof WSNsrdquo Neurocomputing vol 117 pp 72ndash80 2013
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
2 ISRN Artificial Intelligence
functions Sigmoid curves are also common in statistics suchas integrals and logistic distribution normal distribution andStudentrsquos probability density functions In our opinion SBFoffers nonlinear effects for large input value and RBF providenonlinear effect at small input value A nonlinear hybrid ofcosine and sine will result in more nonlinear blending acrossthe entire region
13 Wireless Sensor Networks (WSN) Wireless sensor net-works (WSN) gather and process data from the environmentand make possible applications in the areas of environmentmonitoring logistics support health care and emergencyresponse systems as well as military operations Transmittingdata wirelessly impacts significant benefits to those inves-tigating buildings allowing them to deploy sensors whichmonitor from a remote location Multihop transmissionin wireless sensor networks conforms to the undergroundtunnel structure and provides more scalability for communi-cation system construction in rescue situations A significantdiscovery in the field of complex networks has shown that alarge number of complex networks including the internet arescale-free and their connectivity distribution is described bythe power-law of the form 120588(119896) sim 119896
minus120601 to allow few nodesof very large degree to exist making it difficult for randomattack A scale-free wireless network topology was thereforeused
This paper proposes a nonlinearHybridNeural Networksusing radial and sigmoid transfer functions in undergroundcommunication based on particle swarm optimisation Analternative to this model is without hybrid either RBF orSBF The SBF is known to be fast while RBF is accurate [5]and therefore blending the two will provide both speed andaccuracy In addition the two models have been examinedin our previous work [1] To this end we model the incidentlocation as a pure random event and calculate the probabilitythat communication chain through particular rock layersto the ground is not broken and allows neural network tomemorize the complicated relationship such that when realaccident happens the neural network resident in the robot isused to predict the probability based on the rock layer it seesinstantly If the result is positive the robot waits to receive therescue signal otherwise it moves deeper to the next layer andrepeats the procedure
Section 2 explains the preliminaries to the study andgenerates the routing path that has the highest survival proba-bility for the neural training Section 3 discusses related workto the study Section 4 discusses the network optimizationmodel based on the nonlinear weight on the compact radialbasis function Section 5 highlights the method used andSection 6 shows the simulation results of the various hybridsand compare with the Gaussian model whiles Section 7 givesa summary of the findings and future work
2 Preliminaries
21 Sensor Deployment Some assumptions such as 20 ofsoftware failure 2 safe exits assumed to be available afteraccidents and additional errors committed after the accidents
and the failure rate of radio frequency identification (RFID)weremade in various sectionsThese assumptions are alreadyincluded in the model and they are there to ensure that thesystem remains reliable
Topological deployment of sensor nodes affects the per-formance of the routing protocol [6 7] The ratio of com-munication range to sensing range as well as the distancebetween sensor nodes can affect the network topology
Let Ω be the sensor sequence for the deployment of totalsensors 119879 = 119909119910119911 = 119871 119877 119862 such that
Ω =
For 119905 = 119905 + 1
node (119905 2) = (
10038171003817100381710038171003817minus (119877 + 1) lowast (1 + (minus1)
tog119869)10038171003817100381710038171003817
2+ 119895)
node (119905 3) = (
10038171003817100381710038171003817minus (119862 + 1) lowast (1 + (minus1)
tog119870)10038171003817100381710038171003817
2+ 119896)
for 119894 = 1 119871 119895 = 1 119877 119896 = 1 119862 and node (119905 1) = 119894
(1)
tog119869 = ceil(119905119862119877) and tog119870 = ceil(119905119862) check source anddestination node respectively
Ω(119894 119895 119896) = 1 1 1 1 2 1 119894th 119895th 119896th for [level 1row 1 column 1] [level 1 row 2 column 1] [ ] and [119894th level119895th row and 119896th column] respectively Therefore 119879 = 119879 times 119879
matrix in an underground mine with dimensions of 119871 = 3119877 = 2 and 119862 = 1 for depth (level) row (length) and width(column) respectively with ldquopmrdquo a sensor apart implies thata minimum of 6 sensors will have to be deployed
22 Communication The Through-The-Earth (TTE) Com-munication system transmits voice and data through solidearth rock and concrete and is suitable for challengingunderground environments such as mines tunnels and sub-waysThere were stationary sensor nodes monitoring carbonmonoxide temperature and so forth as well as mobilesensors (humans and vehicles) distributed uniformly Bothstationary and mobile sensor nodes were connected to eitherthe Access Point (AP) andor Access Point Heads (APHeads)based on transmission range requirements [6]TheAPHeadsserve as cluster leaders and are located in areas where therock is relatively soft or signal penetration is better to ensurethat nodes are able to transmit the information they receivefrom APs and sensor nodes The APs are connected to otherAPs orThrough-The-Earth (TTE) which is dropped througha drilled hole down 300 meters apart based on the rock typeThe depth and rock type determine the required numberof TTEs needed Next the data-mule is discharged to carryitems such as food water and equipments to the minersunderground and return with underground information torescue team
23 SignalTransmission Reach Major challenges of sen-sor networks include battery constraints energy efficiencynetwork lifetime harsh underground characteristics bettertransmission range and topology design among othersSeveral routing approaches for safety evacuation have been
ISRN Artificial Intelligence 3
Table 1 Common rocks found in typical mines in relation to hardness or softness
Nonlinear mapping Mica Coal Granite Feldspar Quartz MineralSoftness 070 080 083 086 0875 090Hardness 2 3 5 6 7 9Distance 750m 470m 390m 315m 278 78m
proposed [8ndash14]Thesewere developed depending on specificemergency situations and management requirements Trans-mitting data wirelessly impacts significant benefits to thoseinvestigating buildings allowing them to deploy sensors andmonitor froma remote location [15 16] To effectively gain theneeded results researchers have come out with a number oftechniques to address the problem of topology control (TC)These include localization of nodes and time error and pathloss transmission range and total load eachnode experiencesas well as energy conservation which is very crucial inoptimizing efficiency and minimizing cost in wireless sensornetworks [17ndash19] Minimizing transmission range of wirelesssensor networks is vital to the efficient routing of the networkbecause the amount of communication energy that eachsensor consumes is highly related to its transmission range[6]The nodes signal reach119873120575was defined as the integrationof the change of the minimum and maximum signal reachtaking into consideration the number of cases (120591) of the rockstructure120573 120573 is the rock hardness119873120575 is the signal reach for anode and 119873120575min 119873120575max are minimum and maximum signalreach respectively The node signal reach is calculated as
119873120575 = 119873120575min +1003817100381710038171003817119873120575min minus 119873120575max
1003817100381710038171003817lowastrandom (
1015840
1205731015840
rock 1 1)
119873120575min = min (119871min (RowCol))
119873120575max = max (119871max (RowCol)) (2)
andwhere120573 is a geometric figure in the range of 07 to 09 Fora connection to be made the absolute difference between 119894 119895should be less than the node signal reach-119873120575The connectionmatrix was given as
120593119896 (119894 119895) = 1 if 1003817100381710038171003817119894 minus 119895
1003817100381710038171003817 le 119873120575
0 otherwise(3)
The relationship between rock hardness and the signal reachis a complicated nonlinear function which is related to theskin depth of the rock with alternating currents concentratedin the outer region of a conductor (skin depth) by opposinginternal magnetic fields as follows
Skin depth = radic2
(120588 lowast 120596 lowast 120590) (4)
120588 is material conductivity 120596 is frequency and 120590 is magneticpermeability
The signal (119861-field) is attenuated by cube of distance (119889)
119861 = (119896)1198893
Signal reach (distance) is equal to 3lowast skin depth
Table 1 identifies 6 common rocks found inmines in rela-tion to hardness or softness of each rock
A routing path was modeled using a number of 119879times119879 sizematrices namely the connection matrix (120593119896) routing matrix(120593119903) explosion matrix (120593119909) failed matrix (120593119891) hope matrix(120593ℎ) optimized matrix (120593119900) and the exit matrix (120593119890) Thehardware survival rate vector (120593ℎ) the survival rate vector(120593V) of each miner and the final average survival rate vector(119877) were also generated A sensor node is named by its 3Dinteger (119909 119910 119911) coordinates where 1 le 119909 le 119877 1 le 119910 le 1198621 le 119911 le 119871 for 119879 = 119877 lowast 119862 lowast 119871 being total number ofnodes If the node (119886 119887 119888) is connected with node (119889 119890 119891)
then the element on ((119886 minus 1) lowast 119862 lowast 119871 + (119887 minus 1) lowast 119871 + 119888)throw and ((119889 minus 1) lowast 119862 lowast 119871 + (119890 minus 1) lowast 119871 + 119891)th columnis 1 otherwise 0 and routing are limited to total number ofmultiple points connections that can be made In arriving atthe final optimized vector for transmission each matrix wasgenerated 120591 times
The 120593119903 sub 120593119896 119872120588
le 120572 119872120588
is even 119872120588
representing themaximum point-to-multipoint connection and 119872
120588
is evenallowing bidirectional communication with 119894119895 checkingsource and destination nodes respectively
120593119903 =
1 if 1003817100381710038171003817119894 minus 1198951003817100381710038171003817 ge (
119872120588
2)
0 otherwise
for 119894 119895 = 1 119879
(5)
24 Hardware Software and Network Fault Tolerant Con-siderations Network security is a critical issue in wirelesssensor networks as it significantly affects the efficiency ofthe communication andmany keymanagement schemes hadbeen proposed to mitigate this constraint [18 20] In an eventof accidents (120595) occurring the routing pathwould be affectedby (1minus120595) where120595 is any randomvaluewithin120573 whichwouldcause explosion on 119877120593 matrix and result in 120593119909 such that theresulting matrix would be the failed matrix (120593119891)
120593119909 = (1 minus 120595) 120593119903
120593119891 (119894 119895) = 1 if 120593119909 (119894 119895) lt 120582119871
0 if 120593119909 (119894 119895) ge 120582119871
(6)
120593119891(119894 119895) = 0 if 120593119909(119894 119895) ge 120582119867 else 120593119891 = 120593119909(119894 119895)120582119867 for (120582119871or 120582119867) representing the lower and higher accident impactthresholds respectively
A new set of routing paths (120593ℎ) and exit matrices (120593119890) fortransmission was calculated as
120593ℎ = 120593119891 lowast 120593119903
120593119890 = 119873119890 minus 120595
(7)
4 ISRN Artificial Intelligence
The mathematical objective here is to find an optimizedrouting matrix (120593119900) that has the maximum survivability Thesurvivability is defined as the entropy of a number of parallelconnections between every node to all the sink(s) The exitmatrix (120593119890) described the success rate from each node to thesink(s) 120593119890 assumes that the number of exits (119873119890) is availablewith an errormargin (120595) Inmost practical applicationsmorethan one sink is used and sink node is either through the fiberorThrough-The-Earth (TTE) link It is important to note thatin real rescue situations the software (relational databasemanagement system (RDBMS)) and hardware includingradio frequency identification (RFID) may fail as a result ofthe effect from the explosion and attack The matrix (120593119904)
was used to describe software survival rate including bugs orattacks
120593119904 = 1 minus ((1
119879 + random) (ldquoGeometricrdquo fail 119879 119879)) (8)
To obtain the final survival vector (120593119877) it was assumed thateach miner will have an RFID a vector 120593119868 was used todescribe its failure rate including risks of running out ofbattery a vector 120593119867 for the hardware failure rate for119879 = totalnumber of nodes
120593119868 = 1 minus ((1
119879 + 119903) lowast (ldquoGeometricrdquo fail 1 119879)) (9)
119903 is a random number generated from the vector 119879 and 119903
0 rarr infin 119879 + random 119879 rarr infin (1(119879 + 119903)) 1119879 rarr 01 minus (1(119879 + 119903)) 1 minus 1119879 rarr 1 is the minimum thereforefor 119879 = 119873 nodes we have 1 minus 1119873 = (119873 minus 1)119873 rarr 1 for(119873 minus 1)119873 rarr node is dead and 1 rarr node is alive
120593119867 = min (1 120593119868 [119883Ψ]) for [119883Ψ]) =120593119890
119872119901
(10)
The survival rate of each miner was (120593V) All these assump-tions happen in real life and must be considered The finalsurvival rate vector (120593119877) was calculated
120593V = 120593119904 lowast 120593119890 (11)
120593119877 = 120593V lowast 120593119904 (12)
119877 = 120593119877 (13)
3 Related Work
31 Artificial Neural Networks (ANN) Having found theoptimum set of routing table that has the highest survivalprobability of communicating with and rescuing miners it isimportant to train the neurons such that the initial error willbe minimized and more importantly have a reliable system[21]The topology of a neural network can be recurrent (withfeedback contained in the network from the output back tothe input) the feedforward (where data flow from the input tothe output units)The data processing can extend over severallayers of units but no feedback connections are presentManyresearchers have come out with neural network predictivemodels in both sigmoid and radial basis functions [22 23]
with applications such as nonlinear transformation [23]extreme learning machine predicting accuracy in gene clas-sification [24] crisp distributed support vectors regression(CDSVR) model was monthly streamflow prediction wasproposed by Valdez et al [25] and other applications includefuzzy inference systems (FIS) which have been successfullyapplied in fields such as automatic control data classificationdecision analysis and expert systems [26] among othersArtificial neural networks (ANN) are learning algorithmusedto minimize the error between the neural network outputand desired output This is important where relationshipsexist between weights within the hidden and output layersand among weights of more hidden layers In addition otherparameters including mean iteration standard variationstandard deviation and convergent time (in sections) wereevaluated The architecture of the learning algorithm andthe activation functions were included in neural networksNeurons are trained to process store recognize and retrievepatterns or database entries to solve combinatorial optimiza-tion problems Assuming that the input layer has 4 neuronsoutput layer 3 neurons and the hidden layer 6 neurons wecan evolve other parameters in the feedforward network toevolve the weight So the particles would be in a group ofweights and there would be 4 lowast 6 + 6 lowast 3 = 42 weightsThis implies that the particle consists of 42 real numbersThe range of weight can be set to [minus100 100] or a fittingrange After encoding the particles the fitness function wasthen determined The goodness of the fit was diagnosedusing mean squared error (MSE) as against mean cubicerror (MCE) and the mean absolute error (MAE) The meancubic error will allow for fast convergence and that willgross over accuracy making the process unstable while themean absolute error is stable but converges slowly A midwaybetween the two is the MSE and is given as
MSE =1
119899119904
119899
sum
119895=1
119904
sum
119894=1
(119884119895119894
minus 119910119895119894
)2
(14)
where 119899 is number of samples 119904 is the number of neurons atoutput layer 119884
119894119895
is the ideal value of 119894th sample at 119895th outputand 119884
119894119895
is the actual value of 119894th sample at 119895th output
32 The Gaussian and Zhang Models The standard or directapproach to interpolation at locations 119909
1
119909119873
sub 119877119889
without the first term using Gaussian kernel is given as
120601 (119903) = 119890minus(120576119903)
2
(15)
Zhang has [27] tried tomodify theGaussianmodel as followsa function 120595 [0infin) rarr R such that 119896(119909 1199091015840) = 120593(119909 minus 119909
1015840
)where 119909 1199091015840 isin 120594 and 119909 sdot 119909
1015840 denotes the Euclidean norm with119896(119909 119909
1015840
) = exp(minus119909 minus 1199091015840
2
1205752
) as an example of the RBFkernels The global support for RBF radials or kernels hasresulted in dense Grammatrices that can affect large datasetsand therefore construct the following two equations119896119862V(119909 119909
1015840
) = 120601119862V(119909 minus 119909
1015840
)119896(119909 1199091015840
) and 120601119862V(119909 minus 119909
1015840
) =
[(1 minus 119909 minus 1199091015840
119862)(sdot)+
]V where119862 gt 0 V ge (119889+1)2 and (sdot)
+
isthe positive part The function 120601C(sdot) is a sparsifying operator
ISRN Artificial Intelligence 5
which thresholds all the entries satisfying (119909 minus 1199091015840
) ge 119862
to zeros in the Gram matrix The new kernel resulting fromthis construction preserves positive definiteness This meansthat given any pair of inputs 119909 and 119909
1015840 where 119909 = 1199091015840 the
shrinkage (the smaller 119862) is imposed on the function value119896(119909 119909
1015840
) the result is that the Gram matrices 119870 and 119870119862V can
be either very similar or quite different depending on thechoice of 119862 However Zhang also ended up with an extrapower parameter
4 The Proposed Hybrid
The previous work looked at the Gaussian model and par-alyzed the computational power parameter The result wascompact radial basis function (CRBF) This was used to runboth SBF and RBF and the results compared The scalabilityand processing efficiency were also analyzed
The novelty of this algorithm the weighed nonlinearhybrid was to find several hybrids and the best for rescueoperationsThe cosine and sine functionswere used to reducehigh level nonlinear and increase small level nonlinearBoth the previous and the current algorithm used the samepreliminary considerations but the results is slightly differentbecause data are random
The sigmoid basis function was given as
log sig (119877) 119877 = 119882 sdot 119875 + 119861
log sig (119882 sdot 119875 + 119861) =1
1 + 119890minus(119882sdot119875+119861)
(16)
Neuron function 119878 (sigmoid) is log sig 119882 is weight matrix119875 is input vector and 119861 is threshold We therefore proposeda compact radial basis function based on the Gaussian radialbasis function and Helensrsquo [28] definition expressed as
[exp (minus 119877)] 119877 = 119882 sdot 119875 + 119861
(exp (minusabs ((119882 sdot 119875 + 119861)))
120601 (out) = [exp (minus 119877)]
(17)
119882 is weight matrix 119875 is input vector and 119861 is threshold Thefocus was to improve on the radial basis function for themineapplication
An optimized vector 119877 was generated as the optimumset of transmission routing table that has the highest survivalprobability for data transmission (13)
119875119873 =
119873
sum
119894=1
119877119894
119878119894
+
2
sum
119894=1
119878119894
+
2
sum
119894=1
119878119894
119878119894
+ 1
for 1 le 119894 le 119878 1 le 119896 le 119877
(18)
119877 1198781
1198782
are number of neurons at input hidden and outputlayers respectively
119875119873 is the position of the 119899th particle
119882119894
= 119878119894
1198771198822
= 119878119894
119877 119882119898
= 119878119898
119877
119875119894
= [119877 (119894 minus 119895) + 119870] = 119882119894
(119894 119896)
(19)
There are two thresholds (1198781
rArr 1198611
)Hidden 119861
1
(119894 119895) = 119875(1198771198781
+ 1198781
1198782
) rArr S1
Output 1198612
(119894 119895) = 119875(1198771198781
+ 1198781
1198782
) rArr 1198782
From (16) (17) and (18) the nonlinear weight of [plusmn cos(119877)] or[plusmn sin(119877)] was imposed on the CRBF before being combinedwith the SBF as (HSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sinHSCR-BF
+ sin)The cosine weight was used to keep high levelnonlinear for small input value and reduce the nonlinear forlarge input values while sine weight was used to keep highlevel nonlinear for large input value and reduce the nonlinearfor small input values
HSCR-BFminus cos
119884 (out) = log sig (119877) + [exp (minus 119877)] sdotlowast
[minus cos (119877)]
HSCR-BF+ cos
119884 (out) = log sig (119877) + [exp (minus 119877)] sdotlowast
[+ cos (119877)]
HSCR-BFminus sin
119884 (out) = log sig (119877) + [exp (minus 119877)] sdotlowast
[minus sin (119877)]
HSCR-BF+ sin
119884 (out) = log sig (119877) + [exp (minus 119877)] sdotlowast
[+ sin (119877)]
(20)
Thenonlinear weight of [+ cos(119877)]was imposed on theGRBFbefore being combined with the SBF
HSGR-YBF+ cos
119884 (out) = log sig (119877) + [exp (minus1198772
)] sdotlowast
[+ cos (119877)]
(21)
5 Particle Swarm Optimization
Particle swarm optimization (PSO) an evolutionary algo-rithm is a population based stochastic optimization tech-nique The idea was conceived by an American researcherand social psychologist James Kennedy in the 1950s Thetheory is inspired by social behavior of bird flocking or fishschooling The method falls within the category of swarmintelligence methods for solving global optimization prob-lems Literature has shown that the PSO is an effectivealternative to established evolutionary algorithms (GA) It isalso established that PSO is easily applicable to real worldcomplex problems with discrete continuous and nonlineardesign parameters and retains the conceptual simplicity ofGA [29 30] Each particle within the swarm is given an initialrandom position and an initial speed of propagation Theposition of the particle represents a solution to the problemas described in a matrix 120591 where 119872 and 119873 represent thenumber of particles in the simulation and the number ofdimensions of the problem respectively [31 32] A randomposition representing a possible solution to the problem withan initial associated velocity representing a function of thedistance from the particlersquos current position to the previous
6 ISRN Artificial Intelligence
position of good fitness value was given A velocity matrix119881el with the same dimensions as matrix 120591
119909
described this
120591119909
= (
12059111
12059112
1205911119873
12059121
12059122
1205912119873
1205911198981
1205911198982
120591119872119873
)
119881el = (
V11 V12 V1119873
V21 V22 V2119873
V1198981 V1198982 V119872119873
)
(22)
The best feasible alternative is the genetic algorithmHoweverthis is generally slow as compared to the PSO and thereforecomputationally expensive hence the PSO
While moving in the search space particles commit tomemorize the position of the best solution they have foundAt each iteration of the algorithm each particle moves witha speed that is a weighed sum of three components the oldspeed a speed component that drives the particles towardsthe location in the search space where it previously foundthe best solution so far and a speed component that drivesthe particle towards the location in the search space wherethe neighbor particles found the best solution so far [7] Thepersonal best position can be represented by an119873times119873matrix120588best and the global best position is an 119873-dimensional vector119866best
120588best = (
12058811
12058812
1205881119873
12058821
12058822
1205882119873
1205881198981
1205881198982
120588119872119873
)
119866best = (119892best1
119892best12
119892best119873
)
(23)
All particles move towards the personal and the globalbest with 120591 120588best 119881el and 119892best containing all the requiredinformation by the particle swarm algorithmThese matricesare updated on each successive iterations
119881119898119899
= 119881119898119899
+ 1205741198881
1205781199031
(119901best119898119899
minus 119883119898119898
)
+ 1205741198882
1205781199032
(119892best119899
minus 119883119898119899
)
119883119898119899
= 119881119898119899
(24)
1205741198881
and 1205741198882
are constants set to 13 and 2 respectively and1205781199031
1205781199032
are random numbers
51 AdaptiveMutation according toThreshold Particle swarmoptimization has been effective in training neural networkssuch as a Parallel Particle Swarm optimization (PPSO)methodwith dynamic parameter adaptation used to optimizecomplex mathematical functions and improved evolutionarymethod in fuzzy logic [30 31 33] To prevent particlesfrom not converging or converging at local minimum an
Input nodes64748
64748
64748
Hidden nodesOutput nodes
S11R
S12R
SR
120573 ge 07
S1mR
120573 ge 07 le 09
120573 ge 07 le 09
Figure 1 The structure AMPSO for CRBF GRBF and SBF transferfunctions
1 2 3 4 5 6 7 8 9 100
02040608
11214
The impact of the accident
Link
avai
labi
lity
The curve of transmission link after accident
The higher the impact the lower the link availability
Region 3Links are completely dead
Region 1Links are not
Region 2Indicate probability of link being able to transmit data
affected
Figure 2 Impact of explosionaccident on transmission link
adaptivemutation according to thresholdwas introduced Analternative to this is the use of the nonadaptive mutation andthis could converge to the local minimum or fail to convergeat all The advantage of the nonadaptation is simple and fastbut may not provide the needed results
Particles positions were updated with new value onlywhen the new value is greater than the previous value 20 ofparticles of those obtaining lower values weremade tomutatefor faster convergence and the structure of adaptive mutationPSO (AMPSO) with threshold can be found in [1] The inputlayer takes the final survival vector (13) with a number ofhidden layers and an output layer The feedforward neuralnetwork was used The structure of Adaptive Mutation PSO(AMPSO) with threshold Neural Network was used Figure 1[32]
6 Results and Discussion
61 Generating the Routing Path Elements 0 1 and 2 in 120593119909
imply that the link(s) were not affected and elements 3 45 and 6 represent a probability for the links being able totransmit data while figures from 7 means the link is totallydead Region 1 that indicates that links are not affected region
ISRN Artificial Intelligence 7
Table 2 Optimum set of routing table redundancy with the highest survival probability (6 cases)
Vectors for each case Rock hardness cases1198771 05409 04281 08834 0585 08134 05163 071198772 04526 05841 04547 08997 08436 02696 081198773 02931 05899 05903 08309 06971 02898 0831198774 04851 06085 06913 04002 05026 0 0861198775 04917 05857 07805 06037 05182 06472 08751198776 02902 07253 0458 05762 01397 0 09
2 gives the probability of link available and the last regionindicates the link is completely down (Figure 2)
Thematrices120593119894 120593119896 120593119903 120593119909 120593ℎ for 120591 = 119899 with dimensionsof 119871 = 3 119877 = 2 and119862 = 1 for depth (level) row (length) andwidth (column) respectively were generated as follows
119878eq = (
(
1 1 1
1 2 1
2 2 1
2 1 1
3 1 1
3 2 1
)
)
119894120593 =
[[[[[[[
[
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
]]]]]]]
]
120593119896 = (
(
1 1 1 1 0 0
1 1 1 0 0 0
0 1 1 1 0 0
0 0 1 1 1 0
0 0 0 1 1 1
0 0 0 1 1 1
)
)
1 = signal points
120593119903 = (
(
1 1 1 0 0 0
1 1 1 0 0 0
0 1 1 1 0 0
0 0 1 1 1 0
0 0 0 1 1 1
0 0 0 1 1 1
)
)
1 = transmission path
120593119909 = (
(
0 1 2 2 0 0
1 5 1 2 0 0
0 8 1 5 9 1
1 1 5 1 1 17
1 1 4 11 0 2
4 4 19 5 3 0
)
)
explosion sizes
120593119891 = (
(
1 1 1 1 1 1
1 6 1 1 1 1
1 0 1 6 0 1
1 1 6 1 1 0
1 1 75 0 1 1
75 75 0 6 1 1
)
)
(25)
120593ℎ = (
(
1 1 1 0 0 0
1 6 1 0 0 0
0 0 1 6 0 0
0 0 6 1 1 0
0 0 0 0 1 1
0 0 0 6 1 1
)
)
120593119900 = (
(
8 8 8 0 0 0
8 48 8 0 0 0
0 0 8 48 0 0
0 0 48 8 8 0
0 0 0 0 8 8
0 0 0 48 8 8
)
)
120593119890 = (
(
144 144 144 0 0 0
144 864 144 0 0 0
0 0 144 864 0 0
0 0 864 144 144 0
0 0 0 0 144 144
0 0 0 864 144 144
)
)
1 or fractional figures are transmission path
(26)
Element ldquo0rdquo on 120593119891 depicts a connection lost while ldquo1rdquo meansthat the connection will never go down and represent theconnection to the fixed sink node(s) along the edge or theemergency connection to the mobile data mule(s)
120593s = (
(
8333 8889 9091 8333 8571 8333
8889 8333 8750 8571 8750 8333
8571 9167 8750 9091 8889 8333
8750 8571 8571 8571 8571 8333
8333 8333 8571 8333 8889 9000
9167 8333 9091 8333 8889 8571
)
)
(27)
The survival rate of RFID hardware and each minervectors was displayed as
120593119868 = 8571 8750 8571 8889 8571 8333
120593119867 = 6236 4976 1 6851 9286 6086
120593V = 5409 4281 8834 5850 8134 5163
(28)
Optimization was done numerically usingMatlab simulationtool to find the optimum set of routing tables throughparticle swarm search for rescue operation as discussed inthe preliminaries From (13) the final survival vector fordimensions of 119871 = 3 119877 = 2 and 119862 = 1 will be an average ofthe 6 vectors as in Table 2 Total cases 119877119899 = 120591 = 119879times119879 (which
8 ISRN Artificial Intelligence
Table 3 Scalability of model to survival probability range robot location and rock type
Location Mica Coal Granite Feldspar Quartz MineralCompact radial basis function with negative nonlinear weight (HSCR-BF
minus cos)(10 6 5) 08769ndash10000 09070ndash09927 08402ndash09943 08023ndash09778 07386ndash09376 07345ndash08740(10 5 4) 09024ndash09696 09032ndash09505 08836ndash09430 08633ndash09527 07944ndash09465 07263ndash08720(6 5 4) 08906ndash09877 08827ndash09845 08384ndash09675 08627ndash09820 07911ndash09142 06392ndash07840(3 1 10) 07302ndash08481 06588ndash07816 06341ndash07473 06503ndash07249 05930ndash07156 04445ndash05621
Gaussian radial basis function with positive nonlinear weight (HSGR-BF+ cos)
(10 6 5) 08732ndash09963 08526ndash09987 08492ndash09950 07867ndash09568 07404ndash09025 06606ndash08315(10 5 4) 08843ndash09593 08843ndash09853 08444ndash09753 07746ndash09828 07304ndash09595 06678ndash08817(6 5 4) 08785ndash09983 08902ndash09996 08200ndash09997 07927ndash09446 07065ndash08565 06572ndash08063(3 1 10) 07608ndash08552 06600ndash08325 06090ndash07371 05482ndash06249 05026ndash06056 04856ndash05453
10 20 30 40 50 60 70 80 90 1000
02
04
06
X 94Y 0009904
Number of iterations
Fitn
ess v
alue
PSO optimized error for hybrid SBF-CRBF with negative nonlinear weight
15 2 25 3 35 4 45 5 55 604
06
08
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(HSCRminuscos )
(a) HSCR-BFminus cos
50 100 150 200 2500
01
02
03
X 249Y 001151
PSO optimized error with nonlinear weight [minussine(R)]
Number of iterations
Fitn
ess v
alue
1 2 3 4 5 6
07
08
09
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(b) HSCR-BFminus sin
Figure 3 Hybrid of SBF and CRBF (HSCR-BF) with (a) HSCR-BFminus cos and (b) HSCR-BF
minus sin nonlinear weight
represent the total nodes deployed) for this scenario Rocksoftness or hardness takes the values of 120573 and each matrix isgenerated 120591 timecases and the final routing vector will be anaverage of the 119899 vectors 120593119877 = Average (1198771 1198776) [Table 2]which becomes the input for the PSO training
The final survival rate is 119877 = (05409 04281 08834
0585 08130 05163) from (13)
62 The Final Survival Vector The 119877 = (05409 04281
08834 0585 08130 05163) from (13) is the maximumsurvival probability for a total of 6 nodes deployed Itdescribes the success rate from each node to the sink(s) Inmost practical applications more than one sink is used andsink node is either through the fiber or TTE connection Thesize of the vector depends on the dimensions of the fieldThe elements 119877 = (05409 04281 08834 0585 08130
05163) represent the probability of 54 43 88 5981 and 52 success of each node transmitting data to andfrom its source or destination It assists decision makers as towhether to send data through one ormore nodes or send each
message twice The total nodes used for the simulation were300 with undergroundmine dimensions of 119871 = 10119877 = 6 and119862 = 5 for depth (level) row (length) and width (column)respectively with ldquopmrdquo a sensor apart pm = 100119872
120588
= 4 and119873119890 = 2 The PSO training used swarm size of 20 maximumposition was set to 100 max velocity = 1 and maximumnumber of iteration = 250 The thresholds 120582119871 and 120582119867 were3 and 6 respectively 120591 = 6 cases (each case representa rock hardness case and also represent the total neuronsused) thus each matrix and vectors were run 6 times beforeneural training The survival probability (bottom) indicatesthat the model survived between 90 and 100 where rockcases were relatively soft (ge07) The survival probabilitydeclines as the rock becomes harder and approaches 08 Atthe hardest rock of 09 the survival probability fell between72 and 84 for the entire hybrids In view of this the APheads had to be deployed at a location where the rock isrelatively soft for maximum signal strength Each computersimulation incorporated all the 6 different cases of rockhardnesssoftness to produce the matrices Detailed locationanalysis or the scalability of the model in relation to survival
ISRN Artificial Intelligence 9
Table 4 Hybrid of SBF and CRBF with negative nonlinear weight on CRBF (HSCR-BFminus cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 231020 343498 0382 91650 00985AVG 23102 34350 0038 9165 00098MinMax 29540 45314 0024 80000 00043
Table 5 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BFminus sin)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 307320 488381 0376 114300 0135AVG 30732 48838 0038 11430 0013MinMax 30740 56926 0007 61500 0021
probability range robot location and different rock types isrecorded in Table 3 Figures 4ndash6 represent different scenariosof SBF and CRBF hybrids with the [plusmn cos(119877)] and [plusmn sin(119877)]
nonlinear weights while Figure 7 represents the SBF andGaussian hybrids for [+ cos(119877)] nonlinear weight The tophalf of each figure indicates the optimized error or the finalerror after the neural training and the bottom half revealsmodel survival probability
Figure 3 discusses the SBF and CRBF hybrids (HSCR-BF) with nonlinear weight Training in negative nonlinearweights (HSCR-BF
minus cos HSCR-BFminus sin) responded well The
HSCR-BFminus cos had a steady and compact routing path which
was consistent through all rock layers with initial survivalprobability between 877 and 100 in soft layers decliningto 735ndash874 at harder rock layers The HSCR-BF
minus sinperformed quite well but was more dispersed (879 to 985at the soft rock and 665 to 818 at the hard rock)
In addition the various parameters mean iterationstandard variance standard deviation and the convergenttime of HSCR-BF
minus cos were optimised This is demonstratedby finding the relationship between the maximum and theminimum values of the parameters and comparing with theaverage figures and the closer the difference is to the averagethe more consistent the data are in the dataset (Table 4) thusHSCR-BF
minus cos provided the best results among the proposedhybrids The HSCR-BF
minus sin on the RBF yielded strength inmean iteration and standard variance with error of 0013(Table 5)
In Figure 4 training with nonlinear weight of positivecosine and sine of HSCR-BF (HSCR-BF
+ cos and HSCR-BF+ sin) had some similarities The two models streamed
well at the initial stages from 896 to 992 and 908to 990 at soft layers of the rock respectively Howeverat the last stages the probability of transmitting effectivelybecame marginal from 795 to 852 and 821 to 854respectively Much as this hybrid could perform well in areaswhere routing conditions aremuch better in less dense rescuesituations this could hamper rescue mission in both casesdue to battery drain or collision from traffic conjunctionThedata formean iteration and convergent time forHSCR-BF
+ cos(Table 6) and mean Iteration standard variance the final
error HSCR-BF+ sin (Table 7) were consistent with the nega-
tive sine having better results than the positive cosineThe survival probability of the positive nonlinear weight
Gaussian hybrids (HSGR-BF+ cos) in Figure 5 is more com-
pact both at the initial (ie 873ndash996) and later (661ndash832) stages declining to harder rock layers with an errorof 001173 (Table 8)
At the initial stage particles are sensitive to inputs asthey moved quickly in the search space towards the targetand while particles peaked closer to the target it became lesssensitive to the input and began to align (Figure 6) At thelater stage more RBF are used to keep the error at minimumfor accuracy It was also noticed that instead of acceler-ating higher and becoming more sensitive to the inputswhile particles were far from the target the hybrids withlowast
[minus sin(119877)] and lowast[ + sin(119877)] were less sensitive to the inputsas lowast[minus sin(119877)] descended from [119909 = minus1 119910 = 05785] to[119909 = 1 119910 = 04215] while lowast[+ sin(119877)] descended from [119909 =
minus3 119910 = 00404] to [119909 = minus1 119910 = minus004062] before becomingconscious of the inputs In addition the hybrids with thelowast
[+ cos(119877)] lagged slightly between [119909 = minus4 119910 = 0006014]
and [119909 = minus2 119910 = 006288] before becoming sensitive toinputs However they all lined up for accuracy in terms of theminimised error
The nonlinearity of the hybrids is presented (Figure 7)the negative cosinesine weight is used to reduce nonlinearfor small input values and this lies between minus1 lt 119909 lt +1 andthe positive cosinesine weight is used to keep high nonlinearfor large inputs for the remaining region
63 CPU Time Efficiency The relationship between varioushybrids with respect to the central processing time (CPU)was profiled for different runs (Table 9) and expressed in asixth-order polynomial given as 119884
119903
= 1205737
1198836
1
+ 1205736
1198835
1
+ 1205735
1198834
1
+
1205734
1198833
1
+ 1205733
1198832
1
+ 1205732
1198831
+ 1205730
where 1198831
is time (seconds) and 120573
is the coefficient of the polynomial (Figure 8) The proposedhybrid has better usage of CPU time with 119877
2
= 09160followed byHSCR-BF
minus sin andHSCR-BF+ cos with119877
2 of 07551and 07244 respectively (Table 10) Applying the proposedalgorithm into Gaussian the CPU usage had 119877
2 of 07345
10 ISRN Artificial Intelligence
0 50 100 150 200
002
004
006
008
X 249Y 001103
PSO optimized error for hybrid SBF-CRBF with positive nonlinear weight
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 607
08
09
1
11 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(HSCR+cos )
(a) HSCR-BF+ cos
0 50 100 150 200
005
01
015
02
X 249Y 001247
PSO optimized error with nonlinear weight [+sine(R)]
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 6
070809
111 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(b) HSCR-BF+ sin
Figure 4 Hybrid of SBF and CRBF (HSCR-BF) with (a) HSCR-BF+ cos and (b) HSCR-BF
+ sin nonlinear weight
Table 6 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BF+ cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 350780 522257 0311 115650 0121AVG 35078 52226 0031 11565 0012MaxMin 54840 63413 0009 93500 0006
and indeedmarginally outperformedHSCR-BF+ cos Detailed
work on SBF CRBF and GRBF with regard to scalabilitymemory usage and the central processing time has beencarried by the authors [1]
7 Conclusion
In summary we made the following contributions First weused the mix of SBF and CRBF to present several hybridswith different nonlinear weights of cosine and sine func-tions on compact radial basis function Next we showedthe performance of the proposed nonlinear weight hybridsHSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sin and HSCR-BF+ sin optimised all the parameters with minimised error
of 00098 0012 0013 and 0013 respectively compared to00117 for Gaussian HSCR-BF
+ cos The analyzed CPU usagewith corresponding 119877
2 values of 09160 075 072440 06731and 07345 for HSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sinHSCR-BF
+ sin and HSCR-BF+ cos respectively demonstrated
that the algorithm is scalable There exist some evacuationmodels that is Goh and Mandic [11] which offered a choicefor travelers and several schemes for the decisionmakersmaynot be applicable in emergency undergroundmine situationsThe proposed nonlinear hybrid algorithm with particle swarmoptimisation has better capability of approximation to underly-ing functions with a fast learning speed and high robustness and
is competitive and more computationally efficient to Gaussianwith the same nonlinear weight The algorithm is new andthat makes it difficult to identify limitations and we intend toinvestigate other hybrids and comparewith genetic algorithm(GA) in the future
Nomenclature
119873120575min 119873120575max Minimum and maximum signalreach
HSCR-BFHSGR-BF Hybrid of SBF with CRBFHybrid ofSBF with GRBF
HSCR-BFminus cos Hybrid of SBF and CRBF minuscos
nonlinear weightHSCR-BF
+ cos Hybrid of SBF and CRBF +cosnonlinear weight
HSCR-BFminus sin Hybrid of SBF and CRBF minussine
nonlinear weightHSCR-BF
+ sin Hybrid of SBF and CRBF +sinenonlinear weight
HSGR-BF+ cos Hybrid of SBF and GRBF +cos
nonlinear weight
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
ISRN Artificial Intelligence 11
Table 7 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BF+ sin)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 260360 407109 03457 99800 0129AVG 26036 40711 00346 9980 0013MinMax 27840 56040 0008 106500 0014
Table 8 Hybrid of SBF and GRBF with nonlinear weight on GRBF (HSGR-BF+ cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 219840 349010 0439 112100 01173AVG 21984 34901 0044 11210 00117MinMax 27000 46580 0008 80000 00165
0 50 100 1500
005
01
015
02
X 144Y 0009595
GRBF-PSO optimized error with nonlinear weight [+cos(R)]
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 604
06
08
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
Figure 5 Hybrid of SBF and GRBF with nonlinear weight on GRBF(HSCR-BF
+ cos)
0 2 4 6 8 10
002040608
1121416
Weighed position of particles
Min
imise
d er
ror
Optimised error in nonlinear weighed SBF-CRBF hybridsParticles
Less sensitive to inputsmuch closer to target
more accurate
Less sensitive to inputsmuch closer more accurate
Neurons are more sensitive to inputsfast and far from target
minus02minus2minus4minus6minus8minus10
[x = minus1 y = 05785]
[x = 1 y = 04215]
[minus1 004062]
[x = minus2 y = 006288]
[x = minus4 y = 0006014]
[minus3 00404]
Hybrid+cosHybrid+sin
HybridminuscosHybridminussin
Figure 6 Particles weighed position
Acknowledgments
The authors would like to appreciate the immense contribu-tion of the mining companies where the study was under-taken The authors are grateful to these individuals for
0 1 2 3 4 5
0
05
1
15
2
X 37Y 0989
Particles position
Opt
imise
d er
ror
Nonlinear weight hybrids
Keep large inputvalues more
nonlinear
Keep small inputvalues lessnonlinear
Large input values
Small input values
Using positive cosine and sine
Using negative cosine and sineminus05
minus5 minus4 minus3 minus2 minus1
HSCR-BF+cosHSGR-BF+cos
HSCR-BF+sin
HSCR-BFminuscos
HSCR-BFminuscos
HSCR-BFminussin
HSCR-BF+cos HSCR-BF+sin HSCR-BFminussinand HSGR-BF+cos
Figure 7 Nonlinear weighed curves for optimized error withoutGaussian negative cosine curve
HSGR-BF+cos
HSGR-BF+cos
HSCR-BF+sinHSCR-BFminuscosHSCR-BFminussin
00020000400006000080000
100000120000
0 2 4 6 8 10 12
CPU
tim
e
Time (s)
Nonlinear hybrids
Figure 8 The trend of hybrids CPU time curves using sixth-OrderPolynomial
their immeasurable contributions to this work Fred Atta-kuma Patrick Addai Isaac Owusu Kwankye Thomas KwawAnnan Willet Agongo Nathaniel Quansah Francis OwusuMensah Clement Owusu-Asamoah Joseph Adu-MensahShadrack Aidoo Martin Anokye F T Oduro Ernest Ekow
12 ISRN Artificial Intelligence
Table 9 CPU time usage efficiency for the hybrids
Runs HSCR-BFminus cos HSCR-BF
minus sin HSCR-BF+ cos HSCR-BF
+ sin HSGR-BF+ cos
Total 255889 254350 422344 643962 438133Average 25589 25435 42234 64396 43813
Table 10 Analysis of CPU time on the various hybrids
1205736
1205735
1205734
1205733
1205732
1205731
1205730
1198772
HSCR-BFminus cos 00159119909 minus02575119909 minus0759119909 33494119909 minus19901119909 40345119909 25322 09160
HSCR-BF+ cos 00903119909 minus23158119909 19932119909 minus58027119909 minus31337119909 32808119909 14747 07244
HSCR-BFminus sin minus0033119909 11354119909 minus15811119909 11368119909 minus4337119909 78008119909 minus20794 07551
HSCR-BF+ sin 02571119909 minus14263119909 17173119909 minus10069119909 2993119909 minus42374119909 27879 06731
HSGR-BF+ cos minus00437119909 10911119909 minus82923119909 66218119909 15934119909 minus53939119909 8191 07345
Abano E Omari-Siaw and Qiang Jia This work was sup-ported by the National Natural Science Foundation of China(no 71271103) and by the Six Talents Peak Foundation ofJiangsu Province
References
[1] M O Ansong H X Yao and J S Huang ldquoRadial and sigmoidbasis functions neural networks in wireless sensor routingtopology control in underground mine rescue operation basedon particle swarm optimizationrdquo International Journal of Dis-tributed Sensor Networks-Hindawi vol 2013 Article ID 37693114 pages 2013
[2] R Neruda and P Kudova ldquoLearning methods for radial basisfunction networksrdquo Future Generation Computer Systems vol21 no 7 pp 1131ndash1142 2005
[3] C T Cheng J Y Lin and Y G C Sun ldquoLong-term predictionof discharges in Manwan hydropower using adaptive-network-based fuzzy inference systems modelsrdquo in Advances in NaturalComputation vol 3498 of Lecture Notes in Computer Sciencepp 1152ndash1161 2005
[4] D Broomhead and D Lowe ldquoMultivariate functional interpo-lation and adaptive networksrdquo Complex System vol 2 pp 321ndash355 1988
[5] Y Chen C Chuah and Q Zhao ldquoNetwork configuration foroptimal utilization efficiency of wireless sensor networksrdquo AdHoc Networks vol 6 no 1 pp 92ndash107 2008
[6] P S Rajpal K S Shishodia and G S Sekhon ldquoAn artificialneural network for modeling reliability availability and main-tainability of a repairable systemrdquo Reliability Engineering andSystem Safety vol 91 no 7 pp 809ndash819 2006
[7] W Jang W M Healy and S J Mirosław ldquoWireless sensornetworks as part of a web-based building environmental mon-itoring systemrdquo Automation in Construction vol 17 no 6 pp729ndash736 2008
[8] M S Pan C H Tsai and Y C Tseng ldquoEmergengy guiding andmonitoring application in indoor 3D environment by WirelessSensor Networkrdquo Internation Journal Os Sensor Netwourk vol1 pp 2ndash10 2006
[9] S Zarifzadeh A Nayyeri and N Yazdani ldquoEfficient construc-tion of network topology to conserve energy in wireless ad hocnetworksrdquo Computer Communications vol 31 no 1 pp 160ndash173 2008
[10] R Riaz A Naureen A Akram A H Akbar K Kim and HFarooq Ahmed ldquoA unified security framework with three keymanagement schemes for wireless sensor networksrdquo ComputerCommunications vol 31 no 18 pp 4269ndash4280 2008
[11] S L Goh and D P Mandic ldquoAn augmented CRTRL forcomplex-valued recurrent neural networksrdquo Neural Networksvol 20 no 10 pp 1061ndash1066 2007
[12] L A B Munoz and J J V Ramosy ldquoSimilarity-based heteroge-neous neural networksrdquo Engineering Letters vol 14 no 2 pp103ndash116 2007
[13] R Taormina K Chau and R Sethi ldquoArtificial neural networksimulation of hourly groundwater levels in a coastal aquifer sys-tem of the Venice lagoonrdquo Engineering Applications of ArtificialIntelligence vol 25 no 8 pp 1670ndash1676 2012
[14] K Leblebicioglu and U Halici ldquoInfinite dimensional radialbasis function neural networks for nonlinear transformationson function spacesrdquo Nonlinear Analysis Theory Methods andApplications vol 30 no 3 pp 1649ndash1654 1997
[15] F Fernandez-Navarro C Hervas-Martınez J Sanchez-Monedero and P A Gutierrez ldquoMELM-GRBF a modifiedversion of the extreme learning machine for generalized radialbasis function neural networksrdquo Neurocomputing vol 74 no16 pp 2502ndash2510 2011
[16] C L Wu KW Chau and Y S Li ldquoPredicting monthly stream-flow using data-drivenmodels coupledwith data-preprocessingtechniquesrdquoWater Resources Research vol 45 no 8 Article IDW08432 2009
[17] C EACheng ldquoLong-termprediction of discharges inManwanReservoir using artificial neural network modelsrdquo in Advancesin Neural NetworksmdashISNN 2005 vol 3498 of Lecture Notes inComputer Science pp 1040ndash1045 2005
[18] H H Zhang M Genton and P Liu ldquoCompactly supportedradial basis function kernelsrdquo 2004
[19] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquo inProceedings of the 1995 IEEE International Conference on NeuralNetworks pp 1942ndash1945 December 1995
[20] J Eberhart R Eberhart and Y Shi ldquoParticle swarm optimiza-tion developments applications and resourcesrdquo IEEE vol 1 pp81ndash86 2001
[21] R Malhotra and A Negi ldquoReliability modeling using ParticleSwarm optimisationrdquo International Journal of System AssuranceEngineering and Management vol 4 no 3 pp 275ndash283 2013
[22] D Gies and Y Rahmat-Samii ldquoParticle swarm optimization forreconfigurable phase-differentiated array designrdquo Microwaveand Optical Technology Letters vol 38 no 3 pp 168ndash175 2003
ISRN Artificial Intelligence 13
[23] F Valdez P Melin and C Oscar ldquoAn improved evolution-ary method with fuzzy logic for combining Particle Swarmoptimization and genetic algorithmsrdquo Applied Soft ComputingJournal vol 11 no 2 pp 2625ndash2632 2011
[24] J Zhang and K Chau ldquoMultilayer ensemble pruning vianovel multi-sub-swarm particle swarm optimizationrdquo Journalof Universal Computer Science vol 15 no 4 pp 840ndash858 2009
[25] F Valdez P Melin and O Castillo ldquoParallel Particle Swarmoptimization with parameters adaptation using fuzzy logicrdquoin Proceedings of the 11th Mexican International Conference onArtificial Intelligence (MICAI rsquo12) vol 2 pp 374ndash385 San LuisPotosi Mexico October 2012
[26] J Peng and C P Y Pan ldquoParticle swarm optimization RBFfor gas emission predictionrdquo Journal of Safety Science andTechnology vol 7 pp 77ndash85 2011
[27] G Ren Z Huang Y Cheng X Zhao and Y Zhang ldquoAnintegrated model for evacuation routing and traffic signaloptimization with background demand uncertaintyrdquo Journal ofAdvanced Transportation vol 47 pp 4ndash27 2013
[28] X Feng Z Xiao and X Cui ldquoImproved RSSI algorithm forwireless sensor networks in 3Drdquo Journal of ComputationalInformation Systems vol 7 no 16 pp 5866ndash5873 2011
[29] F Valdez and P Melin ldquoNeural network optimization witha hybrid evolutionary method that combines particle swarmand genetic algorithms with fuzzy rulesrdquo in Proceedings ofthe Annual Meeting of the North American Fuzzy InformationProcessing Society (NAFIPS rsquo08) NewYork NYUSAMay 2008
[30] F Fernandez-Navarro C Hervas-Martınez P A Gutierrez JM Pena-Barragan and F Lopez-Granados ldquoParameter estima-tion of q-Gaussian Radial Basis Functions Neural Networkswith a Hybrid Algorithm for binary classificationrdquo Neurocom-puting vol 75 pp 123ndash134 2012
[31] H Soh S Lim T Zhang et al ldquoWeighted complex networkanalysis of travel routes on the Singapore public transportationsystemrdquo Physica A vol 389 no 24 pp 5852ndash5863 2010
[32] C K S Kumar R Sukumar and M Nageswari ldquoSensorslifetime enhancement techniques in wireless sensor networksmdasha critical reviewrdquo International Journal of Computer Science andInformation Technology amp Security vol 3 pp 159ndash164 2013
[33] S Li and F Qin ldquoA dynamic neural network approach forsolving nonlinear inequalities defined on a graph and itsapplication to distributed routing-free range-free localizationof WSNsrdquo Neurocomputing vol 117 pp 72ndash80 2013
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ISRN Artificial Intelligence 3
Table 1 Common rocks found in typical mines in relation to hardness or softness
Nonlinear mapping Mica Coal Granite Feldspar Quartz MineralSoftness 070 080 083 086 0875 090Hardness 2 3 5 6 7 9Distance 750m 470m 390m 315m 278 78m
proposed [8ndash14]Thesewere developed depending on specificemergency situations and management requirements Trans-mitting data wirelessly impacts significant benefits to thoseinvestigating buildings allowing them to deploy sensors andmonitor froma remote location [15 16] To effectively gain theneeded results researchers have come out with a number oftechniques to address the problem of topology control (TC)These include localization of nodes and time error and pathloss transmission range and total load eachnode experiencesas well as energy conservation which is very crucial inoptimizing efficiency and minimizing cost in wireless sensornetworks [17ndash19] Minimizing transmission range of wirelesssensor networks is vital to the efficient routing of the networkbecause the amount of communication energy that eachsensor consumes is highly related to its transmission range[6]The nodes signal reach119873120575was defined as the integrationof the change of the minimum and maximum signal reachtaking into consideration the number of cases (120591) of the rockstructure120573 120573 is the rock hardness119873120575 is the signal reach for anode and 119873120575min 119873120575max are minimum and maximum signalreach respectively The node signal reach is calculated as
119873120575 = 119873120575min +1003817100381710038171003817119873120575min minus 119873120575max
1003817100381710038171003817lowastrandom (
1015840
1205731015840
rock 1 1)
119873120575min = min (119871min (RowCol))
119873120575max = max (119871max (RowCol)) (2)
andwhere120573 is a geometric figure in the range of 07 to 09 Fora connection to be made the absolute difference between 119894 119895should be less than the node signal reach-119873120575The connectionmatrix was given as
120593119896 (119894 119895) = 1 if 1003817100381710038171003817119894 minus 119895
1003817100381710038171003817 le 119873120575
0 otherwise(3)
The relationship between rock hardness and the signal reachis a complicated nonlinear function which is related to theskin depth of the rock with alternating currents concentratedin the outer region of a conductor (skin depth) by opposinginternal magnetic fields as follows
Skin depth = radic2
(120588 lowast 120596 lowast 120590) (4)
120588 is material conductivity 120596 is frequency and 120590 is magneticpermeability
The signal (119861-field) is attenuated by cube of distance (119889)
119861 = (119896)1198893
Signal reach (distance) is equal to 3lowast skin depth
Table 1 identifies 6 common rocks found inmines in rela-tion to hardness or softness of each rock
A routing path was modeled using a number of 119879times119879 sizematrices namely the connection matrix (120593119896) routing matrix(120593119903) explosion matrix (120593119909) failed matrix (120593119891) hope matrix(120593ℎ) optimized matrix (120593119900) and the exit matrix (120593119890) Thehardware survival rate vector (120593ℎ) the survival rate vector(120593V) of each miner and the final average survival rate vector(119877) were also generated A sensor node is named by its 3Dinteger (119909 119910 119911) coordinates where 1 le 119909 le 119877 1 le 119910 le 1198621 le 119911 le 119871 for 119879 = 119877 lowast 119862 lowast 119871 being total number ofnodes If the node (119886 119887 119888) is connected with node (119889 119890 119891)
then the element on ((119886 minus 1) lowast 119862 lowast 119871 + (119887 minus 1) lowast 119871 + 119888)throw and ((119889 minus 1) lowast 119862 lowast 119871 + (119890 minus 1) lowast 119871 + 119891)th columnis 1 otherwise 0 and routing are limited to total number ofmultiple points connections that can be made In arriving atthe final optimized vector for transmission each matrix wasgenerated 120591 times
The 120593119903 sub 120593119896 119872120588
le 120572 119872120588
is even 119872120588
representing themaximum point-to-multipoint connection and 119872
120588
is evenallowing bidirectional communication with 119894119895 checkingsource and destination nodes respectively
120593119903 =
1 if 1003817100381710038171003817119894 minus 1198951003817100381710038171003817 ge (
119872120588
2)
0 otherwise
for 119894 119895 = 1 119879
(5)
24 Hardware Software and Network Fault Tolerant Con-siderations Network security is a critical issue in wirelesssensor networks as it significantly affects the efficiency ofthe communication andmany keymanagement schemes hadbeen proposed to mitigate this constraint [18 20] In an eventof accidents (120595) occurring the routing pathwould be affectedby (1minus120595) where120595 is any randomvaluewithin120573 whichwouldcause explosion on 119877120593 matrix and result in 120593119909 such that theresulting matrix would be the failed matrix (120593119891)
120593119909 = (1 minus 120595) 120593119903
120593119891 (119894 119895) = 1 if 120593119909 (119894 119895) lt 120582119871
0 if 120593119909 (119894 119895) ge 120582119871
(6)
120593119891(119894 119895) = 0 if 120593119909(119894 119895) ge 120582119867 else 120593119891 = 120593119909(119894 119895)120582119867 for (120582119871or 120582119867) representing the lower and higher accident impactthresholds respectively
A new set of routing paths (120593ℎ) and exit matrices (120593119890) fortransmission was calculated as
120593ℎ = 120593119891 lowast 120593119903
120593119890 = 119873119890 minus 120595
(7)
4 ISRN Artificial Intelligence
The mathematical objective here is to find an optimizedrouting matrix (120593119900) that has the maximum survivability Thesurvivability is defined as the entropy of a number of parallelconnections between every node to all the sink(s) The exitmatrix (120593119890) described the success rate from each node to thesink(s) 120593119890 assumes that the number of exits (119873119890) is availablewith an errormargin (120595) Inmost practical applicationsmorethan one sink is used and sink node is either through the fiberorThrough-The-Earth (TTE) link It is important to note thatin real rescue situations the software (relational databasemanagement system (RDBMS)) and hardware includingradio frequency identification (RFID) may fail as a result ofthe effect from the explosion and attack The matrix (120593119904)
was used to describe software survival rate including bugs orattacks
120593119904 = 1 minus ((1
119879 + random) (ldquoGeometricrdquo fail 119879 119879)) (8)
To obtain the final survival vector (120593119877) it was assumed thateach miner will have an RFID a vector 120593119868 was used todescribe its failure rate including risks of running out ofbattery a vector 120593119867 for the hardware failure rate for119879 = totalnumber of nodes
120593119868 = 1 minus ((1
119879 + 119903) lowast (ldquoGeometricrdquo fail 1 119879)) (9)
119903 is a random number generated from the vector 119879 and 119903
0 rarr infin 119879 + random 119879 rarr infin (1(119879 + 119903)) 1119879 rarr 01 minus (1(119879 + 119903)) 1 minus 1119879 rarr 1 is the minimum thereforefor 119879 = 119873 nodes we have 1 minus 1119873 = (119873 minus 1)119873 rarr 1 for(119873 minus 1)119873 rarr node is dead and 1 rarr node is alive
120593119867 = min (1 120593119868 [119883Ψ]) for [119883Ψ]) =120593119890
119872119901
(10)
The survival rate of each miner was (120593V) All these assump-tions happen in real life and must be considered The finalsurvival rate vector (120593119877) was calculated
120593V = 120593119904 lowast 120593119890 (11)
120593119877 = 120593V lowast 120593119904 (12)
119877 = 120593119877 (13)
3 Related Work
31 Artificial Neural Networks (ANN) Having found theoptimum set of routing table that has the highest survivalprobability of communicating with and rescuing miners it isimportant to train the neurons such that the initial error willbe minimized and more importantly have a reliable system[21]The topology of a neural network can be recurrent (withfeedback contained in the network from the output back tothe input) the feedforward (where data flow from the input tothe output units)The data processing can extend over severallayers of units but no feedback connections are presentManyresearchers have come out with neural network predictivemodels in both sigmoid and radial basis functions [22 23]
with applications such as nonlinear transformation [23]extreme learning machine predicting accuracy in gene clas-sification [24] crisp distributed support vectors regression(CDSVR) model was monthly streamflow prediction wasproposed by Valdez et al [25] and other applications includefuzzy inference systems (FIS) which have been successfullyapplied in fields such as automatic control data classificationdecision analysis and expert systems [26] among othersArtificial neural networks (ANN) are learning algorithmusedto minimize the error between the neural network outputand desired output This is important where relationshipsexist between weights within the hidden and output layersand among weights of more hidden layers In addition otherparameters including mean iteration standard variationstandard deviation and convergent time (in sections) wereevaluated The architecture of the learning algorithm andthe activation functions were included in neural networksNeurons are trained to process store recognize and retrievepatterns or database entries to solve combinatorial optimiza-tion problems Assuming that the input layer has 4 neuronsoutput layer 3 neurons and the hidden layer 6 neurons wecan evolve other parameters in the feedforward network toevolve the weight So the particles would be in a group ofweights and there would be 4 lowast 6 + 6 lowast 3 = 42 weightsThis implies that the particle consists of 42 real numbersThe range of weight can be set to [minus100 100] or a fittingrange After encoding the particles the fitness function wasthen determined The goodness of the fit was diagnosedusing mean squared error (MSE) as against mean cubicerror (MCE) and the mean absolute error (MAE) The meancubic error will allow for fast convergence and that willgross over accuracy making the process unstable while themean absolute error is stable but converges slowly A midwaybetween the two is the MSE and is given as
MSE =1
119899119904
119899
sum
119895=1
119904
sum
119894=1
(119884119895119894
minus 119910119895119894
)2
(14)
where 119899 is number of samples 119904 is the number of neurons atoutput layer 119884
119894119895
is the ideal value of 119894th sample at 119895th outputand 119884
119894119895
is the actual value of 119894th sample at 119895th output
32 The Gaussian and Zhang Models The standard or directapproach to interpolation at locations 119909
1
119909119873
sub 119877119889
without the first term using Gaussian kernel is given as
120601 (119903) = 119890minus(120576119903)
2
(15)
Zhang has [27] tried tomodify theGaussianmodel as followsa function 120595 [0infin) rarr R such that 119896(119909 1199091015840) = 120593(119909 minus 119909
1015840
)where 119909 1199091015840 isin 120594 and 119909 sdot 119909
1015840 denotes the Euclidean norm with119896(119909 119909
1015840
) = exp(minus119909 minus 1199091015840
2
1205752
) as an example of the RBFkernels The global support for RBF radials or kernels hasresulted in dense Grammatrices that can affect large datasetsand therefore construct the following two equations119896119862V(119909 119909
1015840
) = 120601119862V(119909 minus 119909
1015840
)119896(119909 1199091015840
) and 120601119862V(119909 minus 119909
1015840
) =
[(1 minus 119909 minus 1199091015840
119862)(sdot)+
]V where119862 gt 0 V ge (119889+1)2 and (sdot)
+
isthe positive part The function 120601C(sdot) is a sparsifying operator
ISRN Artificial Intelligence 5
which thresholds all the entries satisfying (119909 minus 1199091015840
) ge 119862
to zeros in the Gram matrix The new kernel resulting fromthis construction preserves positive definiteness This meansthat given any pair of inputs 119909 and 119909
1015840 where 119909 = 1199091015840 the
shrinkage (the smaller 119862) is imposed on the function value119896(119909 119909
1015840
) the result is that the Gram matrices 119870 and 119870119862V can
be either very similar or quite different depending on thechoice of 119862 However Zhang also ended up with an extrapower parameter
4 The Proposed Hybrid
The previous work looked at the Gaussian model and par-alyzed the computational power parameter The result wascompact radial basis function (CRBF) This was used to runboth SBF and RBF and the results compared The scalabilityand processing efficiency were also analyzed
The novelty of this algorithm the weighed nonlinearhybrid was to find several hybrids and the best for rescueoperationsThe cosine and sine functionswere used to reducehigh level nonlinear and increase small level nonlinearBoth the previous and the current algorithm used the samepreliminary considerations but the results is slightly differentbecause data are random
The sigmoid basis function was given as
log sig (119877) 119877 = 119882 sdot 119875 + 119861
log sig (119882 sdot 119875 + 119861) =1
1 + 119890minus(119882sdot119875+119861)
(16)
Neuron function 119878 (sigmoid) is log sig 119882 is weight matrix119875 is input vector and 119861 is threshold We therefore proposeda compact radial basis function based on the Gaussian radialbasis function and Helensrsquo [28] definition expressed as
[exp (minus 119877)] 119877 = 119882 sdot 119875 + 119861
(exp (minusabs ((119882 sdot 119875 + 119861)))
120601 (out) = [exp (minus 119877)]
(17)
119882 is weight matrix 119875 is input vector and 119861 is threshold Thefocus was to improve on the radial basis function for themineapplication
An optimized vector 119877 was generated as the optimumset of transmission routing table that has the highest survivalprobability for data transmission (13)
119875119873 =
119873
sum
119894=1
119877119894
119878119894
+
2
sum
119894=1
119878119894
+
2
sum
119894=1
119878119894
119878119894
+ 1
for 1 le 119894 le 119878 1 le 119896 le 119877
(18)
119877 1198781
1198782
are number of neurons at input hidden and outputlayers respectively
119875119873 is the position of the 119899th particle
119882119894
= 119878119894
1198771198822
= 119878119894
119877 119882119898
= 119878119898
119877
119875119894
= [119877 (119894 minus 119895) + 119870] = 119882119894
(119894 119896)
(19)
There are two thresholds (1198781
rArr 1198611
)Hidden 119861
1
(119894 119895) = 119875(1198771198781
+ 1198781
1198782
) rArr S1
Output 1198612
(119894 119895) = 119875(1198771198781
+ 1198781
1198782
) rArr 1198782
From (16) (17) and (18) the nonlinear weight of [plusmn cos(119877)] or[plusmn sin(119877)] was imposed on the CRBF before being combinedwith the SBF as (HSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sinHSCR-BF
+ sin)The cosine weight was used to keep high levelnonlinear for small input value and reduce the nonlinear forlarge input values while sine weight was used to keep highlevel nonlinear for large input value and reduce the nonlinearfor small input values
HSCR-BFminus cos
119884 (out) = log sig (119877) + [exp (minus 119877)] sdotlowast
[minus cos (119877)]
HSCR-BF+ cos
119884 (out) = log sig (119877) + [exp (minus 119877)] sdotlowast
[+ cos (119877)]
HSCR-BFminus sin
119884 (out) = log sig (119877) + [exp (minus 119877)] sdotlowast
[minus sin (119877)]
HSCR-BF+ sin
119884 (out) = log sig (119877) + [exp (minus 119877)] sdotlowast
[+ sin (119877)]
(20)
Thenonlinear weight of [+ cos(119877)]was imposed on theGRBFbefore being combined with the SBF
HSGR-YBF+ cos
119884 (out) = log sig (119877) + [exp (minus1198772
)] sdotlowast
[+ cos (119877)]
(21)
5 Particle Swarm Optimization
Particle swarm optimization (PSO) an evolutionary algo-rithm is a population based stochastic optimization tech-nique The idea was conceived by an American researcherand social psychologist James Kennedy in the 1950s Thetheory is inspired by social behavior of bird flocking or fishschooling The method falls within the category of swarmintelligence methods for solving global optimization prob-lems Literature has shown that the PSO is an effectivealternative to established evolutionary algorithms (GA) It isalso established that PSO is easily applicable to real worldcomplex problems with discrete continuous and nonlineardesign parameters and retains the conceptual simplicity ofGA [29 30] Each particle within the swarm is given an initialrandom position and an initial speed of propagation Theposition of the particle represents a solution to the problemas described in a matrix 120591 where 119872 and 119873 represent thenumber of particles in the simulation and the number ofdimensions of the problem respectively [31 32] A randomposition representing a possible solution to the problem withan initial associated velocity representing a function of thedistance from the particlersquos current position to the previous
6 ISRN Artificial Intelligence
position of good fitness value was given A velocity matrix119881el with the same dimensions as matrix 120591
119909
described this
120591119909
= (
12059111
12059112
1205911119873
12059121
12059122
1205912119873
1205911198981
1205911198982
120591119872119873
)
119881el = (
V11 V12 V1119873
V21 V22 V2119873
V1198981 V1198982 V119872119873
)
(22)
The best feasible alternative is the genetic algorithmHoweverthis is generally slow as compared to the PSO and thereforecomputationally expensive hence the PSO
While moving in the search space particles commit tomemorize the position of the best solution they have foundAt each iteration of the algorithm each particle moves witha speed that is a weighed sum of three components the oldspeed a speed component that drives the particles towardsthe location in the search space where it previously foundthe best solution so far and a speed component that drivesthe particle towards the location in the search space wherethe neighbor particles found the best solution so far [7] Thepersonal best position can be represented by an119873times119873matrix120588best and the global best position is an 119873-dimensional vector119866best
120588best = (
12058811
12058812
1205881119873
12058821
12058822
1205882119873
1205881198981
1205881198982
120588119872119873
)
119866best = (119892best1
119892best12
119892best119873
)
(23)
All particles move towards the personal and the globalbest with 120591 120588best 119881el and 119892best containing all the requiredinformation by the particle swarm algorithmThese matricesare updated on each successive iterations
119881119898119899
= 119881119898119899
+ 1205741198881
1205781199031
(119901best119898119899
minus 119883119898119898
)
+ 1205741198882
1205781199032
(119892best119899
minus 119883119898119899
)
119883119898119899
= 119881119898119899
(24)
1205741198881
and 1205741198882
are constants set to 13 and 2 respectively and1205781199031
1205781199032
are random numbers
51 AdaptiveMutation according toThreshold Particle swarmoptimization has been effective in training neural networkssuch as a Parallel Particle Swarm optimization (PPSO)methodwith dynamic parameter adaptation used to optimizecomplex mathematical functions and improved evolutionarymethod in fuzzy logic [30 31 33] To prevent particlesfrom not converging or converging at local minimum an
Input nodes64748
64748
64748
Hidden nodesOutput nodes
S11R
S12R
SR
120573 ge 07
S1mR
120573 ge 07 le 09
120573 ge 07 le 09
Figure 1 The structure AMPSO for CRBF GRBF and SBF transferfunctions
1 2 3 4 5 6 7 8 9 100
02040608
11214
The impact of the accident
Link
avai
labi
lity
The curve of transmission link after accident
The higher the impact the lower the link availability
Region 3Links are completely dead
Region 1Links are not
Region 2Indicate probability of link being able to transmit data
affected
Figure 2 Impact of explosionaccident on transmission link
adaptivemutation according to thresholdwas introduced Analternative to this is the use of the nonadaptive mutation andthis could converge to the local minimum or fail to convergeat all The advantage of the nonadaptation is simple and fastbut may not provide the needed results
Particles positions were updated with new value onlywhen the new value is greater than the previous value 20 ofparticles of those obtaining lower values weremade tomutatefor faster convergence and the structure of adaptive mutationPSO (AMPSO) with threshold can be found in [1] The inputlayer takes the final survival vector (13) with a number ofhidden layers and an output layer The feedforward neuralnetwork was used The structure of Adaptive Mutation PSO(AMPSO) with threshold Neural Network was used Figure 1[32]
6 Results and Discussion
61 Generating the Routing Path Elements 0 1 and 2 in 120593119909
imply that the link(s) were not affected and elements 3 45 and 6 represent a probability for the links being able totransmit data while figures from 7 means the link is totallydead Region 1 that indicates that links are not affected region
ISRN Artificial Intelligence 7
Table 2 Optimum set of routing table redundancy with the highest survival probability (6 cases)
Vectors for each case Rock hardness cases1198771 05409 04281 08834 0585 08134 05163 071198772 04526 05841 04547 08997 08436 02696 081198773 02931 05899 05903 08309 06971 02898 0831198774 04851 06085 06913 04002 05026 0 0861198775 04917 05857 07805 06037 05182 06472 08751198776 02902 07253 0458 05762 01397 0 09
2 gives the probability of link available and the last regionindicates the link is completely down (Figure 2)
Thematrices120593119894 120593119896 120593119903 120593119909 120593ℎ for 120591 = 119899 with dimensionsof 119871 = 3 119877 = 2 and119862 = 1 for depth (level) row (length) andwidth (column) respectively were generated as follows
119878eq = (
(
1 1 1
1 2 1
2 2 1
2 1 1
3 1 1
3 2 1
)
)
119894120593 =
[[[[[[[
[
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
]]]]]]]
]
120593119896 = (
(
1 1 1 1 0 0
1 1 1 0 0 0
0 1 1 1 0 0
0 0 1 1 1 0
0 0 0 1 1 1
0 0 0 1 1 1
)
)
1 = signal points
120593119903 = (
(
1 1 1 0 0 0
1 1 1 0 0 0
0 1 1 1 0 0
0 0 1 1 1 0
0 0 0 1 1 1
0 0 0 1 1 1
)
)
1 = transmission path
120593119909 = (
(
0 1 2 2 0 0
1 5 1 2 0 0
0 8 1 5 9 1
1 1 5 1 1 17
1 1 4 11 0 2
4 4 19 5 3 0
)
)
explosion sizes
120593119891 = (
(
1 1 1 1 1 1
1 6 1 1 1 1
1 0 1 6 0 1
1 1 6 1 1 0
1 1 75 0 1 1
75 75 0 6 1 1
)
)
(25)
120593ℎ = (
(
1 1 1 0 0 0
1 6 1 0 0 0
0 0 1 6 0 0
0 0 6 1 1 0
0 0 0 0 1 1
0 0 0 6 1 1
)
)
120593119900 = (
(
8 8 8 0 0 0
8 48 8 0 0 0
0 0 8 48 0 0
0 0 48 8 8 0
0 0 0 0 8 8
0 0 0 48 8 8
)
)
120593119890 = (
(
144 144 144 0 0 0
144 864 144 0 0 0
0 0 144 864 0 0
0 0 864 144 144 0
0 0 0 0 144 144
0 0 0 864 144 144
)
)
1 or fractional figures are transmission path
(26)
Element ldquo0rdquo on 120593119891 depicts a connection lost while ldquo1rdquo meansthat the connection will never go down and represent theconnection to the fixed sink node(s) along the edge or theemergency connection to the mobile data mule(s)
120593s = (
(
8333 8889 9091 8333 8571 8333
8889 8333 8750 8571 8750 8333
8571 9167 8750 9091 8889 8333
8750 8571 8571 8571 8571 8333
8333 8333 8571 8333 8889 9000
9167 8333 9091 8333 8889 8571
)
)
(27)
The survival rate of RFID hardware and each minervectors was displayed as
120593119868 = 8571 8750 8571 8889 8571 8333
120593119867 = 6236 4976 1 6851 9286 6086
120593V = 5409 4281 8834 5850 8134 5163
(28)
Optimization was done numerically usingMatlab simulationtool to find the optimum set of routing tables throughparticle swarm search for rescue operation as discussed inthe preliminaries From (13) the final survival vector fordimensions of 119871 = 3 119877 = 2 and 119862 = 1 will be an average ofthe 6 vectors as in Table 2 Total cases 119877119899 = 120591 = 119879times119879 (which
8 ISRN Artificial Intelligence
Table 3 Scalability of model to survival probability range robot location and rock type
Location Mica Coal Granite Feldspar Quartz MineralCompact radial basis function with negative nonlinear weight (HSCR-BF
minus cos)(10 6 5) 08769ndash10000 09070ndash09927 08402ndash09943 08023ndash09778 07386ndash09376 07345ndash08740(10 5 4) 09024ndash09696 09032ndash09505 08836ndash09430 08633ndash09527 07944ndash09465 07263ndash08720(6 5 4) 08906ndash09877 08827ndash09845 08384ndash09675 08627ndash09820 07911ndash09142 06392ndash07840(3 1 10) 07302ndash08481 06588ndash07816 06341ndash07473 06503ndash07249 05930ndash07156 04445ndash05621
Gaussian radial basis function with positive nonlinear weight (HSGR-BF+ cos)
(10 6 5) 08732ndash09963 08526ndash09987 08492ndash09950 07867ndash09568 07404ndash09025 06606ndash08315(10 5 4) 08843ndash09593 08843ndash09853 08444ndash09753 07746ndash09828 07304ndash09595 06678ndash08817(6 5 4) 08785ndash09983 08902ndash09996 08200ndash09997 07927ndash09446 07065ndash08565 06572ndash08063(3 1 10) 07608ndash08552 06600ndash08325 06090ndash07371 05482ndash06249 05026ndash06056 04856ndash05453
10 20 30 40 50 60 70 80 90 1000
02
04
06
X 94Y 0009904
Number of iterations
Fitn
ess v
alue
PSO optimized error for hybrid SBF-CRBF with negative nonlinear weight
15 2 25 3 35 4 45 5 55 604
06
08
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(HSCRminuscos )
(a) HSCR-BFminus cos
50 100 150 200 2500
01
02
03
X 249Y 001151
PSO optimized error with nonlinear weight [minussine(R)]
Number of iterations
Fitn
ess v
alue
1 2 3 4 5 6
07
08
09
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(b) HSCR-BFminus sin
Figure 3 Hybrid of SBF and CRBF (HSCR-BF) with (a) HSCR-BFminus cos and (b) HSCR-BF
minus sin nonlinear weight
represent the total nodes deployed) for this scenario Rocksoftness or hardness takes the values of 120573 and each matrix isgenerated 120591 timecases and the final routing vector will be anaverage of the 119899 vectors 120593119877 = Average (1198771 1198776) [Table 2]which becomes the input for the PSO training
The final survival rate is 119877 = (05409 04281 08834
0585 08130 05163) from (13)
62 The Final Survival Vector The 119877 = (05409 04281
08834 0585 08130 05163) from (13) is the maximumsurvival probability for a total of 6 nodes deployed Itdescribes the success rate from each node to the sink(s) Inmost practical applications more than one sink is used andsink node is either through the fiber or TTE connection Thesize of the vector depends on the dimensions of the fieldThe elements 119877 = (05409 04281 08834 0585 08130
05163) represent the probability of 54 43 88 5981 and 52 success of each node transmitting data to andfrom its source or destination It assists decision makers as towhether to send data through one ormore nodes or send each
message twice The total nodes used for the simulation were300 with undergroundmine dimensions of 119871 = 10119877 = 6 and119862 = 5 for depth (level) row (length) and width (column)respectively with ldquopmrdquo a sensor apart pm = 100119872
120588
= 4 and119873119890 = 2 The PSO training used swarm size of 20 maximumposition was set to 100 max velocity = 1 and maximumnumber of iteration = 250 The thresholds 120582119871 and 120582119867 were3 and 6 respectively 120591 = 6 cases (each case representa rock hardness case and also represent the total neuronsused) thus each matrix and vectors were run 6 times beforeneural training The survival probability (bottom) indicatesthat the model survived between 90 and 100 where rockcases were relatively soft (ge07) The survival probabilitydeclines as the rock becomes harder and approaches 08 Atthe hardest rock of 09 the survival probability fell between72 and 84 for the entire hybrids In view of this the APheads had to be deployed at a location where the rock isrelatively soft for maximum signal strength Each computersimulation incorporated all the 6 different cases of rockhardnesssoftness to produce the matrices Detailed locationanalysis or the scalability of the model in relation to survival
ISRN Artificial Intelligence 9
Table 4 Hybrid of SBF and CRBF with negative nonlinear weight on CRBF (HSCR-BFminus cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 231020 343498 0382 91650 00985AVG 23102 34350 0038 9165 00098MinMax 29540 45314 0024 80000 00043
Table 5 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BFminus sin)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 307320 488381 0376 114300 0135AVG 30732 48838 0038 11430 0013MinMax 30740 56926 0007 61500 0021
probability range robot location and different rock types isrecorded in Table 3 Figures 4ndash6 represent different scenariosof SBF and CRBF hybrids with the [plusmn cos(119877)] and [plusmn sin(119877)]
nonlinear weights while Figure 7 represents the SBF andGaussian hybrids for [+ cos(119877)] nonlinear weight The tophalf of each figure indicates the optimized error or the finalerror after the neural training and the bottom half revealsmodel survival probability
Figure 3 discusses the SBF and CRBF hybrids (HSCR-BF) with nonlinear weight Training in negative nonlinearweights (HSCR-BF
minus cos HSCR-BFminus sin) responded well The
HSCR-BFminus cos had a steady and compact routing path which
was consistent through all rock layers with initial survivalprobability between 877 and 100 in soft layers decliningto 735ndash874 at harder rock layers The HSCR-BF
minus sinperformed quite well but was more dispersed (879 to 985at the soft rock and 665 to 818 at the hard rock)
In addition the various parameters mean iterationstandard variance standard deviation and the convergenttime of HSCR-BF
minus cos were optimised This is demonstratedby finding the relationship between the maximum and theminimum values of the parameters and comparing with theaverage figures and the closer the difference is to the averagethe more consistent the data are in the dataset (Table 4) thusHSCR-BF
minus cos provided the best results among the proposedhybrids The HSCR-BF
minus sin on the RBF yielded strength inmean iteration and standard variance with error of 0013(Table 5)
In Figure 4 training with nonlinear weight of positivecosine and sine of HSCR-BF (HSCR-BF
+ cos and HSCR-BF+ sin) had some similarities The two models streamed
well at the initial stages from 896 to 992 and 908to 990 at soft layers of the rock respectively Howeverat the last stages the probability of transmitting effectivelybecame marginal from 795 to 852 and 821 to 854respectively Much as this hybrid could perform well in areaswhere routing conditions aremuch better in less dense rescuesituations this could hamper rescue mission in both casesdue to battery drain or collision from traffic conjunctionThedata formean iteration and convergent time forHSCR-BF
+ cos(Table 6) and mean Iteration standard variance the final
error HSCR-BF+ sin (Table 7) were consistent with the nega-
tive sine having better results than the positive cosineThe survival probability of the positive nonlinear weight
Gaussian hybrids (HSGR-BF+ cos) in Figure 5 is more com-
pact both at the initial (ie 873ndash996) and later (661ndash832) stages declining to harder rock layers with an errorof 001173 (Table 8)
At the initial stage particles are sensitive to inputs asthey moved quickly in the search space towards the targetand while particles peaked closer to the target it became lesssensitive to the input and began to align (Figure 6) At thelater stage more RBF are used to keep the error at minimumfor accuracy It was also noticed that instead of acceler-ating higher and becoming more sensitive to the inputswhile particles were far from the target the hybrids withlowast
[minus sin(119877)] and lowast[ + sin(119877)] were less sensitive to the inputsas lowast[minus sin(119877)] descended from [119909 = minus1 119910 = 05785] to[119909 = 1 119910 = 04215] while lowast[+ sin(119877)] descended from [119909 =
minus3 119910 = 00404] to [119909 = minus1 119910 = minus004062] before becomingconscious of the inputs In addition the hybrids with thelowast
[+ cos(119877)] lagged slightly between [119909 = minus4 119910 = 0006014]
and [119909 = minus2 119910 = 006288] before becoming sensitive toinputs However they all lined up for accuracy in terms of theminimised error
The nonlinearity of the hybrids is presented (Figure 7)the negative cosinesine weight is used to reduce nonlinearfor small input values and this lies between minus1 lt 119909 lt +1 andthe positive cosinesine weight is used to keep high nonlinearfor large inputs for the remaining region
63 CPU Time Efficiency The relationship between varioushybrids with respect to the central processing time (CPU)was profiled for different runs (Table 9) and expressed in asixth-order polynomial given as 119884
119903
= 1205737
1198836
1
+ 1205736
1198835
1
+ 1205735
1198834
1
+
1205734
1198833
1
+ 1205733
1198832
1
+ 1205732
1198831
+ 1205730
where 1198831
is time (seconds) and 120573
is the coefficient of the polynomial (Figure 8) The proposedhybrid has better usage of CPU time with 119877
2
= 09160followed byHSCR-BF
minus sin andHSCR-BF+ cos with119877
2 of 07551and 07244 respectively (Table 10) Applying the proposedalgorithm into Gaussian the CPU usage had 119877
2 of 07345
10 ISRN Artificial Intelligence
0 50 100 150 200
002
004
006
008
X 249Y 001103
PSO optimized error for hybrid SBF-CRBF with positive nonlinear weight
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 607
08
09
1
11 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(HSCR+cos )
(a) HSCR-BF+ cos
0 50 100 150 200
005
01
015
02
X 249Y 001247
PSO optimized error with nonlinear weight [+sine(R)]
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 6
070809
111 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(b) HSCR-BF+ sin
Figure 4 Hybrid of SBF and CRBF (HSCR-BF) with (a) HSCR-BF+ cos and (b) HSCR-BF
+ sin nonlinear weight
Table 6 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BF+ cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 350780 522257 0311 115650 0121AVG 35078 52226 0031 11565 0012MaxMin 54840 63413 0009 93500 0006
and indeedmarginally outperformedHSCR-BF+ cos Detailed
work on SBF CRBF and GRBF with regard to scalabilitymemory usage and the central processing time has beencarried by the authors [1]
7 Conclusion
In summary we made the following contributions First weused the mix of SBF and CRBF to present several hybridswith different nonlinear weights of cosine and sine func-tions on compact radial basis function Next we showedthe performance of the proposed nonlinear weight hybridsHSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sin and HSCR-BF+ sin optimised all the parameters with minimised error
of 00098 0012 0013 and 0013 respectively compared to00117 for Gaussian HSCR-BF
+ cos The analyzed CPU usagewith corresponding 119877
2 values of 09160 075 072440 06731and 07345 for HSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sinHSCR-BF
+ sin and HSCR-BF+ cos respectively demonstrated
that the algorithm is scalable There exist some evacuationmodels that is Goh and Mandic [11] which offered a choicefor travelers and several schemes for the decisionmakersmaynot be applicable in emergency undergroundmine situationsThe proposed nonlinear hybrid algorithm with particle swarmoptimisation has better capability of approximation to underly-ing functions with a fast learning speed and high robustness and
is competitive and more computationally efficient to Gaussianwith the same nonlinear weight The algorithm is new andthat makes it difficult to identify limitations and we intend toinvestigate other hybrids and comparewith genetic algorithm(GA) in the future
Nomenclature
119873120575min 119873120575max Minimum and maximum signalreach
HSCR-BFHSGR-BF Hybrid of SBF with CRBFHybrid ofSBF with GRBF
HSCR-BFminus cos Hybrid of SBF and CRBF minuscos
nonlinear weightHSCR-BF
+ cos Hybrid of SBF and CRBF +cosnonlinear weight
HSCR-BFminus sin Hybrid of SBF and CRBF minussine
nonlinear weightHSCR-BF
+ sin Hybrid of SBF and CRBF +sinenonlinear weight
HSGR-BF+ cos Hybrid of SBF and GRBF +cos
nonlinear weight
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
ISRN Artificial Intelligence 11
Table 7 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BF+ sin)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 260360 407109 03457 99800 0129AVG 26036 40711 00346 9980 0013MinMax 27840 56040 0008 106500 0014
Table 8 Hybrid of SBF and GRBF with nonlinear weight on GRBF (HSGR-BF+ cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 219840 349010 0439 112100 01173AVG 21984 34901 0044 11210 00117MinMax 27000 46580 0008 80000 00165
0 50 100 1500
005
01
015
02
X 144Y 0009595
GRBF-PSO optimized error with nonlinear weight [+cos(R)]
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 604
06
08
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
Figure 5 Hybrid of SBF and GRBF with nonlinear weight on GRBF(HSCR-BF
+ cos)
0 2 4 6 8 10
002040608
1121416
Weighed position of particles
Min
imise
d er
ror
Optimised error in nonlinear weighed SBF-CRBF hybridsParticles
Less sensitive to inputsmuch closer to target
more accurate
Less sensitive to inputsmuch closer more accurate
Neurons are more sensitive to inputsfast and far from target
minus02minus2minus4minus6minus8minus10
[x = minus1 y = 05785]
[x = 1 y = 04215]
[minus1 004062]
[x = minus2 y = 006288]
[x = minus4 y = 0006014]
[minus3 00404]
Hybrid+cosHybrid+sin
HybridminuscosHybridminussin
Figure 6 Particles weighed position
Acknowledgments
The authors would like to appreciate the immense contribu-tion of the mining companies where the study was under-taken The authors are grateful to these individuals for
0 1 2 3 4 5
0
05
1
15
2
X 37Y 0989
Particles position
Opt
imise
d er
ror
Nonlinear weight hybrids
Keep large inputvalues more
nonlinear
Keep small inputvalues lessnonlinear
Large input values
Small input values
Using positive cosine and sine
Using negative cosine and sineminus05
minus5 minus4 minus3 minus2 minus1
HSCR-BF+cosHSGR-BF+cos
HSCR-BF+sin
HSCR-BFminuscos
HSCR-BFminuscos
HSCR-BFminussin
HSCR-BF+cos HSCR-BF+sin HSCR-BFminussinand HSGR-BF+cos
Figure 7 Nonlinear weighed curves for optimized error withoutGaussian negative cosine curve
HSGR-BF+cos
HSGR-BF+cos
HSCR-BF+sinHSCR-BFminuscosHSCR-BFminussin
00020000400006000080000
100000120000
0 2 4 6 8 10 12
CPU
tim
e
Time (s)
Nonlinear hybrids
Figure 8 The trend of hybrids CPU time curves using sixth-OrderPolynomial
their immeasurable contributions to this work Fred Atta-kuma Patrick Addai Isaac Owusu Kwankye Thomas KwawAnnan Willet Agongo Nathaniel Quansah Francis OwusuMensah Clement Owusu-Asamoah Joseph Adu-MensahShadrack Aidoo Martin Anokye F T Oduro Ernest Ekow
12 ISRN Artificial Intelligence
Table 9 CPU time usage efficiency for the hybrids
Runs HSCR-BFminus cos HSCR-BF
minus sin HSCR-BF+ cos HSCR-BF
+ sin HSGR-BF+ cos
Total 255889 254350 422344 643962 438133Average 25589 25435 42234 64396 43813
Table 10 Analysis of CPU time on the various hybrids
1205736
1205735
1205734
1205733
1205732
1205731
1205730
1198772
HSCR-BFminus cos 00159119909 minus02575119909 minus0759119909 33494119909 minus19901119909 40345119909 25322 09160
HSCR-BF+ cos 00903119909 minus23158119909 19932119909 minus58027119909 minus31337119909 32808119909 14747 07244
HSCR-BFminus sin minus0033119909 11354119909 minus15811119909 11368119909 minus4337119909 78008119909 minus20794 07551
HSCR-BF+ sin 02571119909 minus14263119909 17173119909 minus10069119909 2993119909 minus42374119909 27879 06731
HSGR-BF+ cos minus00437119909 10911119909 minus82923119909 66218119909 15934119909 minus53939119909 8191 07345
Abano E Omari-Siaw and Qiang Jia This work was sup-ported by the National Natural Science Foundation of China(no 71271103) and by the Six Talents Peak Foundation ofJiangsu Province
References
[1] M O Ansong H X Yao and J S Huang ldquoRadial and sigmoidbasis functions neural networks in wireless sensor routingtopology control in underground mine rescue operation basedon particle swarm optimizationrdquo International Journal of Dis-tributed Sensor Networks-Hindawi vol 2013 Article ID 37693114 pages 2013
[2] R Neruda and P Kudova ldquoLearning methods for radial basisfunction networksrdquo Future Generation Computer Systems vol21 no 7 pp 1131ndash1142 2005
[3] C T Cheng J Y Lin and Y G C Sun ldquoLong-term predictionof discharges in Manwan hydropower using adaptive-network-based fuzzy inference systems modelsrdquo in Advances in NaturalComputation vol 3498 of Lecture Notes in Computer Sciencepp 1152ndash1161 2005
[4] D Broomhead and D Lowe ldquoMultivariate functional interpo-lation and adaptive networksrdquo Complex System vol 2 pp 321ndash355 1988
[5] Y Chen C Chuah and Q Zhao ldquoNetwork configuration foroptimal utilization efficiency of wireless sensor networksrdquo AdHoc Networks vol 6 no 1 pp 92ndash107 2008
[6] P S Rajpal K S Shishodia and G S Sekhon ldquoAn artificialneural network for modeling reliability availability and main-tainability of a repairable systemrdquo Reliability Engineering andSystem Safety vol 91 no 7 pp 809ndash819 2006
[7] W Jang W M Healy and S J Mirosław ldquoWireless sensornetworks as part of a web-based building environmental mon-itoring systemrdquo Automation in Construction vol 17 no 6 pp729ndash736 2008
[8] M S Pan C H Tsai and Y C Tseng ldquoEmergengy guiding andmonitoring application in indoor 3D environment by WirelessSensor Networkrdquo Internation Journal Os Sensor Netwourk vol1 pp 2ndash10 2006
[9] S Zarifzadeh A Nayyeri and N Yazdani ldquoEfficient construc-tion of network topology to conserve energy in wireless ad hocnetworksrdquo Computer Communications vol 31 no 1 pp 160ndash173 2008
[10] R Riaz A Naureen A Akram A H Akbar K Kim and HFarooq Ahmed ldquoA unified security framework with three keymanagement schemes for wireless sensor networksrdquo ComputerCommunications vol 31 no 18 pp 4269ndash4280 2008
[11] S L Goh and D P Mandic ldquoAn augmented CRTRL forcomplex-valued recurrent neural networksrdquo Neural Networksvol 20 no 10 pp 1061ndash1066 2007
[12] L A B Munoz and J J V Ramosy ldquoSimilarity-based heteroge-neous neural networksrdquo Engineering Letters vol 14 no 2 pp103ndash116 2007
[13] R Taormina K Chau and R Sethi ldquoArtificial neural networksimulation of hourly groundwater levels in a coastal aquifer sys-tem of the Venice lagoonrdquo Engineering Applications of ArtificialIntelligence vol 25 no 8 pp 1670ndash1676 2012
[14] K Leblebicioglu and U Halici ldquoInfinite dimensional radialbasis function neural networks for nonlinear transformationson function spacesrdquo Nonlinear Analysis Theory Methods andApplications vol 30 no 3 pp 1649ndash1654 1997
[15] F Fernandez-Navarro C Hervas-Martınez J Sanchez-Monedero and P A Gutierrez ldquoMELM-GRBF a modifiedversion of the extreme learning machine for generalized radialbasis function neural networksrdquo Neurocomputing vol 74 no16 pp 2502ndash2510 2011
[16] C L Wu KW Chau and Y S Li ldquoPredicting monthly stream-flow using data-drivenmodels coupledwith data-preprocessingtechniquesrdquoWater Resources Research vol 45 no 8 Article IDW08432 2009
[17] C EACheng ldquoLong-termprediction of discharges inManwanReservoir using artificial neural network modelsrdquo in Advancesin Neural NetworksmdashISNN 2005 vol 3498 of Lecture Notes inComputer Science pp 1040ndash1045 2005
[18] H H Zhang M Genton and P Liu ldquoCompactly supportedradial basis function kernelsrdquo 2004
[19] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquo inProceedings of the 1995 IEEE International Conference on NeuralNetworks pp 1942ndash1945 December 1995
[20] J Eberhart R Eberhart and Y Shi ldquoParticle swarm optimiza-tion developments applications and resourcesrdquo IEEE vol 1 pp81ndash86 2001
[21] R Malhotra and A Negi ldquoReliability modeling using ParticleSwarm optimisationrdquo International Journal of System AssuranceEngineering and Management vol 4 no 3 pp 275ndash283 2013
[22] D Gies and Y Rahmat-Samii ldquoParticle swarm optimization forreconfigurable phase-differentiated array designrdquo Microwaveand Optical Technology Letters vol 38 no 3 pp 168ndash175 2003
ISRN Artificial Intelligence 13
[23] F Valdez P Melin and C Oscar ldquoAn improved evolution-ary method with fuzzy logic for combining Particle Swarmoptimization and genetic algorithmsrdquo Applied Soft ComputingJournal vol 11 no 2 pp 2625ndash2632 2011
[24] J Zhang and K Chau ldquoMultilayer ensemble pruning vianovel multi-sub-swarm particle swarm optimizationrdquo Journalof Universal Computer Science vol 15 no 4 pp 840ndash858 2009
[25] F Valdez P Melin and O Castillo ldquoParallel Particle Swarmoptimization with parameters adaptation using fuzzy logicrdquoin Proceedings of the 11th Mexican International Conference onArtificial Intelligence (MICAI rsquo12) vol 2 pp 374ndash385 San LuisPotosi Mexico October 2012
[26] J Peng and C P Y Pan ldquoParticle swarm optimization RBFfor gas emission predictionrdquo Journal of Safety Science andTechnology vol 7 pp 77ndash85 2011
[27] G Ren Z Huang Y Cheng X Zhao and Y Zhang ldquoAnintegrated model for evacuation routing and traffic signaloptimization with background demand uncertaintyrdquo Journal ofAdvanced Transportation vol 47 pp 4ndash27 2013
[28] X Feng Z Xiao and X Cui ldquoImproved RSSI algorithm forwireless sensor networks in 3Drdquo Journal of ComputationalInformation Systems vol 7 no 16 pp 5866ndash5873 2011
[29] F Valdez and P Melin ldquoNeural network optimization witha hybrid evolutionary method that combines particle swarmand genetic algorithms with fuzzy rulesrdquo in Proceedings ofthe Annual Meeting of the North American Fuzzy InformationProcessing Society (NAFIPS rsquo08) NewYork NYUSAMay 2008
[30] F Fernandez-Navarro C Hervas-Martınez P A Gutierrez JM Pena-Barragan and F Lopez-Granados ldquoParameter estima-tion of q-Gaussian Radial Basis Functions Neural Networkswith a Hybrid Algorithm for binary classificationrdquo Neurocom-puting vol 75 pp 123ndash134 2012
[31] H Soh S Lim T Zhang et al ldquoWeighted complex networkanalysis of travel routes on the Singapore public transportationsystemrdquo Physica A vol 389 no 24 pp 5852ndash5863 2010
[32] C K S Kumar R Sukumar and M Nageswari ldquoSensorslifetime enhancement techniques in wireless sensor networksmdasha critical reviewrdquo International Journal of Computer Science andInformation Technology amp Security vol 3 pp 159ndash164 2013
[33] S Li and F Qin ldquoA dynamic neural network approach forsolving nonlinear inequalities defined on a graph and itsapplication to distributed routing-free range-free localizationof WSNsrdquo Neurocomputing vol 117 pp 72ndash80 2013
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
4 ISRN Artificial Intelligence
The mathematical objective here is to find an optimizedrouting matrix (120593119900) that has the maximum survivability Thesurvivability is defined as the entropy of a number of parallelconnections between every node to all the sink(s) The exitmatrix (120593119890) described the success rate from each node to thesink(s) 120593119890 assumes that the number of exits (119873119890) is availablewith an errormargin (120595) Inmost practical applicationsmorethan one sink is used and sink node is either through the fiberorThrough-The-Earth (TTE) link It is important to note thatin real rescue situations the software (relational databasemanagement system (RDBMS)) and hardware includingradio frequency identification (RFID) may fail as a result ofthe effect from the explosion and attack The matrix (120593119904)
was used to describe software survival rate including bugs orattacks
120593119904 = 1 minus ((1
119879 + random) (ldquoGeometricrdquo fail 119879 119879)) (8)
To obtain the final survival vector (120593119877) it was assumed thateach miner will have an RFID a vector 120593119868 was used todescribe its failure rate including risks of running out ofbattery a vector 120593119867 for the hardware failure rate for119879 = totalnumber of nodes
120593119868 = 1 minus ((1
119879 + 119903) lowast (ldquoGeometricrdquo fail 1 119879)) (9)
119903 is a random number generated from the vector 119879 and 119903
0 rarr infin 119879 + random 119879 rarr infin (1(119879 + 119903)) 1119879 rarr 01 minus (1(119879 + 119903)) 1 minus 1119879 rarr 1 is the minimum thereforefor 119879 = 119873 nodes we have 1 minus 1119873 = (119873 minus 1)119873 rarr 1 for(119873 minus 1)119873 rarr node is dead and 1 rarr node is alive
120593119867 = min (1 120593119868 [119883Ψ]) for [119883Ψ]) =120593119890
119872119901
(10)
The survival rate of each miner was (120593V) All these assump-tions happen in real life and must be considered The finalsurvival rate vector (120593119877) was calculated
120593V = 120593119904 lowast 120593119890 (11)
120593119877 = 120593V lowast 120593119904 (12)
119877 = 120593119877 (13)
3 Related Work
31 Artificial Neural Networks (ANN) Having found theoptimum set of routing table that has the highest survivalprobability of communicating with and rescuing miners it isimportant to train the neurons such that the initial error willbe minimized and more importantly have a reliable system[21]The topology of a neural network can be recurrent (withfeedback contained in the network from the output back tothe input) the feedforward (where data flow from the input tothe output units)The data processing can extend over severallayers of units but no feedback connections are presentManyresearchers have come out with neural network predictivemodels in both sigmoid and radial basis functions [22 23]
with applications such as nonlinear transformation [23]extreme learning machine predicting accuracy in gene clas-sification [24] crisp distributed support vectors regression(CDSVR) model was monthly streamflow prediction wasproposed by Valdez et al [25] and other applications includefuzzy inference systems (FIS) which have been successfullyapplied in fields such as automatic control data classificationdecision analysis and expert systems [26] among othersArtificial neural networks (ANN) are learning algorithmusedto minimize the error between the neural network outputand desired output This is important where relationshipsexist between weights within the hidden and output layersand among weights of more hidden layers In addition otherparameters including mean iteration standard variationstandard deviation and convergent time (in sections) wereevaluated The architecture of the learning algorithm andthe activation functions were included in neural networksNeurons are trained to process store recognize and retrievepatterns or database entries to solve combinatorial optimiza-tion problems Assuming that the input layer has 4 neuronsoutput layer 3 neurons and the hidden layer 6 neurons wecan evolve other parameters in the feedforward network toevolve the weight So the particles would be in a group ofweights and there would be 4 lowast 6 + 6 lowast 3 = 42 weightsThis implies that the particle consists of 42 real numbersThe range of weight can be set to [minus100 100] or a fittingrange After encoding the particles the fitness function wasthen determined The goodness of the fit was diagnosedusing mean squared error (MSE) as against mean cubicerror (MCE) and the mean absolute error (MAE) The meancubic error will allow for fast convergence and that willgross over accuracy making the process unstable while themean absolute error is stable but converges slowly A midwaybetween the two is the MSE and is given as
MSE =1
119899119904
119899
sum
119895=1
119904
sum
119894=1
(119884119895119894
minus 119910119895119894
)2
(14)
where 119899 is number of samples 119904 is the number of neurons atoutput layer 119884
119894119895
is the ideal value of 119894th sample at 119895th outputand 119884
119894119895
is the actual value of 119894th sample at 119895th output
32 The Gaussian and Zhang Models The standard or directapproach to interpolation at locations 119909
1
119909119873
sub 119877119889
without the first term using Gaussian kernel is given as
120601 (119903) = 119890minus(120576119903)
2
(15)
Zhang has [27] tried tomodify theGaussianmodel as followsa function 120595 [0infin) rarr R such that 119896(119909 1199091015840) = 120593(119909 minus 119909
1015840
)where 119909 1199091015840 isin 120594 and 119909 sdot 119909
1015840 denotes the Euclidean norm with119896(119909 119909
1015840
) = exp(minus119909 minus 1199091015840
2
1205752
) as an example of the RBFkernels The global support for RBF radials or kernels hasresulted in dense Grammatrices that can affect large datasetsand therefore construct the following two equations119896119862V(119909 119909
1015840
) = 120601119862V(119909 minus 119909
1015840
)119896(119909 1199091015840
) and 120601119862V(119909 minus 119909
1015840
) =
[(1 minus 119909 minus 1199091015840
119862)(sdot)+
]V where119862 gt 0 V ge (119889+1)2 and (sdot)
+
isthe positive part The function 120601C(sdot) is a sparsifying operator
ISRN Artificial Intelligence 5
which thresholds all the entries satisfying (119909 minus 1199091015840
) ge 119862
to zeros in the Gram matrix The new kernel resulting fromthis construction preserves positive definiteness This meansthat given any pair of inputs 119909 and 119909
1015840 where 119909 = 1199091015840 the
shrinkage (the smaller 119862) is imposed on the function value119896(119909 119909
1015840
) the result is that the Gram matrices 119870 and 119870119862V can
be either very similar or quite different depending on thechoice of 119862 However Zhang also ended up with an extrapower parameter
4 The Proposed Hybrid
The previous work looked at the Gaussian model and par-alyzed the computational power parameter The result wascompact radial basis function (CRBF) This was used to runboth SBF and RBF and the results compared The scalabilityand processing efficiency were also analyzed
The novelty of this algorithm the weighed nonlinearhybrid was to find several hybrids and the best for rescueoperationsThe cosine and sine functionswere used to reducehigh level nonlinear and increase small level nonlinearBoth the previous and the current algorithm used the samepreliminary considerations but the results is slightly differentbecause data are random
The sigmoid basis function was given as
log sig (119877) 119877 = 119882 sdot 119875 + 119861
log sig (119882 sdot 119875 + 119861) =1
1 + 119890minus(119882sdot119875+119861)
(16)
Neuron function 119878 (sigmoid) is log sig 119882 is weight matrix119875 is input vector and 119861 is threshold We therefore proposeda compact radial basis function based on the Gaussian radialbasis function and Helensrsquo [28] definition expressed as
[exp (minus 119877)] 119877 = 119882 sdot 119875 + 119861
(exp (minusabs ((119882 sdot 119875 + 119861)))
120601 (out) = [exp (minus 119877)]
(17)
119882 is weight matrix 119875 is input vector and 119861 is threshold Thefocus was to improve on the radial basis function for themineapplication
An optimized vector 119877 was generated as the optimumset of transmission routing table that has the highest survivalprobability for data transmission (13)
119875119873 =
119873
sum
119894=1
119877119894
119878119894
+
2
sum
119894=1
119878119894
+
2
sum
119894=1
119878119894
119878119894
+ 1
for 1 le 119894 le 119878 1 le 119896 le 119877
(18)
119877 1198781
1198782
are number of neurons at input hidden and outputlayers respectively
119875119873 is the position of the 119899th particle
119882119894
= 119878119894
1198771198822
= 119878119894
119877 119882119898
= 119878119898
119877
119875119894
= [119877 (119894 minus 119895) + 119870] = 119882119894
(119894 119896)
(19)
There are two thresholds (1198781
rArr 1198611
)Hidden 119861
1
(119894 119895) = 119875(1198771198781
+ 1198781
1198782
) rArr S1
Output 1198612
(119894 119895) = 119875(1198771198781
+ 1198781
1198782
) rArr 1198782
From (16) (17) and (18) the nonlinear weight of [plusmn cos(119877)] or[plusmn sin(119877)] was imposed on the CRBF before being combinedwith the SBF as (HSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sinHSCR-BF
+ sin)The cosine weight was used to keep high levelnonlinear for small input value and reduce the nonlinear forlarge input values while sine weight was used to keep highlevel nonlinear for large input value and reduce the nonlinearfor small input values
HSCR-BFminus cos
119884 (out) = log sig (119877) + [exp (minus 119877)] sdotlowast
[minus cos (119877)]
HSCR-BF+ cos
119884 (out) = log sig (119877) + [exp (minus 119877)] sdotlowast
[+ cos (119877)]
HSCR-BFminus sin
119884 (out) = log sig (119877) + [exp (minus 119877)] sdotlowast
[minus sin (119877)]
HSCR-BF+ sin
119884 (out) = log sig (119877) + [exp (minus 119877)] sdotlowast
[+ sin (119877)]
(20)
Thenonlinear weight of [+ cos(119877)]was imposed on theGRBFbefore being combined with the SBF
HSGR-YBF+ cos
119884 (out) = log sig (119877) + [exp (minus1198772
)] sdotlowast
[+ cos (119877)]
(21)
5 Particle Swarm Optimization
Particle swarm optimization (PSO) an evolutionary algo-rithm is a population based stochastic optimization tech-nique The idea was conceived by an American researcherand social psychologist James Kennedy in the 1950s Thetheory is inspired by social behavior of bird flocking or fishschooling The method falls within the category of swarmintelligence methods for solving global optimization prob-lems Literature has shown that the PSO is an effectivealternative to established evolutionary algorithms (GA) It isalso established that PSO is easily applicable to real worldcomplex problems with discrete continuous and nonlineardesign parameters and retains the conceptual simplicity ofGA [29 30] Each particle within the swarm is given an initialrandom position and an initial speed of propagation Theposition of the particle represents a solution to the problemas described in a matrix 120591 where 119872 and 119873 represent thenumber of particles in the simulation and the number ofdimensions of the problem respectively [31 32] A randomposition representing a possible solution to the problem withan initial associated velocity representing a function of thedistance from the particlersquos current position to the previous
6 ISRN Artificial Intelligence
position of good fitness value was given A velocity matrix119881el with the same dimensions as matrix 120591
119909
described this
120591119909
= (
12059111
12059112
1205911119873
12059121
12059122
1205912119873
1205911198981
1205911198982
120591119872119873
)
119881el = (
V11 V12 V1119873
V21 V22 V2119873
V1198981 V1198982 V119872119873
)
(22)
The best feasible alternative is the genetic algorithmHoweverthis is generally slow as compared to the PSO and thereforecomputationally expensive hence the PSO
While moving in the search space particles commit tomemorize the position of the best solution they have foundAt each iteration of the algorithm each particle moves witha speed that is a weighed sum of three components the oldspeed a speed component that drives the particles towardsthe location in the search space where it previously foundthe best solution so far and a speed component that drivesthe particle towards the location in the search space wherethe neighbor particles found the best solution so far [7] Thepersonal best position can be represented by an119873times119873matrix120588best and the global best position is an 119873-dimensional vector119866best
120588best = (
12058811
12058812
1205881119873
12058821
12058822
1205882119873
1205881198981
1205881198982
120588119872119873
)
119866best = (119892best1
119892best12
119892best119873
)
(23)
All particles move towards the personal and the globalbest with 120591 120588best 119881el and 119892best containing all the requiredinformation by the particle swarm algorithmThese matricesare updated on each successive iterations
119881119898119899
= 119881119898119899
+ 1205741198881
1205781199031
(119901best119898119899
minus 119883119898119898
)
+ 1205741198882
1205781199032
(119892best119899
minus 119883119898119899
)
119883119898119899
= 119881119898119899
(24)
1205741198881
and 1205741198882
are constants set to 13 and 2 respectively and1205781199031
1205781199032
are random numbers
51 AdaptiveMutation according toThreshold Particle swarmoptimization has been effective in training neural networkssuch as a Parallel Particle Swarm optimization (PPSO)methodwith dynamic parameter adaptation used to optimizecomplex mathematical functions and improved evolutionarymethod in fuzzy logic [30 31 33] To prevent particlesfrom not converging or converging at local minimum an
Input nodes64748
64748
64748
Hidden nodesOutput nodes
S11R
S12R
SR
120573 ge 07
S1mR
120573 ge 07 le 09
120573 ge 07 le 09
Figure 1 The structure AMPSO for CRBF GRBF and SBF transferfunctions
1 2 3 4 5 6 7 8 9 100
02040608
11214
The impact of the accident
Link
avai
labi
lity
The curve of transmission link after accident
The higher the impact the lower the link availability
Region 3Links are completely dead
Region 1Links are not
Region 2Indicate probability of link being able to transmit data
affected
Figure 2 Impact of explosionaccident on transmission link
adaptivemutation according to thresholdwas introduced Analternative to this is the use of the nonadaptive mutation andthis could converge to the local minimum or fail to convergeat all The advantage of the nonadaptation is simple and fastbut may not provide the needed results
Particles positions were updated with new value onlywhen the new value is greater than the previous value 20 ofparticles of those obtaining lower values weremade tomutatefor faster convergence and the structure of adaptive mutationPSO (AMPSO) with threshold can be found in [1] The inputlayer takes the final survival vector (13) with a number ofhidden layers and an output layer The feedforward neuralnetwork was used The structure of Adaptive Mutation PSO(AMPSO) with threshold Neural Network was used Figure 1[32]
6 Results and Discussion
61 Generating the Routing Path Elements 0 1 and 2 in 120593119909
imply that the link(s) were not affected and elements 3 45 and 6 represent a probability for the links being able totransmit data while figures from 7 means the link is totallydead Region 1 that indicates that links are not affected region
ISRN Artificial Intelligence 7
Table 2 Optimum set of routing table redundancy with the highest survival probability (6 cases)
Vectors for each case Rock hardness cases1198771 05409 04281 08834 0585 08134 05163 071198772 04526 05841 04547 08997 08436 02696 081198773 02931 05899 05903 08309 06971 02898 0831198774 04851 06085 06913 04002 05026 0 0861198775 04917 05857 07805 06037 05182 06472 08751198776 02902 07253 0458 05762 01397 0 09
2 gives the probability of link available and the last regionindicates the link is completely down (Figure 2)
Thematrices120593119894 120593119896 120593119903 120593119909 120593ℎ for 120591 = 119899 with dimensionsof 119871 = 3 119877 = 2 and119862 = 1 for depth (level) row (length) andwidth (column) respectively were generated as follows
119878eq = (
(
1 1 1
1 2 1
2 2 1
2 1 1
3 1 1
3 2 1
)
)
119894120593 =
[[[[[[[
[
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
]]]]]]]
]
120593119896 = (
(
1 1 1 1 0 0
1 1 1 0 0 0
0 1 1 1 0 0
0 0 1 1 1 0
0 0 0 1 1 1
0 0 0 1 1 1
)
)
1 = signal points
120593119903 = (
(
1 1 1 0 0 0
1 1 1 0 0 0
0 1 1 1 0 0
0 0 1 1 1 0
0 0 0 1 1 1
0 0 0 1 1 1
)
)
1 = transmission path
120593119909 = (
(
0 1 2 2 0 0
1 5 1 2 0 0
0 8 1 5 9 1
1 1 5 1 1 17
1 1 4 11 0 2
4 4 19 5 3 0
)
)
explosion sizes
120593119891 = (
(
1 1 1 1 1 1
1 6 1 1 1 1
1 0 1 6 0 1
1 1 6 1 1 0
1 1 75 0 1 1
75 75 0 6 1 1
)
)
(25)
120593ℎ = (
(
1 1 1 0 0 0
1 6 1 0 0 0
0 0 1 6 0 0
0 0 6 1 1 0
0 0 0 0 1 1
0 0 0 6 1 1
)
)
120593119900 = (
(
8 8 8 0 0 0
8 48 8 0 0 0
0 0 8 48 0 0
0 0 48 8 8 0
0 0 0 0 8 8
0 0 0 48 8 8
)
)
120593119890 = (
(
144 144 144 0 0 0
144 864 144 0 0 0
0 0 144 864 0 0
0 0 864 144 144 0
0 0 0 0 144 144
0 0 0 864 144 144
)
)
1 or fractional figures are transmission path
(26)
Element ldquo0rdquo on 120593119891 depicts a connection lost while ldquo1rdquo meansthat the connection will never go down and represent theconnection to the fixed sink node(s) along the edge or theemergency connection to the mobile data mule(s)
120593s = (
(
8333 8889 9091 8333 8571 8333
8889 8333 8750 8571 8750 8333
8571 9167 8750 9091 8889 8333
8750 8571 8571 8571 8571 8333
8333 8333 8571 8333 8889 9000
9167 8333 9091 8333 8889 8571
)
)
(27)
The survival rate of RFID hardware and each minervectors was displayed as
120593119868 = 8571 8750 8571 8889 8571 8333
120593119867 = 6236 4976 1 6851 9286 6086
120593V = 5409 4281 8834 5850 8134 5163
(28)
Optimization was done numerically usingMatlab simulationtool to find the optimum set of routing tables throughparticle swarm search for rescue operation as discussed inthe preliminaries From (13) the final survival vector fordimensions of 119871 = 3 119877 = 2 and 119862 = 1 will be an average ofthe 6 vectors as in Table 2 Total cases 119877119899 = 120591 = 119879times119879 (which
8 ISRN Artificial Intelligence
Table 3 Scalability of model to survival probability range robot location and rock type
Location Mica Coal Granite Feldspar Quartz MineralCompact radial basis function with negative nonlinear weight (HSCR-BF
minus cos)(10 6 5) 08769ndash10000 09070ndash09927 08402ndash09943 08023ndash09778 07386ndash09376 07345ndash08740(10 5 4) 09024ndash09696 09032ndash09505 08836ndash09430 08633ndash09527 07944ndash09465 07263ndash08720(6 5 4) 08906ndash09877 08827ndash09845 08384ndash09675 08627ndash09820 07911ndash09142 06392ndash07840(3 1 10) 07302ndash08481 06588ndash07816 06341ndash07473 06503ndash07249 05930ndash07156 04445ndash05621
Gaussian radial basis function with positive nonlinear weight (HSGR-BF+ cos)
(10 6 5) 08732ndash09963 08526ndash09987 08492ndash09950 07867ndash09568 07404ndash09025 06606ndash08315(10 5 4) 08843ndash09593 08843ndash09853 08444ndash09753 07746ndash09828 07304ndash09595 06678ndash08817(6 5 4) 08785ndash09983 08902ndash09996 08200ndash09997 07927ndash09446 07065ndash08565 06572ndash08063(3 1 10) 07608ndash08552 06600ndash08325 06090ndash07371 05482ndash06249 05026ndash06056 04856ndash05453
10 20 30 40 50 60 70 80 90 1000
02
04
06
X 94Y 0009904
Number of iterations
Fitn
ess v
alue
PSO optimized error for hybrid SBF-CRBF with negative nonlinear weight
15 2 25 3 35 4 45 5 55 604
06
08
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(HSCRminuscos )
(a) HSCR-BFminus cos
50 100 150 200 2500
01
02
03
X 249Y 001151
PSO optimized error with nonlinear weight [minussine(R)]
Number of iterations
Fitn
ess v
alue
1 2 3 4 5 6
07
08
09
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(b) HSCR-BFminus sin
Figure 3 Hybrid of SBF and CRBF (HSCR-BF) with (a) HSCR-BFminus cos and (b) HSCR-BF
minus sin nonlinear weight
represent the total nodes deployed) for this scenario Rocksoftness or hardness takes the values of 120573 and each matrix isgenerated 120591 timecases and the final routing vector will be anaverage of the 119899 vectors 120593119877 = Average (1198771 1198776) [Table 2]which becomes the input for the PSO training
The final survival rate is 119877 = (05409 04281 08834
0585 08130 05163) from (13)
62 The Final Survival Vector The 119877 = (05409 04281
08834 0585 08130 05163) from (13) is the maximumsurvival probability for a total of 6 nodes deployed Itdescribes the success rate from each node to the sink(s) Inmost practical applications more than one sink is used andsink node is either through the fiber or TTE connection Thesize of the vector depends on the dimensions of the fieldThe elements 119877 = (05409 04281 08834 0585 08130
05163) represent the probability of 54 43 88 5981 and 52 success of each node transmitting data to andfrom its source or destination It assists decision makers as towhether to send data through one ormore nodes or send each
message twice The total nodes used for the simulation were300 with undergroundmine dimensions of 119871 = 10119877 = 6 and119862 = 5 for depth (level) row (length) and width (column)respectively with ldquopmrdquo a sensor apart pm = 100119872
120588
= 4 and119873119890 = 2 The PSO training used swarm size of 20 maximumposition was set to 100 max velocity = 1 and maximumnumber of iteration = 250 The thresholds 120582119871 and 120582119867 were3 and 6 respectively 120591 = 6 cases (each case representa rock hardness case and also represent the total neuronsused) thus each matrix and vectors were run 6 times beforeneural training The survival probability (bottom) indicatesthat the model survived between 90 and 100 where rockcases were relatively soft (ge07) The survival probabilitydeclines as the rock becomes harder and approaches 08 Atthe hardest rock of 09 the survival probability fell between72 and 84 for the entire hybrids In view of this the APheads had to be deployed at a location where the rock isrelatively soft for maximum signal strength Each computersimulation incorporated all the 6 different cases of rockhardnesssoftness to produce the matrices Detailed locationanalysis or the scalability of the model in relation to survival
ISRN Artificial Intelligence 9
Table 4 Hybrid of SBF and CRBF with negative nonlinear weight on CRBF (HSCR-BFminus cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 231020 343498 0382 91650 00985AVG 23102 34350 0038 9165 00098MinMax 29540 45314 0024 80000 00043
Table 5 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BFminus sin)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 307320 488381 0376 114300 0135AVG 30732 48838 0038 11430 0013MinMax 30740 56926 0007 61500 0021
probability range robot location and different rock types isrecorded in Table 3 Figures 4ndash6 represent different scenariosof SBF and CRBF hybrids with the [plusmn cos(119877)] and [plusmn sin(119877)]
nonlinear weights while Figure 7 represents the SBF andGaussian hybrids for [+ cos(119877)] nonlinear weight The tophalf of each figure indicates the optimized error or the finalerror after the neural training and the bottom half revealsmodel survival probability
Figure 3 discusses the SBF and CRBF hybrids (HSCR-BF) with nonlinear weight Training in negative nonlinearweights (HSCR-BF
minus cos HSCR-BFminus sin) responded well The
HSCR-BFminus cos had a steady and compact routing path which
was consistent through all rock layers with initial survivalprobability between 877 and 100 in soft layers decliningto 735ndash874 at harder rock layers The HSCR-BF
minus sinperformed quite well but was more dispersed (879 to 985at the soft rock and 665 to 818 at the hard rock)
In addition the various parameters mean iterationstandard variance standard deviation and the convergenttime of HSCR-BF
minus cos were optimised This is demonstratedby finding the relationship between the maximum and theminimum values of the parameters and comparing with theaverage figures and the closer the difference is to the averagethe more consistent the data are in the dataset (Table 4) thusHSCR-BF
minus cos provided the best results among the proposedhybrids The HSCR-BF
minus sin on the RBF yielded strength inmean iteration and standard variance with error of 0013(Table 5)
In Figure 4 training with nonlinear weight of positivecosine and sine of HSCR-BF (HSCR-BF
+ cos and HSCR-BF+ sin) had some similarities The two models streamed
well at the initial stages from 896 to 992 and 908to 990 at soft layers of the rock respectively Howeverat the last stages the probability of transmitting effectivelybecame marginal from 795 to 852 and 821 to 854respectively Much as this hybrid could perform well in areaswhere routing conditions aremuch better in less dense rescuesituations this could hamper rescue mission in both casesdue to battery drain or collision from traffic conjunctionThedata formean iteration and convergent time forHSCR-BF
+ cos(Table 6) and mean Iteration standard variance the final
error HSCR-BF+ sin (Table 7) were consistent with the nega-
tive sine having better results than the positive cosineThe survival probability of the positive nonlinear weight
Gaussian hybrids (HSGR-BF+ cos) in Figure 5 is more com-
pact both at the initial (ie 873ndash996) and later (661ndash832) stages declining to harder rock layers with an errorof 001173 (Table 8)
At the initial stage particles are sensitive to inputs asthey moved quickly in the search space towards the targetand while particles peaked closer to the target it became lesssensitive to the input and began to align (Figure 6) At thelater stage more RBF are used to keep the error at minimumfor accuracy It was also noticed that instead of acceler-ating higher and becoming more sensitive to the inputswhile particles were far from the target the hybrids withlowast
[minus sin(119877)] and lowast[ + sin(119877)] were less sensitive to the inputsas lowast[minus sin(119877)] descended from [119909 = minus1 119910 = 05785] to[119909 = 1 119910 = 04215] while lowast[+ sin(119877)] descended from [119909 =
minus3 119910 = 00404] to [119909 = minus1 119910 = minus004062] before becomingconscious of the inputs In addition the hybrids with thelowast
[+ cos(119877)] lagged slightly between [119909 = minus4 119910 = 0006014]
and [119909 = minus2 119910 = 006288] before becoming sensitive toinputs However they all lined up for accuracy in terms of theminimised error
The nonlinearity of the hybrids is presented (Figure 7)the negative cosinesine weight is used to reduce nonlinearfor small input values and this lies between minus1 lt 119909 lt +1 andthe positive cosinesine weight is used to keep high nonlinearfor large inputs for the remaining region
63 CPU Time Efficiency The relationship between varioushybrids with respect to the central processing time (CPU)was profiled for different runs (Table 9) and expressed in asixth-order polynomial given as 119884
119903
= 1205737
1198836
1
+ 1205736
1198835
1
+ 1205735
1198834
1
+
1205734
1198833
1
+ 1205733
1198832
1
+ 1205732
1198831
+ 1205730
where 1198831
is time (seconds) and 120573
is the coefficient of the polynomial (Figure 8) The proposedhybrid has better usage of CPU time with 119877
2
= 09160followed byHSCR-BF
minus sin andHSCR-BF+ cos with119877
2 of 07551and 07244 respectively (Table 10) Applying the proposedalgorithm into Gaussian the CPU usage had 119877
2 of 07345
10 ISRN Artificial Intelligence
0 50 100 150 200
002
004
006
008
X 249Y 001103
PSO optimized error for hybrid SBF-CRBF with positive nonlinear weight
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 607
08
09
1
11 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(HSCR+cos )
(a) HSCR-BF+ cos
0 50 100 150 200
005
01
015
02
X 249Y 001247
PSO optimized error with nonlinear weight [+sine(R)]
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 6
070809
111 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(b) HSCR-BF+ sin
Figure 4 Hybrid of SBF and CRBF (HSCR-BF) with (a) HSCR-BF+ cos and (b) HSCR-BF
+ sin nonlinear weight
Table 6 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BF+ cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 350780 522257 0311 115650 0121AVG 35078 52226 0031 11565 0012MaxMin 54840 63413 0009 93500 0006
and indeedmarginally outperformedHSCR-BF+ cos Detailed
work on SBF CRBF and GRBF with regard to scalabilitymemory usage and the central processing time has beencarried by the authors [1]
7 Conclusion
In summary we made the following contributions First weused the mix of SBF and CRBF to present several hybridswith different nonlinear weights of cosine and sine func-tions on compact radial basis function Next we showedthe performance of the proposed nonlinear weight hybridsHSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sin and HSCR-BF+ sin optimised all the parameters with minimised error
of 00098 0012 0013 and 0013 respectively compared to00117 for Gaussian HSCR-BF
+ cos The analyzed CPU usagewith corresponding 119877
2 values of 09160 075 072440 06731and 07345 for HSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sinHSCR-BF
+ sin and HSCR-BF+ cos respectively demonstrated
that the algorithm is scalable There exist some evacuationmodels that is Goh and Mandic [11] which offered a choicefor travelers and several schemes for the decisionmakersmaynot be applicable in emergency undergroundmine situationsThe proposed nonlinear hybrid algorithm with particle swarmoptimisation has better capability of approximation to underly-ing functions with a fast learning speed and high robustness and
is competitive and more computationally efficient to Gaussianwith the same nonlinear weight The algorithm is new andthat makes it difficult to identify limitations and we intend toinvestigate other hybrids and comparewith genetic algorithm(GA) in the future
Nomenclature
119873120575min 119873120575max Minimum and maximum signalreach
HSCR-BFHSGR-BF Hybrid of SBF with CRBFHybrid ofSBF with GRBF
HSCR-BFminus cos Hybrid of SBF and CRBF minuscos
nonlinear weightHSCR-BF
+ cos Hybrid of SBF and CRBF +cosnonlinear weight
HSCR-BFminus sin Hybrid of SBF and CRBF minussine
nonlinear weightHSCR-BF
+ sin Hybrid of SBF and CRBF +sinenonlinear weight
HSGR-BF+ cos Hybrid of SBF and GRBF +cos
nonlinear weight
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
ISRN Artificial Intelligence 11
Table 7 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BF+ sin)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 260360 407109 03457 99800 0129AVG 26036 40711 00346 9980 0013MinMax 27840 56040 0008 106500 0014
Table 8 Hybrid of SBF and GRBF with nonlinear weight on GRBF (HSGR-BF+ cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 219840 349010 0439 112100 01173AVG 21984 34901 0044 11210 00117MinMax 27000 46580 0008 80000 00165
0 50 100 1500
005
01
015
02
X 144Y 0009595
GRBF-PSO optimized error with nonlinear weight [+cos(R)]
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 604
06
08
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
Figure 5 Hybrid of SBF and GRBF with nonlinear weight on GRBF(HSCR-BF
+ cos)
0 2 4 6 8 10
002040608
1121416
Weighed position of particles
Min
imise
d er
ror
Optimised error in nonlinear weighed SBF-CRBF hybridsParticles
Less sensitive to inputsmuch closer to target
more accurate
Less sensitive to inputsmuch closer more accurate
Neurons are more sensitive to inputsfast and far from target
minus02minus2minus4minus6minus8minus10
[x = minus1 y = 05785]
[x = 1 y = 04215]
[minus1 004062]
[x = minus2 y = 006288]
[x = minus4 y = 0006014]
[minus3 00404]
Hybrid+cosHybrid+sin
HybridminuscosHybridminussin
Figure 6 Particles weighed position
Acknowledgments
The authors would like to appreciate the immense contribu-tion of the mining companies where the study was under-taken The authors are grateful to these individuals for
0 1 2 3 4 5
0
05
1
15
2
X 37Y 0989
Particles position
Opt
imise
d er
ror
Nonlinear weight hybrids
Keep large inputvalues more
nonlinear
Keep small inputvalues lessnonlinear
Large input values
Small input values
Using positive cosine and sine
Using negative cosine and sineminus05
minus5 minus4 minus3 minus2 minus1
HSCR-BF+cosHSGR-BF+cos
HSCR-BF+sin
HSCR-BFminuscos
HSCR-BFminuscos
HSCR-BFminussin
HSCR-BF+cos HSCR-BF+sin HSCR-BFminussinand HSGR-BF+cos
Figure 7 Nonlinear weighed curves for optimized error withoutGaussian negative cosine curve
HSGR-BF+cos
HSGR-BF+cos
HSCR-BF+sinHSCR-BFminuscosHSCR-BFminussin
00020000400006000080000
100000120000
0 2 4 6 8 10 12
CPU
tim
e
Time (s)
Nonlinear hybrids
Figure 8 The trend of hybrids CPU time curves using sixth-OrderPolynomial
their immeasurable contributions to this work Fred Atta-kuma Patrick Addai Isaac Owusu Kwankye Thomas KwawAnnan Willet Agongo Nathaniel Quansah Francis OwusuMensah Clement Owusu-Asamoah Joseph Adu-MensahShadrack Aidoo Martin Anokye F T Oduro Ernest Ekow
12 ISRN Artificial Intelligence
Table 9 CPU time usage efficiency for the hybrids
Runs HSCR-BFminus cos HSCR-BF
minus sin HSCR-BF+ cos HSCR-BF
+ sin HSGR-BF+ cos
Total 255889 254350 422344 643962 438133Average 25589 25435 42234 64396 43813
Table 10 Analysis of CPU time on the various hybrids
1205736
1205735
1205734
1205733
1205732
1205731
1205730
1198772
HSCR-BFminus cos 00159119909 minus02575119909 minus0759119909 33494119909 minus19901119909 40345119909 25322 09160
HSCR-BF+ cos 00903119909 minus23158119909 19932119909 minus58027119909 minus31337119909 32808119909 14747 07244
HSCR-BFminus sin minus0033119909 11354119909 minus15811119909 11368119909 minus4337119909 78008119909 minus20794 07551
HSCR-BF+ sin 02571119909 minus14263119909 17173119909 minus10069119909 2993119909 minus42374119909 27879 06731
HSGR-BF+ cos minus00437119909 10911119909 minus82923119909 66218119909 15934119909 minus53939119909 8191 07345
Abano E Omari-Siaw and Qiang Jia This work was sup-ported by the National Natural Science Foundation of China(no 71271103) and by the Six Talents Peak Foundation ofJiangsu Province
References
[1] M O Ansong H X Yao and J S Huang ldquoRadial and sigmoidbasis functions neural networks in wireless sensor routingtopology control in underground mine rescue operation basedon particle swarm optimizationrdquo International Journal of Dis-tributed Sensor Networks-Hindawi vol 2013 Article ID 37693114 pages 2013
[2] R Neruda and P Kudova ldquoLearning methods for radial basisfunction networksrdquo Future Generation Computer Systems vol21 no 7 pp 1131ndash1142 2005
[3] C T Cheng J Y Lin and Y G C Sun ldquoLong-term predictionof discharges in Manwan hydropower using adaptive-network-based fuzzy inference systems modelsrdquo in Advances in NaturalComputation vol 3498 of Lecture Notes in Computer Sciencepp 1152ndash1161 2005
[4] D Broomhead and D Lowe ldquoMultivariate functional interpo-lation and adaptive networksrdquo Complex System vol 2 pp 321ndash355 1988
[5] Y Chen C Chuah and Q Zhao ldquoNetwork configuration foroptimal utilization efficiency of wireless sensor networksrdquo AdHoc Networks vol 6 no 1 pp 92ndash107 2008
[6] P S Rajpal K S Shishodia and G S Sekhon ldquoAn artificialneural network for modeling reliability availability and main-tainability of a repairable systemrdquo Reliability Engineering andSystem Safety vol 91 no 7 pp 809ndash819 2006
[7] W Jang W M Healy and S J Mirosław ldquoWireless sensornetworks as part of a web-based building environmental mon-itoring systemrdquo Automation in Construction vol 17 no 6 pp729ndash736 2008
[8] M S Pan C H Tsai and Y C Tseng ldquoEmergengy guiding andmonitoring application in indoor 3D environment by WirelessSensor Networkrdquo Internation Journal Os Sensor Netwourk vol1 pp 2ndash10 2006
[9] S Zarifzadeh A Nayyeri and N Yazdani ldquoEfficient construc-tion of network topology to conserve energy in wireless ad hocnetworksrdquo Computer Communications vol 31 no 1 pp 160ndash173 2008
[10] R Riaz A Naureen A Akram A H Akbar K Kim and HFarooq Ahmed ldquoA unified security framework with three keymanagement schemes for wireless sensor networksrdquo ComputerCommunications vol 31 no 18 pp 4269ndash4280 2008
[11] S L Goh and D P Mandic ldquoAn augmented CRTRL forcomplex-valued recurrent neural networksrdquo Neural Networksvol 20 no 10 pp 1061ndash1066 2007
[12] L A B Munoz and J J V Ramosy ldquoSimilarity-based heteroge-neous neural networksrdquo Engineering Letters vol 14 no 2 pp103ndash116 2007
[13] R Taormina K Chau and R Sethi ldquoArtificial neural networksimulation of hourly groundwater levels in a coastal aquifer sys-tem of the Venice lagoonrdquo Engineering Applications of ArtificialIntelligence vol 25 no 8 pp 1670ndash1676 2012
[14] K Leblebicioglu and U Halici ldquoInfinite dimensional radialbasis function neural networks for nonlinear transformationson function spacesrdquo Nonlinear Analysis Theory Methods andApplications vol 30 no 3 pp 1649ndash1654 1997
[15] F Fernandez-Navarro C Hervas-Martınez J Sanchez-Monedero and P A Gutierrez ldquoMELM-GRBF a modifiedversion of the extreme learning machine for generalized radialbasis function neural networksrdquo Neurocomputing vol 74 no16 pp 2502ndash2510 2011
[16] C L Wu KW Chau and Y S Li ldquoPredicting monthly stream-flow using data-drivenmodels coupledwith data-preprocessingtechniquesrdquoWater Resources Research vol 45 no 8 Article IDW08432 2009
[17] C EACheng ldquoLong-termprediction of discharges inManwanReservoir using artificial neural network modelsrdquo in Advancesin Neural NetworksmdashISNN 2005 vol 3498 of Lecture Notes inComputer Science pp 1040ndash1045 2005
[18] H H Zhang M Genton and P Liu ldquoCompactly supportedradial basis function kernelsrdquo 2004
[19] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquo inProceedings of the 1995 IEEE International Conference on NeuralNetworks pp 1942ndash1945 December 1995
[20] J Eberhart R Eberhart and Y Shi ldquoParticle swarm optimiza-tion developments applications and resourcesrdquo IEEE vol 1 pp81ndash86 2001
[21] R Malhotra and A Negi ldquoReliability modeling using ParticleSwarm optimisationrdquo International Journal of System AssuranceEngineering and Management vol 4 no 3 pp 275ndash283 2013
[22] D Gies and Y Rahmat-Samii ldquoParticle swarm optimization forreconfigurable phase-differentiated array designrdquo Microwaveand Optical Technology Letters vol 38 no 3 pp 168ndash175 2003
ISRN Artificial Intelligence 13
[23] F Valdez P Melin and C Oscar ldquoAn improved evolution-ary method with fuzzy logic for combining Particle Swarmoptimization and genetic algorithmsrdquo Applied Soft ComputingJournal vol 11 no 2 pp 2625ndash2632 2011
[24] J Zhang and K Chau ldquoMultilayer ensemble pruning vianovel multi-sub-swarm particle swarm optimizationrdquo Journalof Universal Computer Science vol 15 no 4 pp 840ndash858 2009
[25] F Valdez P Melin and O Castillo ldquoParallel Particle Swarmoptimization with parameters adaptation using fuzzy logicrdquoin Proceedings of the 11th Mexican International Conference onArtificial Intelligence (MICAI rsquo12) vol 2 pp 374ndash385 San LuisPotosi Mexico October 2012
[26] J Peng and C P Y Pan ldquoParticle swarm optimization RBFfor gas emission predictionrdquo Journal of Safety Science andTechnology vol 7 pp 77ndash85 2011
[27] G Ren Z Huang Y Cheng X Zhao and Y Zhang ldquoAnintegrated model for evacuation routing and traffic signaloptimization with background demand uncertaintyrdquo Journal ofAdvanced Transportation vol 47 pp 4ndash27 2013
[28] X Feng Z Xiao and X Cui ldquoImproved RSSI algorithm forwireless sensor networks in 3Drdquo Journal of ComputationalInformation Systems vol 7 no 16 pp 5866ndash5873 2011
[29] F Valdez and P Melin ldquoNeural network optimization witha hybrid evolutionary method that combines particle swarmand genetic algorithms with fuzzy rulesrdquo in Proceedings ofthe Annual Meeting of the North American Fuzzy InformationProcessing Society (NAFIPS rsquo08) NewYork NYUSAMay 2008
[30] F Fernandez-Navarro C Hervas-Martınez P A Gutierrez JM Pena-Barragan and F Lopez-Granados ldquoParameter estima-tion of q-Gaussian Radial Basis Functions Neural Networkswith a Hybrid Algorithm for binary classificationrdquo Neurocom-puting vol 75 pp 123ndash134 2012
[31] H Soh S Lim T Zhang et al ldquoWeighted complex networkanalysis of travel routes on the Singapore public transportationsystemrdquo Physica A vol 389 no 24 pp 5852ndash5863 2010
[32] C K S Kumar R Sukumar and M Nageswari ldquoSensorslifetime enhancement techniques in wireless sensor networksmdasha critical reviewrdquo International Journal of Computer Science andInformation Technology amp Security vol 3 pp 159ndash164 2013
[33] S Li and F Qin ldquoA dynamic neural network approach forsolving nonlinear inequalities defined on a graph and itsapplication to distributed routing-free range-free localizationof WSNsrdquo Neurocomputing vol 117 pp 72ndash80 2013
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ISRN Artificial Intelligence 5
which thresholds all the entries satisfying (119909 minus 1199091015840
) ge 119862
to zeros in the Gram matrix The new kernel resulting fromthis construction preserves positive definiteness This meansthat given any pair of inputs 119909 and 119909
1015840 where 119909 = 1199091015840 the
shrinkage (the smaller 119862) is imposed on the function value119896(119909 119909
1015840
) the result is that the Gram matrices 119870 and 119870119862V can
be either very similar or quite different depending on thechoice of 119862 However Zhang also ended up with an extrapower parameter
4 The Proposed Hybrid
The previous work looked at the Gaussian model and par-alyzed the computational power parameter The result wascompact radial basis function (CRBF) This was used to runboth SBF and RBF and the results compared The scalabilityand processing efficiency were also analyzed
The novelty of this algorithm the weighed nonlinearhybrid was to find several hybrids and the best for rescueoperationsThe cosine and sine functionswere used to reducehigh level nonlinear and increase small level nonlinearBoth the previous and the current algorithm used the samepreliminary considerations but the results is slightly differentbecause data are random
The sigmoid basis function was given as
log sig (119877) 119877 = 119882 sdot 119875 + 119861
log sig (119882 sdot 119875 + 119861) =1
1 + 119890minus(119882sdot119875+119861)
(16)
Neuron function 119878 (sigmoid) is log sig 119882 is weight matrix119875 is input vector and 119861 is threshold We therefore proposeda compact radial basis function based on the Gaussian radialbasis function and Helensrsquo [28] definition expressed as
[exp (minus 119877)] 119877 = 119882 sdot 119875 + 119861
(exp (minusabs ((119882 sdot 119875 + 119861)))
120601 (out) = [exp (minus 119877)]
(17)
119882 is weight matrix 119875 is input vector and 119861 is threshold Thefocus was to improve on the radial basis function for themineapplication
An optimized vector 119877 was generated as the optimumset of transmission routing table that has the highest survivalprobability for data transmission (13)
119875119873 =
119873
sum
119894=1
119877119894
119878119894
+
2
sum
119894=1
119878119894
+
2
sum
119894=1
119878119894
119878119894
+ 1
for 1 le 119894 le 119878 1 le 119896 le 119877
(18)
119877 1198781
1198782
are number of neurons at input hidden and outputlayers respectively
119875119873 is the position of the 119899th particle
119882119894
= 119878119894
1198771198822
= 119878119894
119877 119882119898
= 119878119898
119877
119875119894
= [119877 (119894 minus 119895) + 119870] = 119882119894
(119894 119896)
(19)
There are two thresholds (1198781
rArr 1198611
)Hidden 119861
1
(119894 119895) = 119875(1198771198781
+ 1198781
1198782
) rArr S1
Output 1198612
(119894 119895) = 119875(1198771198781
+ 1198781
1198782
) rArr 1198782
From (16) (17) and (18) the nonlinear weight of [plusmn cos(119877)] or[plusmn sin(119877)] was imposed on the CRBF before being combinedwith the SBF as (HSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sinHSCR-BF
+ sin)The cosine weight was used to keep high levelnonlinear for small input value and reduce the nonlinear forlarge input values while sine weight was used to keep highlevel nonlinear for large input value and reduce the nonlinearfor small input values
HSCR-BFminus cos
119884 (out) = log sig (119877) + [exp (minus 119877)] sdotlowast
[minus cos (119877)]
HSCR-BF+ cos
119884 (out) = log sig (119877) + [exp (minus 119877)] sdotlowast
[+ cos (119877)]
HSCR-BFminus sin
119884 (out) = log sig (119877) + [exp (minus 119877)] sdotlowast
[minus sin (119877)]
HSCR-BF+ sin
119884 (out) = log sig (119877) + [exp (minus 119877)] sdotlowast
[+ sin (119877)]
(20)
Thenonlinear weight of [+ cos(119877)]was imposed on theGRBFbefore being combined with the SBF
HSGR-YBF+ cos
119884 (out) = log sig (119877) + [exp (minus1198772
)] sdotlowast
[+ cos (119877)]
(21)
5 Particle Swarm Optimization
Particle swarm optimization (PSO) an evolutionary algo-rithm is a population based stochastic optimization tech-nique The idea was conceived by an American researcherand social psychologist James Kennedy in the 1950s Thetheory is inspired by social behavior of bird flocking or fishschooling The method falls within the category of swarmintelligence methods for solving global optimization prob-lems Literature has shown that the PSO is an effectivealternative to established evolutionary algorithms (GA) It isalso established that PSO is easily applicable to real worldcomplex problems with discrete continuous and nonlineardesign parameters and retains the conceptual simplicity ofGA [29 30] Each particle within the swarm is given an initialrandom position and an initial speed of propagation Theposition of the particle represents a solution to the problemas described in a matrix 120591 where 119872 and 119873 represent thenumber of particles in the simulation and the number ofdimensions of the problem respectively [31 32] A randomposition representing a possible solution to the problem withan initial associated velocity representing a function of thedistance from the particlersquos current position to the previous
6 ISRN Artificial Intelligence
position of good fitness value was given A velocity matrix119881el with the same dimensions as matrix 120591
119909
described this
120591119909
= (
12059111
12059112
1205911119873
12059121
12059122
1205912119873
1205911198981
1205911198982
120591119872119873
)
119881el = (
V11 V12 V1119873
V21 V22 V2119873
V1198981 V1198982 V119872119873
)
(22)
The best feasible alternative is the genetic algorithmHoweverthis is generally slow as compared to the PSO and thereforecomputationally expensive hence the PSO
While moving in the search space particles commit tomemorize the position of the best solution they have foundAt each iteration of the algorithm each particle moves witha speed that is a weighed sum of three components the oldspeed a speed component that drives the particles towardsthe location in the search space where it previously foundthe best solution so far and a speed component that drivesthe particle towards the location in the search space wherethe neighbor particles found the best solution so far [7] Thepersonal best position can be represented by an119873times119873matrix120588best and the global best position is an 119873-dimensional vector119866best
120588best = (
12058811
12058812
1205881119873
12058821
12058822
1205882119873
1205881198981
1205881198982
120588119872119873
)
119866best = (119892best1
119892best12
119892best119873
)
(23)
All particles move towards the personal and the globalbest with 120591 120588best 119881el and 119892best containing all the requiredinformation by the particle swarm algorithmThese matricesare updated on each successive iterations
119881119898119899
= 119881119898119899
+ 1205741198881
1205781199031
(119901best119898119899
minus 119883119898119898
)
+ 1205741198882
1205781199032
(119892best119899
minus 119883119898119899
)
119883119898119899
= 119881119898119899
(24)
1205741198881
and 1205741198882
are constants set to 13 and 2 respectively and1205781199031
1205781199032
are random numbers
51 AdaptiveMutation according toThreshold Particle swarmoptimization has been effective in training neural networkssuch as a Parallel Particle Swarm optimization (PPSO)methodwith dynamic parameter adaptation used to optimizecomplex mathematical functions and improved evolutionarymethod in fuzzy logic [30 31 33] To prevent particlesfrom not converging or converging at local minimum an
Input nodes64748
64748
64748
Hidden nodesOutput nodes
S11R
S12R
SR
120573 ge 07
S1mR
120573 ge 07 le 09
120573 ge 07 le 09
Figure 1 The structure AMPSO for CRBF GRBF and SBF transferfunctions
1 2 3 4 5 6 7 8 9 100
02040608
11214
The impact of the accident
Link
avai
labi
lity
The curve of transmission link after accident
The higher the impact the lower the link availability
Region 3Links are completely dead
Region 1Links are not
Region 2Indicate probability of link being able to transmit data
affected
Figure 2 Impact of explosionaccident on transmission link
adaptivemutation according to thresholdwas introduced Analternative to this is the use of the nonadaptive mutation andthis could converge to the local minimum or fail to convergeat all The advantage of the nonadaptation is simple and fastbut may not provide the needed results
Particles positions were updated with new value onlywhen the new value is greater than the previous value 20 ofparticles of those obtaining lower values weremade tomutatefor faster convergence and the structure of adaptive mutationPSO (AMPSO) with threshold can be found in [1] The inputlayer takes the final survival vector (13) with a number ofhidden layers and an output layer The feedforward neuralnetwork was used The structure of Adaptive Mutation PSO(AMPSO) with threshold Neural Network was used Figure 1[32]
6 Results and Discussion
61 Generating the Routing Path Elements 0 1 and 2 in 120593119909
imply that the link(s) were not affected and elements 3 45 and 6 represent a probability for the links being able totransmit data while figures from 7 means the link is totallydead Region 1 that indicates that links are not affected region
ISRN Artificial Intelligence 7
Table 2 Optimum set of routing table redundancy with the highest survival probability (6 cases)
Vectors for each case Rock hardness cases1198771 05409 04281 08834 0585 08134 05163 071198772 04526 05841 04547 08997 08436 02696 081198773 02931 05899 05903 08309 06971 02898 0831198774 04851 06085 06913 04002 05026 0 0861198775 04917 05857 07805 06037 05182 06472 08751198776 02902 07253 0458 05762 01397 0 09
2 gives the probability of link available and the last regionindicates the link is completely down (Figure 2)
Thematrices120593119894 120593119896 120593119903 120593119909 120593ℎ for 120591 = 119899 with dimensionsof 119871 = 3 119877 = 2 and119862 = 1 for depth (level) row (length) andwidth (column) respectively were generated as follows
119878eq = (
(
1 1 1
1 2 1
2 2 1
2 1 1
3 1 1
3 2 1
)
)
119894120593 =
[[[[[[[
[
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
]]]]]]]
]
120593119896 = (
(
1 1 1 1 0 0
1 1 1 0 0 0
0 1 1 1 0 0
0 0 1 1 1 0
0 0 0 1 1 1
0 0 0 1 1 1
)
)
1 = signal points
120593119903 = (
(
1 1 1 0 0 0
1 1 1 0 0 0
0 1 1 1 0 0
0 0 1 1 1 0
0 0 0 1 1 1
0 0 0 1 1 1
)
)
1 = transmission path
120593119909 = (
(
0 1 2 2 0 0
1 5 1 2 0 0
0 8 1 5 9 1
1 1 5 1 1 17
1 1 4 11 0 2
4 4 19 5 3 0
)
)
explosion sizes
120593119891 = (
(
1 1 1 1 1 1
1 6 1 1 1 1
1 0 1 6 0 1
1 1 6 1 1 0
1 1 75 0 1 1
75 75 0 6 1 1
)
)
(25)
120593ℎ = (
(
1 1 1 0 0 0
1 6 1 0 0 0
0 0 1 6 0 0
0 0 6 1 1 0
0 0 0 0 1 1
0 0 0 6 1 1
)
)
120593119900 = (
(
8 8 8 0 0 0
8 48 8 0 0 0
0 0 8 48 0 0
0 0 48 8 8 0
0 0 0 0 8 8
0 0 0 48 8 8
)
)
120593119890 = (
(
144 144 144 0 0 0
144 864 144 0 0 0
0 0 144 864 0 0
0 0 864 144 144 0
0 0 0 0 144 144
0 0 0 864 144 144
)
)
1 or fractional figures are transmission path
(26)
Element ldquo0rdquo on 120593119891 depicts a connection lost while ldquo1rdquo meansthat the connection will never go down and represent theconnection to the fixed sink node(s) along the edge or theemergency connection to the mobile data mule(s)
120593s = (
(
8333 8889 9091 8333 8571 8333
8889 8333 8750 8571 8750 8333
8571 9167 8750 9091 8889 8333
8750 8571 8571 8571 8571 8333
8333 8333 8571 8333 8889 9000
9167 8333 9091 8333 8889 8571
)
)
(27)
The survival rate of RFID hardware and each minervectors was displayed as
120593119868 = 8571 8750 8571 8889 8571 8333
120593119867 = 6236 4976 1 6851 9286 6086
120593V = 5409 4281 8834 5850 8134 5163
(28)
Optimization was done numerically usingMatlab simulationtool to find the optimum set of routing tables throughparticle swarm search for rescue operation as discussed inthe preliminaries From (13) the final survival vector fordimensions of 119871 = 3 119877 = 2 and 119862 = 1 will be an average ofthe 6 vectors as in Table 2 Total cases 119877119899 = 120591 = 119879times119879 (which
8 ISRN Artificial Intelligence
Table 3 Scalability of model to survival probability range robot location and rock type
Location Mica Coal Granite Feldspar Quartz MineralCompact radial basis function with negative nonlinear weight (HSCR-BF
minus cos)(10 6 5) 08769ndash10000 09070ndash09927 08402ndash09943 08023ndash09778 07386ndash09376 07345ndash08740(10 5 4) 09024ndash09696 09032ndash09505 08836ndash09430 08633ndash09527 07944ndash09465 07263ndash08720(6 5 4) 08906ndash09877 08827ndash09845 08384ndash09675 08627ndash09820 07911ndash09142 06392ndash07840(3 1 10) 07302ndash08481 06588ndash07816 06341ndash07473 06503ndash07249 05930ndash07156 04445ndash05621
Gaussian radial basis function with positive nonlinear weight (HSGR-BF+ cos)
(10 6 5) 08732ndash09963 08526ndash09987 08492ndash09950 07867ndash09568 07404ndash09025 06606ndash08315(10 5 4) 08843ndash09593 08843ndash09853 08444ndash09753 07746ndash09828 07304ndash09595 06678ndash08817(6 5 4) 08785ndash09983 08902ndash09996 08200ndash09997 07927ndash09446 07065ndash08565 06572ndash08063(3 1 10) 07608ndash08552 06600ndash08325 06090ndash07371 05482ndash06249 05026ndash06056 04856ndash05453
10 20 30 40 50 60 70 80 90 1000
02
04
06
X 94Y 0009904
Number of iterations
Fitn
ess v
alue
PSO optimized error for hybrid SBF-CRBF with negative nonlinear weight
15 2 25 3 35 4 45 5 55 604
06
08
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(HSCRminuscos )
(a) HSCR-BFminus cos
50 100 150 200 2500
01
02
03
X 249Y 001151
PSO optimized error with nonlinear weight [minussine(R)]
Number of iterations
Fitn
ess v
alue
1 2 3 4 5 6
07
08
09
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(b) HSCR-BFminus sin
Figure 3 Hybrid of SBF and CRBF (HSCR-BF) with (a) HSCR-BFminus cos and (b) HSCR-BF
minus sin nonlinear weight
represent the total nodes deployed) for this scenario Rocksoftness or hardness takes the values of 120573 and each matrix isgenerated 120591 timecases and the final routing vector will be anaverage of the 119899 vectors 120593119877 = Average (1198771 1198776) [Table 2]which becomes the input for the PSO training
The final survival rate is 119877 = (05409 04281 08834
0585 08130 05163) from (13)
62 The Final Survival Vector The 119877 = (05409 04281
08834 0585 08130 05163) from (13) is the maximumsurvival probability for a total of 6 nodes deployed Itdescribes the success rate from each node to the sink(s) Inmost practical applications more than one sink is used andsink node is either through the fiber or TTE connection Thesize of the vector depends on the dimensions of the fieldThe elements 119877 = (05409 04281 08834 0585 08130
05163) represent the probability of 54 43 88 5981 and 52 success of each node transmitting data to andfrom its source or destination It assists decision makers as towhether to send data through one ormore nodes or send each
message twice The total nodes used for the simulation were300 with undergroundmine dimensions of 119871 = 10119877 = 6 and119862 = 5 for depth (level) row (length) and width (column)respectively with ldquopmrdquo a sensor apart pm = 100119872
120588
= 4 and119873119890 = 2 The PSO training used swarm size of 20 maximumposition was set to 100 max velocity = 1 and maximumnumber of iteration = 250 The thresholds 120582119871 and 120582119867 were3 and 6 respectively 120591 = 6 cases (each case representa rock hardness case and also represent the total neuronsused) thus each matrix and vectors were run 6 times beforeneural training The survival probability (bottom) indicatesthat the model survived between 90 and 100 where rockcases were relatively soft (ge07) The survival probabilitydeclines as the rock becomes harder and approaches 08 Atthe hardest rock of 09 the survival probability fell between72 and 84 for the entire hybrids In view of this the APheads had to be deployed at a location where the rock isrelatively soft for maximum signal strength Each computersimulation incorporated all the 6 different cases of rockhardnesssoftness to produce the matrices Detailed locationanalysis or the scalability of the model in relation to survival
ISRN Artificial Intelligence 9
Table 4 Hybrid of SBF and CRBF with negative nonlinear weight on CRBF (HSCR-BFminus cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 231020 343498 0382 91650 00985AVG 23102 34350 0038 9165 00098MinMax 29540 45314 0024 80000 00043
Table 5 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BFminus sin)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 307320 488381 0376 114300 0135AVG 30732 48838 0038 11430 0013MinMax 30740 56926 0007 61500 0021
probability range robot location and different rock types isrecorded in Table 3 Figures 4ndash6 represent different scenariosof SBF and CRBF hybrids with the [plusmn cos(119877)] and [plusmn sin(119877)]
nonlinear weights while Figure 7 represents the SBF andGaussian hybrids for [+ cos(119877)] nonlinear weight The tophalf of each figure indicates the optimized error or the finalerror after the neural training and the bottom half revealsmodel survival probability
Figure 3 discusses the SBF and CRBF hybrids (HSCR-BF) with nonlinear weight Training in negative nonlinearweights (HSCR-BF
minus cos HSCR-BFminus sin) responded well The
HSCR-BFminus cos had a steady and compact routing path which
was consistent through all rock layers with initial survivalprobability between 877 and 100 in soft layers decliningto 735ndash874 at harder rock layers The HSCR-BF
minus sinperformed quite well but was more dispersed (879 to 985at the soft rock and 665 to 818 at the hard rock)
In addition the various parameters mean iterationstandard variance standard deviation and the convergenttime of HSCR-BF
minus cos were optimised This is demonstratedby finding the relationship between the maximum and theminimum values of the parameters and comparing with theaverage figures and the closer the difference is to the averagethe more consistent the data are in the dataset (Table 4) thusHSCR-BF
minus cos provided the best results among the proposedhybrids The HSCR-BF
minus sin on the RBF yielded strength inmean iteration and standard variance with error of 0013(Table 5)
In Figure 4 training with nonlinear weight of positivecosine and sine of HSCR-BF (HSCR-BF
+ cos and HSCR-BF+ sin) had some similarities The two models streamed
well at the initial stages from 896 to 992 and 908to 990 at soft layers of the rock respectively Howeverat the last stages the probability of transmitting effectivelybecame marginal from 795 to 852 and 821 to 854respectively Much as this hybrid could perform well in areaswhere routing conditions aremuch better in less dense rescuesituations this could hamper rescue mission in both casesdue to battery drain or collision from traffic conjunctionThedata formean iteration and convergent time forHSCR-BF
+ cos(Table 6) and mean Iteration standard variance the final
error HSCR-BF+ sin (Table 7) were consistent with the nega-
tive sine having better results than the positive cosineThe survival probability of the positive nonlinear weight
Gaussian hybrids (HSGR-BF+ cos) in Figure 5 is more com-
pact both at the initial (ie 873ndash996) and later (661ndash832) stages declining to harder rock layers with an errorof 001173 (Table 8)
At the initial stage particles are sensitive to inputs asthey moved quickly in the search space towards the targetand while particles peaked closer to the target it became lesssensitive to the input and began to align (Figure 6) At thelater stage more RBF are used to keep the error at minimumfor accuracy It was also noticed that instead of acceler-ating higher and becoming more sensitive to the inputswhile particles were far from the target the hybrids withlowast
[minus sin(119877)] and lowast[ + sin(119877)] were less sensitive to the inputsas lowast[minus sin(119877)] descended from [119909 = minus1 119910 = 05785] to[119909 = 1 119910 = 04215] while lowast[+ sin(119877)] descended from [119909 =
minus3 119910 = 00404] to [119909 = minus1 119910 = minus004062] before becomingconscious of the inputs In addition the hybrids with thelowast
[+ cos(119877)] lagged slightly between [119909 = minus4 119910 = 0006014]
and [119909 = minus2 119910 = 006288] before becoming sensitive toinputs However they all lined up for accuracy in terms of theminimised error
The nonlinearity of the hybrids is presented (Figure 7)the negative cosinesine weight is used to reduce nonlinearfor small input values and this lies between minus1 lt 119909 lt +1 andthe positive cosinesine weight is used to keep high nonlinearfor large inputs for the remaining region
63 CPU Time Efficiency The relationship between varioushybrids with respect to the central processing time (CPU)was profiled for different runs (Table 9) and expressed in asixth-order polynomial given as 119884
119903
= 1205737
1198836
1
+ 1205736
1198835
1
+ 1205735
1198834
1
+
1205734
1198833
1
+ 1205733
1198832
1
+ 1205732
1198831
+ 1205730
where 1198831
is time (seconds) and 120573
is the coefficient of the polynomial (Figure 8) The proposedhybrid has better usage of CPU time with 119877
2
= 09160followed byHSCR-BF
minus sin andHSCR-BF+ cos with119877
2 of 07551and 07244 respectively (Table 10) Applying the proposedalgorithm into Gaussian the CPU usage had 119877
2 of 07345
10 ISRN Artificial Intelligence
0 50 100 150 200
002
004
006
008
X 249Y 001103
PSO optimized error for hybrid SBF-CRBF with positive nonlinear weight
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 607
08
09
1
11 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(HSCR+cos )
(a) HSCR-BF+ cos
0 50 100 150 200
005
01
015
02
X 249Y 001247
PSO optimized error with nonlinear weight [+sine(R)]
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 6
070809
111 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(b) HSCR-BF+ sin
Figure 4 Hybrid of SBF and CRBF (HSCR-BF) with (a) HSCR-BF+ cos and (b) HSCR-BF
+ sin nonlinear weight
Table 6 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BF+ cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 350780 522257 0311 115650 0121AVG 35078 52226 0031 11565 0012MaxMin 54840 63413 0009 93500 0006
and indeedmarginally outperformedHSCR-BF+ cos Detailed
work on SBF CRBF and GRBF with regard to scalabilitymemory usage and the central processing time has beencarried by the authors [1]
7 Conclusion
In summary we made the following contributions First weused the mix of SBF and CRBF to present several hybridswith different nonlinear weights of cosine and sine func-tions on compact radial basis function Next we showedthe performance of the proposed nonlinear weight hybridsHSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sin and HSCR-BF+ sin optimised all the parameters with minimised error
of 00098 0012 0013 and 0013 respectively compared to00117 for Gaussian HSCR-BF
+ cos The analyzed CPU usagewith corresponding 119877
2 values of 09160 075 072440 06731and 07345 for HSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sinHSCR-BF
+ sin and HSCR-BF+ cos respectively demonstrated
that the algorithm is scalable There exist some evacuationmodels that is Goh and Mandic [11] which offered a choicefor travelers and several schemes for the decisionmakersmaynot be applicable in emergency undergroundmine situationsThe proposed nonlinear hybrid algorithm with particle swarmoptimisation has better capability of approximation to underly-ing functions with a fast learning speed and high robustness and
is competitive and more computationally efficient to Gaussianwith the same nonlinear weight The algorithm is new andthat makes it difficult to identify limitations and we intend toinvestigate other hybrids and comparewith genetic algorithm(GA) in the future
Nomenclature
119873120575min 119873120575max Minimum and maximum signalreach
HSCR-BFHSGR-BF Hybrid of SBF with CRBFHybrid ofSBF with GRBF
HSCR-BFminus cos Hybrid of SBF and CRBF minuscos
nonlinear weightHSCR-BF
+ cos Hybrid of SBF and CRBF +cosnonlinear weight
HSCR-BFminus sin Hybrid of SBF and CRBF minussine
nonlinear weightHSCR-BF
+ sin Hybrid of SBF and CRBF +sinenonlinear weight
HSGR-BF+ cos Hybrid of SBF and GRBF +cos
nonlinear weight
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
ISRN Artificial Intelligence 11
Table 7 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BF+ sin)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 260360 407109 03457 99800 0129AVG 26036 40711 00346 9980 0013MinMax 27840 56040 0008 106500 0014
Table 8 Hybrid of SBF and GRBF with nonlinear weight on GRBF (HSGR-BF+ cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 219840 349010 0439 112100 01173AVG 21984 34901 0044 11210 00117MinMax 27000 46580 0008 80000 00165
0 50 100 1500
005
01
015
02
X 144Y 0009595
GRBF-PSO optimized error with nonlinear weight [+cos(R)]
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 604
06
08
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
Figure 5 Hybrid of SBF and GRBF with nonlinear weight on GRBF(HSCR-BF
+ cos)
0 2 4 6 8 10
002040608
1121416
Weighed position of particles
Min
imise
d er
ror
Optimised error in nonlinear weighed SBF-CRBF hybridsParticles
Less sensitive to inputsmuch closer to target
more accurate
Less sensitive to inputsmuch closer more accurate
Neurons are more sensitive to inputsfast and far from target
minus02minus2minus4minus6minus8minus10
[x = minus1 y = 05785]
[x = 1 y = 04215]
[minus1 004062]
[x = minus2 y = 006288]
[x = minus4 y = 0006014]
[minus3 00404]
Hybrid+cosHybrid+sin
HybridminuscosHybridminussin
Figure 6 Particles weighed position
Acknowledgments
The authors would like to appreciate the immense contribu-tion of the mining companies where the study was under-taken The authors are grateful to these individuals for
0 1 2 3 4 5
0
05
1
15
2
X 37Y 0989
Particles position
Opt
imise
d er
ror
Nonlinear weight hybrids
Keep large inputvalues more
nonlinear
Keep small inputvalues lessnonlinear
Large input values
Small input values
Using positive cosine and sine
Using negative cosine and sineminus05
minus5 minus4 minus3 minus2 minus1
HSCR-BF+cosHSGR-BF+cos
HSCR-BF+sin
HSCR-BFminuscos
HSCR-BFminuscos
HSCR-BFminussin
HSCR-BF+cos HSCR-BF+sin HSCR-BFminussinand HSGR-BF+cos
Figure 7 Nonlinear weighed curves for optimized error withoutGaussian negative cosine curve
HSGR-BF+cos
HSGR-BF+cos
HSCR-BF+sinHSCR-BFminuscosHSCR-BFminussin
00020000400006000080000
100000120000
0 2 4 6 8 10 12
CPU
tim
e
Time (s)
Nonlinear hybrids
Figure 8 The trend of hybrids CPU time curves using sixth-OrderPolynomial
their immeasurable contributions to this work Fred Atta-kuma Patrick Addai Isaac Owusu Kwankye Thomas KwawAnnan Willet Agongo Nathaniel Quansah Francis OwusuMensah Clement Owusu-Asamoah Joseph Adu-MensahShadrack Aidoo Martin Anokye F T Oduro Ernest Ekow
12 ISRN Artificial Intelligence
Table 9 CPU time usage efficiency for the hybrids
Runs HSCR-BFminus cos HSCR-BF
minus sin HSCR-BF+ cos HSCR-BF
+ sin HSGR-BF+ cos
Total 255889 254350 422344 643962 438133Average 25589 25435 42234 64396 43813
Table 10 Analysis of CPU time on the various hybrids
1205736
1205735
1205734
1205733
1205732
1205731
1205730
1198772
HSCR-BFminus cos 00159119909 minus02575119909 minus0759119909 33494119909 minus19901119909 40345119909 25322 09160
HSCR-BF+ cos 00903119909 minus23158119909 19932119909 minus58027119909 minus31337119909 32808119909 14747 07244
HSCR-BFminus sin minus0033119909 11354119909 minus15811119909 11368119909 minus4337119909 78008119909 minus20794 07551
HSCR-BF+ sin 02571119909 minus14263119909 17173119909 minus10069119909 2993119909 minus42374119909 27879 06731
HSGR-BF+ cos minus00437119909 10911119909 minus82923119909 66218119909 15934119909 minus53939119909 8191 07345
Abano E Omari-Siaw and Qiang Jia This work was sup-ported by the National Natural Science Foundation of China(no 71271103) and by the Six Talents Peak Foundation ofJiangsu Province
References
[1] M O Ansong H X Yao and J S Huang ldquoRadial and sigmoidbasis functions neural networks in wireless sensor routingtopology control in underground mine rescue operation basedon particle swarm optimizationrdquo International Journal of Dis-tributed Sensor Networks-Hindawi vol 2013 Article ID 37693114 pages 2013
[2] R Neruda and P Kudova ldquoLearning methods for radial basisfunction networksrdquo Future Generation Computer Systems vol21 no 7 pp 1131ndash1142 2005
[3] C T Cheng J Y Lin and Y G C Sun ldquoLong-term predictionof discharges in Manwan hydropower using adaptive-network-based fuzzy inference systems modelsrdquo in Advances in NaturalComputation vol 3498 of Lecture Notes in Computer Sciencepp 1152ndash1161 2005
[4] D Broomhead and D Lowe ldquoMultivariate functional interpo-lation and adaptive networksrdquo Complex System vol 2 pp 321ndash355 1988
[5] Y Chen C Chuah and Q Zhao ldquoNetwork configuration foroptimal utilization efficiency of wireless sensor networksrdquo AdHoc Networks vol 6 no 1 pp 92ndash107 2008
[6] P S Rajpal K S Shishodia and G S Sekhon ldquoAn artificialneural network for modeling reliability availability and main-tainability of a repairable systemrdquo Reliability Engineering andSystem Safety vol 91 no 7 pp 809ndash819 2006
[7] W Jang W M Healy and S J Mirosław ldquoWireless sensornetworks as part of a web-based building environmental mon-itoring systemrdquo Automation in Construction vol 17 no 6 pp729ndash736 2008
[8] M S Pan C H Tsai and Y C Tseng ldquoEmergengy guiding andmonitoring application in indoor 3D environment by WirelessSensor Networkrdquo Internation Journal Os Sensor Netwourk vol1 pp 2ndash10 2006
[9] S Zarifzadeh A Nayyeri and N Yazdani ldquoEfficient construc-tion of network topology to conserve energy in wireless ad hocnetworksrdquo Computer Communications vol 31 no 1 pp 160ndash173 2008
[10] R Riaz A Naureen A Akram A H Akbar K Kim and HFarooq Ahmed ldquoA unified security framework with three keymanagement schemes for wireless sensor networksrdquo ComputerCommunications vol 31 no 18 pp 4269ndash4280 2008
[11] S L Goh and D P Mandic ldquoAn augmented CRTRL forcomplex-valued recurrent neural networksrdquo Neural Networksvol 20 no 10 pp 1061ndash1066 2007
[12] L A B Munoz and J J V Ramosy ldquoSimilarity-based heteroge-neous neural networksrdquo Engineering Letters vol 14 no 2 pp103ndash116 2007
[13] R Taormina K Chau and R Sethi ldquoArtificial neural networksimulation of hourly groundwater levels in a coastal aquifer sys-tem of the Venice lagoonrdquo Engineering Applications of ArtificialIntelligence vol 25 no 8 pp 1670ndash1676 2012
[14] K Leblebicioglu and U Halici ldquoInfinite dimensional radialbasis function neural networks for nonlinear transformationson function spacesrdquo Nonlinear Analysis Theory Methods andApplications vol 30 no 3 pp 1649ndash1654 1997
[15] F Fernandez-Navarro C Hervas-Martınez J Sanchez-Monedero and P A Gutierrez ldquoMELM-GRBF a modifiedversion of the extreme learning machine for generalized radialbasis function neural networksrdquo Neurocomputing vol 74 no16 pp 2502ndash2510 2011
[16] C L Wu KW Chau and Y S Li ldquoPredicting monthly stream-flow using data-drivenmodels coupledwith data-preprocessingtechniquesrdquoWater Resources Research vol 45 no 8 Article IDW08432 2009
[17] C EACheng ldquoLong-termprediction of discharges inManwanReservoir using artificial neural network modelsrdquo in Advancesin Neural NetworksmdashISNN 2005 vol 3498 of Lecture Notes inComputer Science pp 1040ndash1045 2005
[18] H H Zhang M Genton and P Liu ldquoCompactly supportedradial basis function kernelsrdquo 2004
[19] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquo inProceedings of the 1995 IEEE International Conference on NeuralNetworks pp 1942ndash1945 December 1995
[20] J Eberhart R Eberhart and Y Shi ldquoParticle swarm optimiza-tion developments applications and resourcesrdquo IEEE vol 1 pp81ndash86 2001
[21] R Malhotra and A Negi ldquoReliability modeling using ParticleSwarm optimisationrdquo International Journal of System AssuranceEngineering and Management vol 4 no 3 pp 275ndash283 2013
[22] D Gies and Y Rahmat-Samii ldquoParticle swarm optimization forreconfigurable phase-differentiated array designrdquo Microwaveand Optical Technology Letters vol 38 no 3 pp 168ndash175 2003
ISRN Artificial Intelligence 13
[23] F Valdez P Melin and C Oscar ldquoAn improved evolution-ary method with fuzzy logic for combining Particle Swarmoptimization and genetic algorithmsrdquo Applied Soft ComputingJournal vol 11 no 2 pp 2625ndash2632 2011
[24] J Zhang and K Chau ldquoMultilayer ensemble pruning vianovel multi-sub-swarm particle swarm optimizationrdquo Journalof Universal Computer Science vol 15 no 4 pp 840ndash858 2009
[25] F Valdez P Melin and O Castillo ldquoParallel Particle Swarmoptimization with parameters adaptation using fuzzy logicrdquoin Proceedings of the 11th Mexican International Conference onArtificial Intelligence (MICAI rsquo12) vol 2 pp 374ndash385 San LuisPotosi Mexico October 2012
[26] J Peng and C P Y Pan ldquoParticle swarm optimization RBFfor gas emission predictionrdquo Journal of Safety Science andTechnology vol 7 pp 77ndash85 2011
[27] G Ren Z Huang Y Cheng X Zhao and Y Zhang ldquoAnintegrated model for evacuation routing and traffic signaloptimization with background demand uncertaintyrdquo Journal ofAdvanced Transportation vol 47 pp 4ndash27 2013
[28] X Feng Z Xiao and X Cui ldquoImproved RSSI algorithm forwireless sensor networks in 3Drdquo Journal of ComputationalInformation Systems vol 7 no 16 pp 5866ndash5873 2011
[29] F Valdez and P Melin ldquoNeural network optimization witha hybrid evolutionary method that combines particle swarmand genetic algorithms with fuzzy rulesrdquo in Proceedings ofthe Annual Meeting of the North American Fuzzy InformationProcessing Society (NAFIPS rsquo08) NewYork NYUSAMay 2008
[30] F Fernandez-Navarro C Hervas-Martınez P A Gutierrez JM Pena-Barragan and F Lopez-Granados ldquoParameter estima-tion of q-Gaussian Radial Basis Functions Neural Networkswith a Hybrid Algorithm for binary classificationrdquo Neurocom-puting vol 75 pp 123ndash134 2012
[31] H Soh S Lim T Zhang et al ldquoWeighted complex networkanalysis of travel routes on the Singapore public transportationsystemrdquo Physica A vol 389 no 24 pp 5852ndash5863 2010
[32] C K S Kumar R Sukumar and M Nageswari ldquoSensorslifetime enhancement techniques in wireless sensor networksmdasha critical reviewrdquo International Journal of Computer Science andInformation Technology amp Security vol 3 pp 159ndash164 2013
[33] S Li and F Qin ldquoA dynamic neural network approach forsolving nonlinear inequalities defined on a graph and itsapplication to distributed routing-free range-free localizationof WSNsrdquo Neurocomputing vol 117 pp 72ndash80 2013
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
6 ISRN Artificial Intelligence
position of good fitness value was given A velocity matrix119881el with the same dimensions as matrix 120591
119909
described this
120591119909
= (
12059111
12059112
1205911119873
12059121
12059122
1205912119873
1205911198981
1205911198982
120591119872119873
)
119881el = (
V11 V12 V1119873
V21 V22 V2119873
V1198981 V1198982 V119872119873
)
(22)
The best feasible alternative is the genetic algorithmHoweverthis is generally slow as compared to the PSO and thereforecomputationally expensive hence the PSO
While moving in the search space particles commit tomemorize the position of the best solution they have foundAt each iteration of the algorithm each particle moves witha speed that is a weighed sum of three components the oldspeed a speed component that drives the particles towardsthe location in the search space where it previously foundthe best solution so far and a speed component that drivesthe particle towards the location in the search space wherethe neighbor particles found the best solution so far [7] Thepersonal best position can be represented by an119873times119873matrix120588best and the global best position is an 119873-dimensional vector119866best
120588best = (
12058811
12058812
1205881119873
12058821
12058822
1205882119873
1205881198981
1205881198982
120588119872119873
)
119866best = (119892best1
119892best12
119892best119873
)
(23)
All particles move towards the personal and the globalbest with 120591 120588best 119881el and 119892best containing all the requiredinformation by the particle swarm algorithmThese matricesare updated on each successive iterations
119881119898119899
= 119881119898119899
+ 1205741198881
1205781199031
(119901best119898119899
minus 119883119898119898
)
+ 1205741198882
1205781199032
(119892best119899
minus 119883119898119899
)
119883119898119899
= 119881119898119899
(24)
1205741198881
and 1205741198882
are constants set to 13 and 2 respectively and1205781199031
1205781199032
are random numbers
51 AdaptiveMutation according toThreshold Particle swarmoptimization has been effective in training neural networkssuch as a Parallel Particle Swarm optimization (PPSO)methodwith dynamic parameter adaptation used to optimizecomplex mathematical functions and improved evolutionarymethod in fuzzy logic [30 31 33] To prevent particlesfrom not converging or converging at local minimum an
Input nodes64748
64748
64748
Hidden nodesOutput nodes
S11R
S12R
SR
120573 ge 07
S1mR
120573 ge 07 le 09
120573 ge 07 le 09
Figure 1 The structure AMPSO for CRBF GRBF and SBF transferfunctions
1 2 3 4 5 6 7 8 9 100
02040608
11214
The impact of the accident
Link
avai
labi
lity
The curve of transmission link after accident
The higher the impact the lower the link availability
Region 3Links are completely dead
Region 1Links are not
Region 2Indicate probability of link being able to transmit data
affected
Figure 2 Impact of explosionaccident on transmission link
adaptivemutation according to thresholdwas introduced Analternative to this is the use of the nonadaptive mutation andthis could converge to the local minimum or fail to convergeat all The advantage of the nonadaptation is simple and fastbut may not provide the needed results
Particles positions were updated with new value onlywhen the new value is greater than the previous value 20 ofparticles of those obtaining lower values weremade tomutatefor faster convergence and the structure of adaptive mutationPSO (AMPSO) with threshold can be found in [1] The inputlayer takes the final survival vector (13) with a number ofhidden layers and an output layer The feedforward neuralnetwork was used The structure of Adaptive Mutation PSO(AMPSO) with threshold Neural Network was used Figure 1[32]
6 Results and Discussion
61 Generating the Routing Path Elements 0 1 and 2 in 120593119909
imply that the link(s) were not affected and elements 3 45 and 6 represent a probability for the links being able totransmit data while figures from 7 means the link is totallydead Region 1 that indicates that links are not affected region
ISRN Artificial Intelligence 7
Table 2 Optimum set of routing table redundancy with the highest survival probability (6 cases)
Vectors for each case Rock hardness cases1198771 05409 04281 08834 0585 08134 05163 071198772 04526 05841 04547 08997 08436 02696 081198773 02931 05899 05903 08309 06971 02898 0831198774 04851 06085 06913 04002 05026 0 0861198775 04917 05857 07805 06037 05182 06472 08751198776 02902 07253 0458 05762 01397 0 09
2 gives the probability of link available and the last regionindicates the link is completely down (Figure 2)
Thematrices120593119894 120593119896 120593119903 120593119909 120593ℎ for 120591 = 119899 with dimensionsof 119871 = 3 119877 = 2 and119862 = 1 for depth (level) row (length) andwidth (column) respectively were generated as follows
119878eq = (
(
1 1 1
1 2 1
2 2 1
2 1 1
3 1 1
3 2 1
)
)
119894120593 =
[[[[[[[
[
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
]]]]]]]
]
120593119896 = (
(
1 1 1 1 0 0
1 1 1 0 0 0
0 1 1 1 0 0
0 0 1 1 1 0
0 0 0 1 1 1
0 0 0 1 1 1
)
)
1 = signal points
120593119903 = (
(
1 1 1 0 0 0
1 1 1 0 0 0
0 1 1 1 0 0
0 0 1 1 1 0
0 0 0 1 1 1
0 0 0 1 1 1
)
)
1 = transmission path
120593119909 = (
(
0 1 2 2 0 0
1 5 1 2 0 0
0 8 1 5 9 1
1 1 5 1 1 17
1 1 4 11 0 2
4 4 19 5 3 0
)
)
explosion sizes
120593119891 = (
(
1 1 1 1 1 1
1 6 1 1 1 1
1 0 1 6 0 1
1 1 6 1 1 0
1 1 75 0 1 1
75 75 0 6 1 1
)
)
(25)
120593ℎ = (
(
1 1 1 0 0 0
1 6 1 0 0 0
0 0 1 6 0 0
0 0 6 1 1 0
0 0 0 0 1 1
0 0 0 6 1 1
)
)
120593119900 = (
(
8 8 8 0 0 0
8 48 8 0 0 0
0 0 8 48 0 0
0 0 48 8 8 0
0 0 0 0 8 8
0 0 0 48 8 8
)
)
120593119890 = (
(
144 144 144 0 0 0
144 864 144 0 0 0
0 0 144 864 0 0
0 0 864 144 144 0
0 0 0 0 144 144
0 0 0 864 144 144
)
)
1 or fractional figures are transmission path
(26)
Element ldquo0rdquo on 120593119891 depicts a connection lost while ldquo1rdquo meansthat the connection will never go down and represent theconnection to the fixed sink node(s) along the edge or theemergency connection to the mobile data mule(s)
120593s = (
(
8333 8889 9091 8333 8571 8333
8889 8333 8750 8571 8750 8333
8571 9167 8750 9091 8889 8333
8750 8571 8571 8571 8571 8333
8333 8333 8571 8333 8889 9000
9167 8333 9091 8333 8889 8571
)
)
(27)
The survival rate of RFID hardware and each minervectors was displayed as
120593119868 = 8571 8750 8571 8889 8571 8333
120593119867 = 6236 4976 1 6851 9286 6086
120593V = 5409 4281 8834 5850 8134 5163
(28)
Optimization was done numerically usingMatlab simulationtool to find the optimum set of routing tables throughparticle swarm search for rescue operation as discussed inthe preliminaries From (13) the final survival vector fordimensions of 119871 = 3 119877 = 2 and 119862 = 1 will be an average ofthe 6 vectors as in Table 2 Total cases 119877119899 = 120591 = 119879times119879 (which
8 ISRN Artificial Intelligence
Table 3 Scalability of model to survival probability range robot location and rock type
Location Mica Coal Granite Feldspar Quartz MineralCompact radial basis function with negative nonlinear weight (HSCR-BF
minus cos)(10 6 5) 08769ndash10000 09070ndash09927 08402ndash09943 08023ndash09778 07386ndash09376 07345ndash08740(10 5 4) 09024ndash09696 09032ndash09505 08836ndash09430 08633ndash09527 07944ndash09465 07263ndash08720(6 5 4) 08906ndash09877 08827ndash09845 08384ndash09675 08627ndash09820 07911ndash09142 06392ndash07840(3 1 10) 07302ndash08481 06588ndash07816 06341ndash07473 06503ndash07249 05930ndash07156 04445ndash05621
Gaussian radial basis function with positive nonlinear weight (HSGR-BF+ cos)
(10 6 5) 08732ndash09963 08526ndash09987 08492ndash09950 07867ndash09568 07404ndash09025 06606ndash08315(10 5 4) 08843ndash09593 08843ndash09853 08444ndash09753 07746ndash09828 07304ndash09595 06678ndash08817(6 5 4) 08785ndash09983 08902ndash09996 08200ndash09997 07927ndash09446 07065ndash08565 06572ndash08063(3 1 10) 07608ndash08552 06600ndash08325 06090ndash07371 05482ndash06249 05026ndash06056 04856ndash05453
10 20 30 40 50 60 70 80 90 1000
02
04
06
X 94Y 0009904
Number of iterations
Fitn
ess v
alue
PSO optimized error for hybrid SBF-CRBF with negative nonlinear weight
15 2 25 3 35 4 45 5 55 604
06
08
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(HSCRminuscos )
(a) HSCR-BFminus cos
50 100 150 200 2500
01
02
03
X 249Y 001151
PSO optimized error with nonlinear weight [minussine(R)]
Number of iterations
Fitn
ess v
alue
1 2 3 4 5 6
07
08
09
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(b) HSCR-BFminus sin
Figure 3 Hybrid of SBF and CRBF (HSCR-BF) with (a) HSCR-BFminus cos and (b) HSCR-BF
minus sin nonlinear weight
represent the total nodes deployed) for this scenario Rocksoftness or hardness takes the values of 120573 and each matrix isgenerated 120591 timecases and the final routing vector will be anaverage of the 119899 vectors 120593119877 = Average (1198771 1198776) [Table 2]which becomes the input for the PSO training
The final survival rate is 119877 = (05409 04281 08834
0585 08130 05163) from (13)
62 The Final Survival Vector The 119877 = (05409 04281
08834 0585 08130 05163) from (13) is the maximumsurvival probability for a total of 6 nodes deployed Itdescribes the success rate from each node to the sink(s) Inmost practical applications more than one sink is used andsink node is either through the fiber or TTE connection Thesize of the vector depends on the dimensions of the fieldThe elements 119877 = (05409 04281 08834 0585 08130
05163) represent the probability of 54 43 88 5981 and 52 success of each node transmitting data to andfrom its source or destination It assists decision makers as towhether to send data through one ormore nodes or send each
message twice The total nodes used for the simulation were300 with undergroundmine dimensions of 119871 = 10119877 = 6 and119862 = 5 for depth (level) row (length) and width (column)respectively with ldquopmrdquo a sensor apart pm = 100119872
120588
= 4 and119873119890 = 2 The PSO training used swarm size of 20 maximumposition was set to 100 max velocity = 1 and maximumnumber of iteration = 250 The thresholds 120582119871 and 120582119867 were3 and 6 respectively 120591 = 6 cases (each case representa rock hardness case and also represent the total neuronsused) thus each matrix and vectors were run 6 times beforeneural training The survival probability (bottom) indicatesthat the model survived between 90 and 100 where rockcases were relatively soft (ge07) The survival probabilitydeclines as the rock becomes harder and approaches 08 Atthe hardest rock of 09 the survival probability fell between72 and 84 for the entire hybrids In view of this the APheads had to be deployed at a location where the rock isrelatively soft for maximum signal strength Each computersimulation incorporated all the 6 different cases of rockhardnesssoftness to produce the matrices Detailed locationanalysis or the scalability of the model in relation to survival
ISRN Artificial Intelligence 9
Table 4 Hybrid of SBF and CRBF with negative nonlinear weight on CRBF (HSCR-BFminus cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 231020 343498 0382 91650 00985AVG 23102 34350 0038 9165 00098MinMax 29540 45314 0024 80000 00043
Table 5 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BFminus sin)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 307320 488381 0376 114300 0135AVG 30732 48838 0038 11430 0013MinMax 30740 56926 0007 61500 0021
probability range robot location and different rock types isrecorded in Table 3 Figures 4ndash6 represent different scenariosof SBF and CRBF hybrids with the [plusmn cos(119877)] and [plusmn sin(119877)]
nonlinear weights while Figure 7 represents the SBF andGaussian hybrids for [+ cos(119877)] nonlinear weight The tophalf of each figure indicates the optimized error or the finalerror after the neural training and the bottom half revealsmodel survival probability
Figure 3 discusses the SBF and CRBF hybrids (HSCR-BF) with nonlinear weight Training in negative nonlinearweights (HSCR-BF
minus cos HSCR-BFminus sin) responded well The
HSCR-BFminus cos had a steady and compact routing path which
was consistent through all rock layers with initial survivalprobability between 877 and 100 in soft layers decliningto 735ndash874 at harder rock layers The HSCR-BF
minus sinperformed quite well but was more dispersed (879 to 985at the soft rock and 665 to 818 at the hard rock)
In addition the various parameters mean iterationstandard variance standard deviation and the convergenttime of HSCR-BF
minus cos were optimised This is demonstratedby finding the relationship between the maximum and theminimum values of the parameters and comparing with theaverage figures and the closer the difference is to the averagethe more consistent the data are in the dataset (Table 4) thusHSCR-BF
minus cos provided the best results among the proposedhybrids The HSCR-BF
minus sin on the RBF yielded strength inmean iteration and standard variance with error of 0013(Table 5)
In Figure 4 training with nonlinear weight of positivecosine and sine of HSCR-BF (HSCR-BF
+ cos and HSCR-BF+ sin) had some similarities The two models streamed
well at the initial stages from 896 to 992 and 908to 990 at soft layers of the rock respectively Howeverat the last stages the probability of transmitting effectivelybecame marginal from 795 to 852 and 821 to 854respectively Much as this hybrid could perform well in areaswhere routing conditions aremuch better in less dense rescuesituations this could hamper rescue mission in both casesdue to battery drain or collision from traffic conjunctionThedata formean iteration and convergent time forHSCR-BF
+ cos(Table 6) and mean Iteration standard variance the final
error HSCR-BF+ sin (Table 7) were consistent with the nega-
tive sine having better results than the positive cosineThe survival probability of the positive nonlinear weight
Gaussian hybrids (HSGR-BF+ cos) in Figure 5 is more com-
pact both at the initial (ie 873ndash996) and later (661ndash832) stages declining to harder rock layers with an errorof 001173 (Table 8)
At the initial stage particles are sensitive to inputs asthey moved quickly in the search space towards the targetand while particles peaked closer to the target it became lesssensitive to the input and began to align (Figure 6) At thelater stage more RBF are used to keep the error at minimumfor accuracy It was also noticed that instead of acceler-ating higher and becoming more sensitive to the inputswhile particles were far from the target the hybrids withlowast
[minus sin(119877)] and lowast[ + sin(119877)] were less sensitive to the inputsas lowast[minus sin(119877)] descended from [119909 = minus1 119910 = 05785] to[119909 = 1 119910 = 04215] while lowast[+ sin(119877)] descended from [119909 =
minus3 119910 = 00404] to [119909 = minus1 119910 = minus004062] before becomingconscious of the inputs In addition the hybrids with thelowast
[+ cos(119877)] lagged slightly between [119909 = minus4 119910 = 0006014]
and [119909 = minus2 119910 = 006288] before becoming sensitive toinputs However they all lined up for accuracy in terms of theminimised error
The nonlinearity of the hybrids is presented (Figure 7)the negative cosinesine weight is used to reduce nonlinearfor small input values and this lies between minus1 lt 119909 lt +1 andthe positive cosinesine weight is used to keep high nonlinearfor large inputs for the remaining region
63 CPU Time Efficiency The relationship between varioushybrids with respect to the central processing time (CPU)was profiled for different runs (Table 9) and expressed in asixth-order polynomial given as 119884
119903
= 1205737
1198836
1
+ 1205736
1198835
1
+ 1205735
1198834
1
+
1205734
1198833
1
+ 1205733
1198832
1
+ 1205732
1198831
+ 1205730
where 1198831
is time (seconds) and 120573
is the coefficient of the polynomial (Figure 8) The proposedhybrid has better usage of CPU time with 119877
2
= 09160followed byHSCR-BF
minus sin andHSCR-BF+ cos with119877
2 of 07551and 07244 respectively (Table 10) Applying the proposedalgorithm into Gaussian the CPU usage had 119877
2 of 07345
10 ISRN Artificial Intelligence
0 50 100 150 200
002
004
006
008
X 249Y 001103
PSO optimized error for hybrid SBF-CRBF with positive nonlinear weight
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 607
08
09
1
11 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(HSCR+cos )
(a) HSCR-BF+ cos
0 50 100 150 200
005
01
015
02
X 249Y 001247
PSO optimized error with nonlinear weight [+sine(R)]
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 6
070809
111 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(b) HSCR-BF+ sin
Figure 4 Hybrid of SBF and CRBF (HSCR-BF) with (a) HSCR-BF+ cos and (b) HSCR-BF
+ sin nonlinear weight
Table 6 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BF+ cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 350780 522257 0311 115650 0121AVG 35078 52226 0031 11565 0012MaxMin 54840 63413 0009 93500 0006
and indeedmarginally outperformedHSCR-BF+ cos Detailed
work on SBF CRBF and GRBF with regard to scalabilitymemory usage and the central processing time has beencarried by the authors [1]
7 Conclusion
In summary we made the following contributions First weused the mix of SBF and CRBF to present several hybridswith different nonlinear weights of cosine and sine func-tions on compact radial basis function Next we showedthe performance of the proposed nonlinear weight hybridsHSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sin and HSCR-BF+ sin optimised all the parameters with minimised error
of 00098 0012 0013 and 0013 respectively compared to00117 for Gaussian HSCR-BF
+ cos The analyzed CPU usagewith corresponding 119877
2 values of 09160 075 072440 06731and 07345 for HSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sinHSCR-BF
+ sin and HSCR-BF+ cos respectively demonstrated
that the algorithm is scalable There exist some evacuationmodels that is Goh and Mandic [11] which offered a choicefor travelers and several schemes for the decisionmakersmaynot be applicable in emergency undergroundmine situationsThe proposed nonlinear hybrid algorithm with particle swarmoptimisation has better capability of approximation to underly-ing functions with a fast learning speed and high robustness and
is competitive and more computationally efficient to Gaussianwith the same nonlinear weight The algorithm is new andthat makes it difficult to identify limitations and we intend toinvestigate other hybrids and comparewith genetic algorithm(GA) in the future
Nomenclature
119873120575min 119873120575max Minimum and maximum signalreach
HSCR-BFHSGR-BF Hybrid of SBF with CRBFHybrid ofSBF with GRBF
HSCR-BFminus cos Hybrid of SBF and CRBF minuscos
nonlinear weightHSCR-BF
+ cos Hybrid of SBF and CRBF +cosnonlinear weight
HSCR-BFminus sin Hybrid of SBF and CRBF minussine
nonlinear weightHSCR-BF
+ sin Hybrid of SBF and CRBF +sinenonlinear weight
HSGR-BF+ cos Hybrid of SBF and GRBF +cos
nonlinear weight
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
ISRN Artificial Intelligence 11
Table 7 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BF+ sin)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 260360 407109 03457 99800 0129AVG 26036 40711 00346 9980 0013MinMax 27840 56040 0008 106500 0014
Table 8 Hybrid of SBF and GRBF with nonlinear weight on GRBF (HSGR-BF+ cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 219840 349010 0439 112100 01173AVG 21984 34901 0044 11210 00117MinMax 27000 46580 0008 80000 00165
0 50 100 1500
005
01
015
02
X 144Y 0009595
GRBF-PSO optimized error with nonlinear weight [+cos(R)]
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 604
06
08
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
Figure 5 Hybrid of SBF and GRBF with nonlinear weight on GRBF(HSCR-BF
+ cos)
0 2 4 6 8 10
002040608
1121416
Weighed position of particles
Min
imise
d er
ror
Optimised error in nonlinear weighed SBF-CRBF hybridsParticles
Less sensitive to inputsmuch closer to target
more accurate
Less sensitive to inputsmuch closer more accurate
Neurons are more sensitive to inputsfast and far from target
minus02minus2minus4minus6minus8minus10
[x = minus1 y = 05785]
[x = 1 y = 04215]
[minus1 004062]
[x = minus2 y = 006288]
[x = minus4 y = 0006014]
[minus3 00404]
Hybrid+cosHybrid+sin
HybridminuscosHybridminussin
Figure 6 Particles weighed position
Acknowledgments
The authors would like to appreciate the immense contribu-tion of the mining companies where the study was under-taken The authors are grateful to these individuals for
0 1 2 3 4 5
0
05
1
15
2
X 37Y 0989
Particles position
Opt
imise
d er
ror
Nonlinear weight hybrids
Keep large inputvalues more
nonlinear
Keep small inputvalues lessnonlinear
Large input values
Small input values
Using positive cosine and sine
Using negative cosine and sineminus05
minus5 minus4 minus3 minus2 minus1
HSCR-BF+cosHSGR-BF+cos
HSCR-BF+sin
HSCR-BFminuscos
HSCR-BFminuscos
HSCR-BFminussin
HSCR-BF+cos HSCR-BF+sin HSCR-BFminussinand HSGR-BF+cos
Figure 7 Nonlinear weighed curves for optimized error withoutGaussian negative cosine curve
HSGR-BF+cos
HSGR-BF+cos
HSCR-BF+sinHSCR-BFminuscosHSCR-BFminussin
00020000400006000080000
100000120000
0 2 4 6 8 10 12
CPU
tim
e
Time (s)
Nonlinear hybrids
Figure 8 The trend of hybrids CPU time curves using sixth-OrderPolynomial
their immeasurable contributions to this work Fred Atta-kuma Patrick Addai Isaac Owusu Kwankye Thomas KwawAnnan Willet Agongo Nathaniel Quansah Francis OwusuMensah Clement Owusu-Asamoah Joseph Adu-MensahShadrack Aidoo Martin Anokye F T Oduro Ernest Ekow
12 ISRN Artificial Intelligence
Table 9 CPU time usage efficiency for the hybrids
Runs HSCR-BFminus cos HSCR-BF
minus sin HSCR-BF+ cos HSCR-BF
+ sin HSGR-BF+ cos
Total 255889 254350 422344 643962 438133Average 25589 25435 42234 64396 43813
Table 10 Analysis of CPU time on the various hybrids
1205736
1205735
1205734
1205733
1205732
1205731
1205730
1198772
HSCR-BFminus cos 00159119909 minus02575119909 minus0759119909 33494119909 minus19901119909 40345119909 25322 09160
HSCR-BF+ cos 00903119909 minus23158119909 19932119909 minus58027119909 minus31337119909 32808119909 14747 07244
HSCR-BFminus sin minus0033119909 11354119909 minus15811119909 11368119909 minus4337119909 78008119909 minus20794 07551
HSCR-BF+ sin 02571119909 minus14263119909 17173119909 minus10069119909 2993119909 minus42374119909 27879 06731
HSGR-BF+ cos minus00437119909 10911119909 minus82923119909 66218119909 15934119909 minus53939119909 8191 07345
Abano E Omari-Siaw and Qiang Jia This work was sup-ported by the National Natural Science Foundation of China(no 71271103) and by the Six Talents Peak Foundation ofJiangsu Province
References
[1] M O Ansong H X Yao and J S Huang ldquoRadial and sigmoidbasis functions neural networks in wireless sensor routingtopology control in underground mine rescue operation basedon particle swarm optimizationrdquo International Journal of Dis-tributed Sensor Networks-Hindawi vol 2013 Article ID 37693114 pages 2013
[2] R Neruda and P Kudova ldquoLearning methods for radial basisfunction networksrdquo Future Generation Computer Systems vol21 no 7 pp 1131ndash1142 2005
[3] C T Cheng J Y Lin and Y G C Sun ldquoLong-term predictionof discharges in Manwan hydropower using adaptive-network-based fuzzy inference systems modelsrdquo in Advances in NaturalComputation vol 3498 of Lecture Notes in Computer Sciencepp 1152ndash1161 2005
[4] D Broomhead and D Lowe ldquoMultivariate functional interpo-lation and adaptive networksrdquo Complex System vol 2 pp 321ndash355 1988
[5] Y Chen C Chuah and Q Zhao ldquoNetwork configuration foroptimal utilization efficiency of wireless sensor networksrdquo AdHoc Networks vol 6 no 1 pp 92ndash107 2008
[6] P S Rajpal K S Shishodia and G S Sekhon ldquoAn artificialneural network for modeling reliability availability and main-tainability of a repairable systemrdquo Reliability Engineering andSystem Safety vol 91 no 7 pp 809ndash819 2006
[7] W Jang W M Healy and S J Mirosław ldquoWireless sensornetworks as part of a web-based building environmental mon-itoring systemrdquo Automation in Construction vol 17 no 6 pp729ndash736 2008
[8] M S Pan C H Tsai and Y C Tseng ldquoEmergengy guiding andmonitoring application in indoor 3D environment by WirelessSensor Networkrdquo Internation Journal Os Sensor Netwourk vol1 pp 2ndash10 2006
[9] S Zarifzadeh A Nayyeri and N Yazdani ldquoEfficient construc-tion of network topology to conserve energy in wireless ad hocnetworksrdquo Computer Communications vol 31 no 1 pp 160ndash173 2008
[10] R Riaz A Naureen A Akram A H Akbar K Kim and HFarooq Ahmed ldquoA unified security framework with three keymanagement schemes for wireless sensor networksrdquo ComputerCommunications vol 31 no 18 pp 4269ndash4280 2008
[11] S L Goh and D P Mandic ldquoAn augmented CRTRL forcomplex-valued recurrent neural networksrdquo Neural Networksvol 20 no 10 pp 1061ndash1066 2007
[12] L A B Munoz and J J V Ramosy ldquoSimilarity-based heteroge-neous neural networksrdquo Engineering Letters vol 14 no 2 pp103ndash116 2007
[13] R Taormina K Chau and R Sethi ldquoArtificial neural networksimulation of hourly groundwater levels in a coastal aquifer sys-tem of the Venice lagoonrdquo Engineering Applications of ArtificialIntelligence vol 25 no 8 pp 1670ndash1676 2012
[14] K Leblebicioglu and U Halici ldquoInfinite dimensional radialbasis function neural networks for nonlinear transformationson function spacesrdquo Nonlinear Analysis Theory Methods andApplications vol 30 no 3 pp 1649ndash1654 1997
[15] F Fernandez-Navarro C Hervas-Martınez J Sanchez-Monedero and P A Gutierrez ldquoMELM-GRBF a modifiedversion of the extreme learning machine for generalized radialbasis function neural networksrdquo Neurocomputing vol 74 no16 pp 2502ndash2510 2011
[16] C L Wu KW Chau and Y S Li ldquoPredicting monthly stream-flow using data-drivenmodels coupledwith data-preprocessingtechniquesrdquoWater Resources Research vol 45 no 8 Article IDW08432 2009
[17] C EACheng ldquoLong-termprediction of discharges inManwanReservoir using artificial neural network modelsrdquo in Advancesin Neural NetworksmdashISNN 2005 vol 3498 of Lecture Notes inComputer Science pp 1040ndash1045 2005
[18] H H Zhang M Genton and P Liu ldquoCompactly supportedradial basis function kernelsrdquo 2004
[19] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquo inProceedings of the 1995 IEEE International Conference on NeuralNetworks pp 1942ndash1945 December 1995
[20] J Eberhart R Eberhart and Y Shi ldquoParticle swarm optimiza-tion developments applications and resourcesrdquo IEEE vol 1 pp81ndash86 2001
[21] R Malhotra and A Negi ldquoReliability modeling using ParticleSwarm optimisationrdquo International Journal of System AssuranceEngineering and Management vol 4 no 3 pp 275ndash283 2013
[22] D Gies and Y Rahmat-Samii ldquoParticle swarm optimization forreconfigurable phase-differentiated array designrdquo Microwaveand Optical Technology Letters vol 38 no 3 pp 168ndash175 2003
ISRN Artificial Intelligence 13
[23] F Valdez P Melin and C Oscar ldquoAn improved evolution-ary method with fuzzy logic for combining Particle Swarmoptimization and genetic algorithmsrdquo Applied Soft ComputingJournal vol 11 no 2 pp 2625ndash2632 2011
[24] J Zhang and K Chau ldquoMultilayer ensemble pruning vianovel multi-sub-swarm particle swarm optimizationrdquo Journalof Universal Computer Science vol 15 no 4 pp 840ndash858 2009
[25] F Valdez P Melin and O Castillo ldquoParallel Particle Swarmoptimization with parameters adaptation using fuzzy logicrdquoin Proceedings of the 11th Mexican International Conference onArtificial Intelligence (MICAI rsquo12) vol 2 pp 374ndash385 San LuisPotosi Mexico October 2012
[26] J Peng and C P Y Pan ldquoParticle swarm optimization RBFfor gas emission predictionrdquo Journal of Safety Science andTechnology vol 7 pp 77ndash85 2011
[27] G Ren Z Huang Y Cheng X Zhao and Y Zhang ldquoAnintegrated model for evacuation routing and traffic signaloptimization with background demand uncertaintyrdquo Journal ofAdvanced Transportation vol 47 pp 4ndash27 2013
[28] X Feng Z Xiao and X Cui ldquoImproved RSSI algorithm forwireless sensor networks in 3Drdquo Journal of ComputationalInformation Systems vol 7 no 16 pp 5866ndash5873 2011
[29] F Valdez and P Melin ldquoNeural network optimization witha hybrid evolutionary method that combines particle swarmand genetic algorithms with fuzzy rulesrdquo in Proceedings ofthe Annual Meeting of the North American Fuzzy InformationProcessing Society (NAFIPS rsquo08) NewYork NYUSAMay 2008
[30] F Fernandez-Navarro C Hervas-Martınez P A Gutierrez JM Pena-Barragan and F Lopez-Granados ldquoParameter estima-tion of q-Gaussian Radial Basis Functions Neural Networkswith a Hybrid Algorithm for binary classificationrdquo Neurocom-puting vol 75 pp 123ndash134 2012
[31] H Soh S Lim T Zhang et al ldquoWeighted complex networkanalysis of travel routes on the Singapore public transportationsystemrdquo Physica A vol 389 no 24 pp 5852ndash5863 2010
[32] C K S Kumar R Sukumar and M Nageswari ldquoSensorslifetime enhancement techniques in wireless sensor networksmdasha critical reviewrdquo International Journal of Computer Science andInformation Technology amp Security vol 3 pp 159ndash164 2013
[33] S Li and F Qin ldquoA dynamic neural network approach forsolving nonlinear inequalities defined on a graph and itsapplication to distributed routing-free range-free localizationof WSNsrdquo Neurocomputing vol 117 pp 72ndash80 2013
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ISRN Artificial Intelligence 7
Table 2 Optimum set of routing table redundancy with the highest survival probability (6 cases)
Vectors for each case Rock hardness cases1198771 05409 04281 08834 0585 08134 05163 071198772 04526 05841 04547 08997 08436 02696 081198773 02931 05899 05903 08309 06971 02898 0831198774 04851 06085 06913 04002 05026 0 0861198775 04917 05857 07805 06037 05182 06472 08751198776 02902 07253 0458 05762 01397 0 09
2 gives the probability of link available and the last regionindicates the link is completely down (Figure 2)
Thematrices120593119894 120593119896 120593119903 120593119909 120593ℎ for 120591 = 119899 with dimensionsof 119871 = 3 119877 = 2 and119862 = 1 for depth (level) row (length) andwidth (column) respectively were generated as follows
119878eq = (
(
1 1 1
1 2 1
2 2 1
2 1 1
3 1 1
3 2 1
)
)
119894120593 =
[[[[[[[
[
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
]]]]]]]
]
120593119896 = (
(
1 1 1 1 0 0
1 1 1 0 0 0
0 1 1 1 0 0
0 0 1 1 1 0
0 0 0 1 1 1
0 0 0 1 1 1
)
)
1 = signal points
120593119903 = (
(
1 1 1 0 0 0
1 1 1 0 0 0
0 1 1 1 0 0
0 0 1 1 1 0
0 0 0 1 1 1
0 0 0 1 1 1
)
)
1 = transmission path
120593119909 = (
(
0 1 2 2 0 0
1 5 1 2 0 0
0 8 1 5 9 1
1 1 5 1 1 17
1 1 4 11 0 2
4 4 19 5 3 0
)
)
explosion sizes
120593119891 = (
(
1 1 1 1 1 1
1 6 1 1 1 1
1 0 1 6 0 1
1 1 6 1 1 0
1 1 75 0 1 1
75 75 0 6 1 1
)
)
(25)
120593ℎ = (
(
1 1 1 0 0 0
1 6 1 0 0 0
0 0 1 6 0 0
0 0 6 1 1 0
0 0 0 0 1 1
0 0 0 6 1 1
)
)
120593119900 = (
(
8 8 8 0 0 0
8 48 8 0 0 0
0 0 8 48 0 0
0 0 48 8 8 0
0 0 0 0 8 8
0 0 0 48 8 8
)
)
120593119890 = (
(
144 144 144 0 0 0
144 864 144 0 0 0
0 0 144 864 0 0
0 0 864 144 144 0
0 0 0 0 144 144
0 0 0 864 144 144
)
)
1 or fractional figures are transmission path
(26)
Element ldquo0rdquo on 120593119891 depicts a connection lost while ldquo1rdquo meansthat the connection will never go down and represent theconnection to the fixed sink node(s) along the edge or theemergency connection to the mobile data mule(s)
120593s = (
(
8333 8889 9091 8333 8571 8333
8889 8333 8750 8571 8750 8333
8571 9167 8750 9091 8889 8333
8750 8571 8571 8571 8571 8333
8333 8333 8571 8333 8889 9000
9167 8333 9091 8333 8889 8571
)
)
(27)
The survival rate of RFID hardware and each minervectors was displayed as
120593119868 = 8571 8750 8571 8889 8571 8333
120593119867 = 6236 4976 1 6851 9286 6086
120593V = 5409 4281 8834 5850 8134 5163
(28)
Optimization was done numerically usingMatlab simulationtool to find the optimum set of routing tables throughparticle swarm search for rescue operation as discussed inthe preliminaries From (13) the final survival vector fordimensions of 119871 = 3 119877 = 2 and 119862 = 1 will be an average ofthe 6 vectors as in Table 2 Total cases 119877119899 = 120591 = 119879times119879 (which
8 ISRN Artificial Intelligence
Table 3 Scalability of model to survival probability range robot location and rock type
Location Mica Coal Granite Feldspar Quartz MineralCompact radial basis function with negative nonlinear weight (HSCR-BF
minus cos)(10 6 5) 08769ndash10000 09070ndash09927 08402ndash09943 08023ndash09778 07386ndash09376 07345ndash08740(10 5 4) 09024ndash09696 09032ndash09505 08836ndash09430 08633ndash09527 07944ndash09465 07263ndash08720(6 5 4) 08906ndash09877 08827ndash09845 08384ndash09675 08627ndash09820 07911ndash09142 06392ndash07840(3 1 10) 07302ndash08481 06588ndash07816 06341ndash07473 06503ndash07249 05930ndash07156 04445ndash05621
Gaussian radial basis function with positive nonlinear weight (HSGR-BF+ cos)
(10 6 5) 08732ndash09963 08526ndash09987 08492ndash09950 07867ndash09568 07404ndash09025 06606ndash08315(10 5 4) 08843ndash09593 08843ndash09853 08444ndash09753 07746ndash09828 07304ndash09595 06678ndash08817(6 5 4) 08785ndash09983 08902ndash09996 08200ndash09997 07927ndash09446 07065ndash08565 06572ndash08063(3 1 10) 07608ndash08552 06600ndash08325 06090ndash07371 05482ndash06249 05026ndash06056 04856ndash05453
10 20 30 40 50 60 70 80 90 1000
02
04
06
X 94Y 0009904
Number of iterations
Fitn
ess v
alue
PSO optimized error for hybrid SBF-CRBF with negative nonlinear weight
15 2 25 3 35 4 45 5 55 604
06
08
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(HSCRminuscos )
(a) HSCR-BFminus cos
50 100 150 200 2500
01
02
03
X 249Y 001151
PSO optimized error with nonlinear weight [minussine(R)]
Number of iterations
Fitn
ess v
alue
1 2 3 4 5 6
07
08
09
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(b) HSCR-BFminus sin
Figure 3 Hybrid of SBF and CRBF (HSCR-BF) with (a) HSCR-BFminus cos and (b) HSCR-BF
minus sin nonlinear weight
represent the total nodes deployed) for this scenario Rocksoftness or hardness takes the values of 120573 and each matrix isgenerated 120591 timecases and the final routing vector will be anaverage of the 119899 vectors 120593119877 = Average (1198771 1198776) [Table 2]which becomes the input for the PSO training
The final survival rate is 119877 = (05409 04281 08834
0585 08130 05163) from (13)
62 The Final Survival Vector The 119877 = (05409 04281
08834 0585 08130 05163) from (13) is the maximumsurvival probability for a total of 6 nodes deployed Itdescribes the success rate from each node to the sink(s) Inmost practical applications more than one sink is used andsink node is either through the fiber or TTE connection Thesize of the vector depends on the dimensions of the fieldThe elements 119877 = (05409 04281 08834 0585 08130
05163) represent the probability of 54 43 88 5981 and 52 success of each node transmitting data to andfrom its source or destination It assists decision makers as towhether to send data through one ormore nodes or send each
message twice The total nodes used for the simulation were300 with undergroundmine dimensions of 119871 = 10119877 = 6 and119862 = 5 for depth (level) row (length) and width (column)respectively with ldquopmrdquo a sensor apart pm = 100119872
120588
= 4 and119873119890 = 2 The PSO training used swarm size of 20 maximumposition was set to 100 max velocity = 1 and maximumnumber of iteration = 250 The thresholds 120582119871 and 120582119867 were3 and 6 respectively 120591 = 6 cases (each case representa rock hardness case and also represent the total neuronsused) thus each matrix and vectors were run 6 times beforeneural training The survival probability (bottom) indicatesthat the model survived between 90 and 100 where rockcases were relatively soft (ge07) The survival probabilitydeclines as the rock becomes harder and approaches 08 Atthe hardest rock of 09 the survival probability fell between72 and 84 for the entire hybrids In view of this the APheads had to be deployed at a location where the rock isrelatively soft for maximum signal strength Each computersimulation incorporated all the 6 different cases of rockhardnesssoftness to produce the matrices Detailed locationanalysis or the scalability of the model in relation to survival
ISRN Artificial Intelligence 9
Table 4 Hybrid of SBF and CRBF with negative nonlinear weight on CRBF (HSCR-BFminus cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 231020 343498 0382 91650 00985AVG 23102 34350 0038 9165 00098MinMax 29540 45314 0024 80000 00043
Table 5 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BFminus sin)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 307320 488381 0376 114300 0135AVG 30732 48838 0038 11430 0013MinMax 30740 56926 0007 61500 0021
probability range robot location and different rock types isrecorded in Table 3 Figures 4ndash6 represent different scenariosof SBF and CRBF hybrids with the [plusmn cos(119877)] and [plusmn sin(119877)]
nonlinear weights while Figure 7 represents the SBF andGaussian hybrids for [+ cos(119877)] nonlinear weight The tophalf of each figure indicates the optimized error or the finalerror after the neural training and the bottom half revealsmodel survival probability
Figure 3 discusses the SBF and CRBF hybrids (HSCR-BF) with nonlinear weight Training in negative nonlinearweights (HSCR-BF
minus cos HSCR-BFminus sin) responded well The
HSCR-BFminus cos had a steady and compact routing path which
was consistent through all rock layers with initial survivalprobability between 877 and 100 in soft layers decliningto 735ndash874 at harder rock layers The HSCR-BF
minus sinperformed quite well but was more dispersed (879 to 985at the soft rock and 665 to 818 at the hard rock)
In addition the various parameters mean iterationstandard variance standard deviation and the convergenttime of HSCR-BF
minus cos were optimised This is demonstratedby finding the relationship between the maximum and theminimum values of the parameters and comparing with theaverage figures and the closer the difference is to the averagethe more consistent the data are in the dataset (Table 4) thusHSCR-BF
minus cos provided the best results among the proposedhybrids The HSCR-BF
minus sin on the RBF yielded strength inmean iteration and standard variance with error of 0013(Table 5)
In Figure 4 training with nonlinear weight of positivecosine and sine of HSCR-BF (HSCR-BF
+ cos and HSCR-BF+ sin) had some similarities The two models streamed
well at the initial stages from 896 to 992 and 908to 990 at soft layers of the rock respectively Howeverat the last stages the probability of transmitting effectivelybecame marginal from 795 to 852 and 821 to 854respectively Much as this hybrid could perform well in areaswhere routing conditions aremuch better in less dense rescuesituations this could hamper rescue mission in both casesdue to battery drain or collision from traffic conjunctionThedata formean iteration and convergent time forHSCR-BF
+ cos(Table 6) and mean Iteration standard variance the final
error HSCR-BF+ sin (Table 7) were consistent with the nega-
tive sine having better results than the positive cosineThe survival probability of the positive nonlinear weight
Gaussian hybrids (HSGR-BF+ cos) in Figure 5 is more com-
pact both at the initial (ie 873ndash996) and later (661ndash832) stages declining to harder rock layers with an errorof 001173 (Table 8)
At the initial stage particles are sensitive to inputs asthey moved quickly in the search space towards the targetand while particles peaked closer to the target it became lesssensitive to the input and began to align (Figure 6) At thelater stage more RBF are used to keep the error at minimumfor accuracy It was also noticed that instead of acceler-ating higher and becoming more sensitive to the inputswhile particles were far from the target the hybrids withlowast
[minus sin(119877)] and lowast[ + sin(119877)] were less sensitive to the inputsas lowast[minus sin(119877)] descended from [119909 = minus1 119910 = 05785] to[119909 = 1 119910 = 04215] while lowast[+ sin(119877)] descended from [119909 =
minus3 119910 = 00404] to [119909 = minus1 119910 = minus004062] before becomingconscious of the inputs In addition the hybrids with thelowast
[+ cos(119877)] lagged slightly between [119909 = minus4 119910 = 0006014]
and [119909 = minus2 119910 = 006288] before becoming sensitive toinputs However they all lined up for accuracy in terms of theminimised error
The nonlinearity of the hybrids is presented (Figure 7)the negative cosinesine weight is used to reduce nonlinearfor small input values and this lies between minus1 lt 119909 lt +1 andthe positive cosinesine weight is used to keep high nonlinearfor large inputs for the remaining region
63 CPU Time Efficiency The relationship between varioushybrids with respect to the central processing time (CPU)was profiled for different runs (Table 9) and expressed in asixth-order polynomial given as 119884
119903
= 1205737
1198836
1
+ 1205736
1198835
1
+ 1205735
1198834
1
+
1205734
1198833
1
+ 1205733
1198832
1
+ 1205732
1198831
+ 1205730
where 1198831
is time (seconds) and 120573
is the coefficient of the polynomial (Figure 8) The proposedhybrid has better usage of CPU time with 119877
2
= 09160followed byHSCR-BF
minus sin andHSCR-BF+ cos with119877
2 of 07551and 07244 respectively (Table 10) Applying the proposedalgorithm into Gaussian the CPU usage had 119877
2 of 07345
10 ISRN Artificial Intelligence
0 50 100 150 200
002
004
006
008
X 249Y 001103
PSO optimized error for hybrid SBF-CRBF with positive nonlinear weight
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 607
08
09
1
11 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(HSCR+cos )
(a) HSCR-BF+ cos
0 50 100 150 200
005
01
015
02
X 249Y 001247
PSO optimized error with nonlinear weight [+sine(R)]
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 6
070809
111 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(b) HSCR-BF+ sin
Figure 4 Hybrid of SBF and CRBF (HSCR-BF) with (a) HSCR-BF+ cos and (b) HSCR-BF
+ sin nonlinear weight
Table 6 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BF+ cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 350780 522257 0311 115650 0121AVG 35078 52226 0031 11565 0012MaxMin 54840 63413 0009 93500 0006
and indeedmarginally outperformedHSCR-BF+ cos Detailed
work on SBF CRBF and GRBF with regard to scalabilitymemory usage and the central processing time has beencarried by the authors [1]
7 Conclusion
In summary we made the following contributions First weused the mix of SBF and CRBF to present several hybridswith different nonlinear weights of cosine and sine func-tions on compact radial basis function Next we showedthe performance of the proposed nonlinear weight hybridsHSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sin and HSCR-BF+ sin optimised all the parameters with minimised error
of 00098 0012 0013 and 0013 respectively compared to00117 for Gaussian HSCR-BF
+ cos The analyzed CPU usagewith corresponding 119877
2 values of 09160 075 072440 06731and 07345 for HSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sinHSCR-BF
+ sin and HSCR-BF+ cos respectively demonstrated
that the algorithm is scalable There exist some evacuationmodels that is Goh and Mandic [11] which offered a choicefor travelers and several schemes for the decisionmakersmaynot be applicable in emergency undergroundmine situationsThe proposed nonlinear hybrid algorithm with particle swarmoptimisation has better capability of approximation to underly-ing functions with a fast learning speed and high robustness and
is competitive and more computationally efficient to Gaussianwith the same nonlinear weight The algorithm is new andthat makes it difficult to identify limitations and we intend toinvestigate other hybrids and comparewith genetic algorithm(GA) in the future
Nomenclature
119873120575min 119873120575max Minimum and maximum signalreach
HSCR-BFHSGR-BF Hybrid of SBF with CRBFHybrid ofSBF with GRBF
HSCR-BFminus cos Hybrid of SBF and CRBF minuscos
nonlinear weightHSCR-BF
+ cos Hybrid of SBF and CRBF +cosnonlinear weight
HSCR-BFminus sin Hybrid of SBF and CRBF minussine
nonlinear weightHSCR-BF
+ sin Hybrid of SBF and CRBF +sinenonlinear weight
HSGR-BF+ cos Hybrid of SBF and GRBF +cos
nonlinear weight
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
ISRN Artificial Intelligence 11
Table 7 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BF+ sin)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 260360 407109 03457 99800 0129AVG 26036 40711 00346 9980 0013MinMax 27840 56040 0008 106500 0014
Table 8 Hybrid of SBF and GRBF with nonlinear weight on GRBF (HSGR-BF+ cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 219840 349010 0439 112100 01173AVG 21984 34901 0044 11210 00117MinMax 27000 46580 0008 80000 00165
0 50 100 1500
005
01
015
02
X 144Y 0009595
GRBF-PSO optimized error with nonlinear weight [+cos(R)]
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 604
06
08
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
Figure 5 Hybrid of SBF and GRBF with nonlinear weight on GRBF(HSCR-BF
+ cos)
0 2 4 6 8 10
002040608
1121416
Weighed position of particles
Min
imise
d er
ror
Optimised error in nonlinear weighed SBF-CRBF hybridsParticles
Less sensitive to inputsmuch closer to target
more accurate
Less sensitive to inputsmuch closer more accurate
Neurons are more sensitive to inputsfast and far from target
minus02minus2minus4minus6minus8minus10
[x = minus1 y = 05785]
[x = 1 y = 04215]
[minus1 004062]
[x = minus2 y = 006288]
[x = minus4 y = 0006014]
[minus3 00404]
Hybrid+cosHybrid+sin
HybridminuscosHybridminussin
Figure 6 Particles weighed position
Acknowledgments
The authors would like to appreciate the immense contribu-tion of the mining companies where the study was under-taken The authors are grateful to these individuals for
0 1 2 3 4 5
0
05
1
15
2
X 37Y 0989
Particles position
Opt
imise
d er
ror
Nonlinear weight hybrids
Keep large inputvalues more
nonlinear
Keep small inputvalues lessnonlinear
Large input values
Small input values
Using positive cosine and sine
Using negative cosine and sineminus05
minus5 minus4 minus3 minus2 minus1
HSCR-BF+cosHSGR-BF+cos
HSCR-BF+sin
HSCR-BFminuscos
HSCR-BFminuscos
HSCR-BFminussin
HSCR-BF+cos HSCR-BF+sin HSCR-BFminussinand HSGR-BF+cos
Figure 7 Nonlinear weighed curves for optimized error withoutGaussian negative cosine curve
HSGR-BF+cos
HSGR-BF+cos
HSCR-BF+sinHSCR-BFminuscosHSCR-BFminussin
00020000400006000080000
100000120000
0 2 4 6 8 10 12
CPU
tim
e
Time (s)
Nonlinear hybrids
Figure 8 The trend of hybrids CPU time curves using sixth-OrderPolynomial
their immeasurable contributions to this work Fred Atta-kuma Patrick Addai Isaac Owusu Kwankye Thomas KwawAnnan Willet Agongo Nathaniel Quansah Francis OwusuMensah Clement Owusu-Asamoah Joseph Adu-MensahShadrack Aidoo Martin Anokye F T Oduro Ernest Ekow
12 ISRN Artificial Intelligence
Table 9 CPU time usage efficiency for the hybrids
Runs HSCR-BFminus cos HSCR-BF
minus sin HSCR-BF+ cos HSCR-BF
+ sin HSGR-BF+ cos
Total 255889 254350 422344 643962 438133Average 25589 25435 42234 64396 43813
Table 10 Analysis of CPU time on the various hybrids
1205736
1205735
1205734
1205733
1205732
1205731
1205730
1198772
HSCR-BFminus cos 00159119909 minus02575119909 minus0759119909 33494119909 minus19901119909 40345119909 25322 09160
HSCR-BF+ cos 00903119909 minus23158119909 19932119909 minus58027119909 minus31337119909 32808119909 14747 07244
HSCR-BFminus sin minus0033119909 11354119909 minus15811119909 11368119909 minus4337119909 78008119909 minus20794 07551
HSCR-BF+ sin 02571119909 minus14263119909 17173119909 minus10069119909 2993119909 minus42374119909 27879 06731
HSGR-BF+ cos minus00437119909 10911119909 minus82923119909 66218119909 15934119909 minus53939119909 8191 07345
Abano E Omari-Siaw and Qiang Jia This work was sup-ported by the National Natural Science Foundation of China(no 71271103) and by the Six Talents Peak Foundation ofJiangsu Province
References
[1] M O Ansong H X Yao and J S Huang ldquoRadial and sigmoidbasis functions neural networks in wireless sensor routingtopology control in underground mine rescue operation basedon particle swarm optimizationrdquo International Journal of Dis-tributed Sensor Networks-Hindawi vol 2013 Article ID 37693114 pages 2013
[2] R Neruda and P Kudova ldquoLearning methods for radial basisfunction networksrdquo Future Generation Computer Systems vol21 no 7 pp 1131ndash1142 2005
[3] C T Cheng J Y Lin and Y G C Sun ldquoLong-term predictionof discharges in Manwan hydropower using adaptive-network-based fuzzy inference systems modelsrdquo in Advances in NaturalComputation vol 3498 of Lecture Notes in Computer Sciencepp 1152ndash1161 2005
[4] D Broomhead and D Lowe ldquoMultivariate functional interpo-lation and adaptive networksrdquo Complex System vol 2 pp 321ndash355 1988
[5] Y Chen C Chuah and Q Zhao ldquoNetwork configuration foroptimal utilization efficiency of wireless sensor networksrdquo AdHoc Networks vol 6 no 1 pp 92ndash107 2008
[6] P S Rajpal K S Shishodia and G S Sekhon ldquoAn artificialneural network for modeling reliability availability and main-tainability of a repairable systemrdquo Reliability Engineering andSystem Safety vol 91 no 7 pp 809ndash819 2006
[7] W Jang W M Healy and S J Mirosław ldquoWireless sensornetworks as part of a web-based building environmental mon-itoring systemrdquo Automation in Construction vol 17 no 6 pp729ndash736 2008
[8] M S Pan C H Tsai and Y C Tseng ldquoEmergengy guiding andmonitoring application in indoor 3D environment by WirelessSensor Networkrdquo Internation Journal Os Sensor Netwourk vol1 pp 2ndash10 2006
[9] S Zarifzadeh A Nayyeri and N Yazdani ldquoEfficient construc-tion of network topology to conserve energy in wireless ad hocnetworksrdquo Computer Communications vol 31 no 1 pp 160ndash173 2008
[10] R Riaz A Naureen A Akram A H Akbar K Kim and HFarooq Ahmed ldquoA unified security framework with three keymanagement schemes for wireless sensor networksrdquo ComputerCommunications vol 31 no 18 pp 4269ndash4280 2008
[11] S L Goh and D P Mandic ldquoAn augmented CRTRL forcomplex-valued recurrent neural networksrdquo Neural Networksvol 20 no 10 pp 1061ndash1066 2007
[12] L A B Munoz and J J V Ramosy ldquoSimilarity-based heteroge-neous neural networksrdquo Engineering Letters vol 14 no 2 pp103ndash116 2007
[13] R Taormina K Chau and R Sethi ldquoArtificial neural networksimulation of hourly groundwater levels in a coastal aquifer sys-tem of the Venice lagoonrdquo Engineering Applications of ArtificialIntelligence vol 25 no 8 pp 1670ndash1676 2012
[14] K Leblebicioglu and U Halici ldquoInfinite dimensional radialbasis function neural networks for nonlinear transformationson function spacesrdquo Nonlinear Analysis Theory Methods andApplications vol 30 no 3 pp 1649ndash1654 1997
[15] F Fernandez-Navarro C Hervas-Martınez J Sanchez-Monedero and P A Gutierrez ldquoMELM-GRBF a modifiedversion of the extreme learning machine for generalized radialbasis function neural networksrdquo Neurocomputing vol 74 no16 pp 2502ndash2510 2011
[16] C L Wu KW Chau and Y S Li ldquoPredicting monthly stream-flow using data-drivenmodels coupledwith data-preprocessingtechniquesrdquoWater Resources Research vol 45 no 8 Article IDW08432 2009
[17] C EACheng ldquoLong-termprediction of discharges inManwanReservoir using artificial neural network modelsrdquo in Advancesin Neural NetworksmdashISNN 2005 vol 3498 of Lecture Notes inComputer Science pp 1040ndash1045 2005
[18] H H Zhang M Genton and P Liu ldquoCompactly supportedradial basis function kernelsrdquo 2004
[19] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquo inProceedings of the 1995 IEEE International Conference on NeuralNetworks pp 1942ndash1945 December 1995
[20] J Eberhart R Eberhart and Y Shi ldquoParticle swarm optimiza-tion developments applications and resourcesrdquo IEEE vol 1 pp81ndash86 2001
[21] R Malhotra and A Negi ldquoReliability modeling using ParticleSwarm optimisationrdquo International Journal of System AssuranceEngineering and Management vol 4 no 3 pp 275ndash283 2013
[22] D Gies and Y Rahmat-Samii ldquoParticle swarm optimization forreconfigurable phase-differentiated array designrdquo Microwaveand Optical Technology Letters vol 38 no 3 pp 168ndash175 2003
ISRN Artificial Intelligence 13
[23] F Valdez P Melin and C Oscar ldquoAn improved evolution-ary method with fuzzy logic for combining Particle Swarmoptimization and genetic algorithmsrdquo Applied Soft ComputingJournal vol 11 no 2 pp 2625ndash2632 2011
[24] J Zhang and K Chau ldquoMultilayer ensemble pruning vianovel multi-sub-swarm particle swarm optimizationrdquo Journalof Universal Computer Science vol 15 no 4 pp 840ndash858 2009
[25] F Valdez P Melin and O Castillo ldquoParallel Particle Swarmoptimization with parameters adaptation using fuzzy logicrdquoin Proceedings of the 11th Mexican International Conference onArtificial Intelligence (MICAI rsquo12) vol 2 pp 374ndash385 San LuisPotosi Mexico October 2012
[26] J Peng and C P Y Pan ldquoParticle swarm optimization RBFfor gas emission predictionrdquo Journal of Safety Science andTechnology vol 7 pp 77ndash85 2011
[27] G Ren Z Huang Y Cheng X Zhao and Y Zhang ldquoAnintegrated model for evacuation routing and traffic signaloptimization with background demand uncertaintyrdquo Journal ofAdvanced Transportation vol 47 pp 4ndash27 2013
[28] X Feng Z Xiao and X Cui ldquoImproved RSSI algorithm forwireless sensor networks in 3Drdquo Journal of ComputationalInformation Systems vol 7 no 16 pp 5866ndash5873 2011
[29] F Valdez and P Melin ldquoNeural network optimization witha hybrid evolutionary method that combines particle swarmand genetic algorithms with fuzzy rulesrdquo in Proceedings ofthe Annual Meeting of the North American Fuzzy InformationProcessing Society (NAFIPS rsquo08) NewYork NYUSAMay 2008
[30] F Fernandez-Navarro C Hervas-Martınez P A Gutierrez JM Pena-Barragan and F Lopez-Granados ldquoParameter estima-tion of q-Gaussian Radial Basis Functions Neural Networkswith a Hybrid Algorithm for binary classificationrdquo Neurocom-puting vol 75 pp 123ndash134 2012
[31] H Soh S Lim T Zhang et al ldquoWeighted complex networkanalysis of travel routes on the Singapore public transportationsystemrdquo Physica A vol 389 no 24 pp 5852ndash5863 2010
[32] C K S Kumar R Sukumar and M Nageswari ldquoSensorslifetime enhancement techniques in wireless sensor networksmdasha critical reviewrdquo International Journal of Computer Science andInformation Technology amp Security vol 3 pp 159ndash164 2013
[33] S Li and F Qin ldquoA dynamic neural network approach forsolving nonlinear inequalities defined on a graph and itsapplication to distributed routing-free range-free localizationof WSNsrdquo Neurocomputing vol 117 pp 72ndash80 2013
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
8 ISRN Artificial Intelligence
Table 3 Scalability of model to survival probability range robot location and rock type
Location Mica Coal Granite Feldspar Quartz MineralCompact radial basis function with negative nonlinear weight (HSCR-BF
minus cos)(10 6 5) 08769ndash10000 09070ndash09927 08402ndash09943 08023ndash09778 07386ndash09376 07345ndash08740(10 5 4) 09024ndash09696 09032ndash09505 08836ndash09430 08633ndash09527 07944ndash09465 07263ndash08720(6 5 4) 08906ndash09877 08827ndash09845 08384ndash09675 08627ndash09820 07911ndash09142 06392ndash07840(3 1 10) 07302ndash08481 06588ndash07816 06341ndash07473 06503ndash07249 05930ndash07156 04445ndash05621
Gaussian radial basis function with positive nonlinear weight (HSGR-BF+ cos)
(10 6 5) 08732ndash09963 08526ndash09987 08492ndash09950 07867ndash09568 07404ndash09025 06606ndash08315(10 5 4) 08843ndash09593 08843ndash09853 08444ndash09753 07746ndash09828 07304ndash09595 06678ndash08817(6 5 4) 08785ndash09983 08902ndash09996 08200ndash09997 07927ndash09446 07065ndash08565 06572ndash08063(3 1 10) 07608ndash08552 06600ndash08325 06090ndash07371 05482ndash06249 05026ndash06056 04856ndash05453
10 20 30 40 50 60 70 80 90 1000
02
04
06
X 94Y 0009904
Number of iterations
Fitn
ess v
alue
PSO optimized error for hybrid SBF-CRBF with negative nonlinear weight
15 2 25 3 35 4 45 5 55 604
06
08
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(HSCRminuscos )
(a) HSCR-BFminus cos
50 100 150 200 2500
01
02
03
X 249Y 001151
PSO optimized error with nonlinear weight [minussine(R)]
Number of iterations
Fitn
ess v
alue
1 2 3 4 5 6
07
08
09
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(b) HSCR-BFminus sin
Figure 3 Hybrid of SBF and CRBF (HSCR-BF) with (a) HSCR-BFminus cos and (b) HSCR-BF
minus sin nonlinear weight
represent the total nodes deployed) for this scenario Rocksoftness or hardness takes the values of 120573 and each matrix isgenerated 120591 timecases and the final routing vector will be anaverage of the 119899 vectors 120593119877 = Average (1198771 1198776) [Table 2]which becomes the input for the PSO training
The final survival rate is 119877 = (05409 04281 08834
0585 08130 05163) from (13)
62 The Final Survival Vector The 119877 = (05409 04281
08834 0585 08130 05163) from (13) is the maximumsurvival probability for a total of 6 nodes deployed Itdescribes the success rate from each node to the sink(s) Inmost practical applications more than one sink is used andsink node is either through the fiber or TTE connection Thesize of the vector depends on the dimensions of the fieldThe elements 119877 = (05409 04281 08834 0585 08130
05163) represent the probability of 54 43 88 5981 and 52 success of each node transmitting data to andfrom its source or destination It assists decision makers as towhether to send data through one ormore nodes or send each
message twice The total nodes used for the simulation were300 with undergroundmine dimensions of 119871 = 10119877 = 6 and119862 = 5 for depth (level) row (length) and width (column)respectively with ldquopmrdquo a sensor apart pm = 100119872
120588
= 4 and119873119890 = 2 The PSO training used swarm size of 20 maximumposition was set to 100 max velocity = 1 and maximumnumber of iteration = 250 The thresholds 120582119871 and 120582119867 were3 and 6 respectively 120591 = 6 cases (each case representa rock hardness case and also represent the total neuronsused) thus each matrix and vectors were run 6 times beforeneural training The survival probability (bottom) indicatesthat the model survived between 90 and 100 where rockcases were relatively soft (ge07) The survival probabilitydeclines as the rock becomes harder and approaches 08 Atthe hardest rock of 09 the survival probability fell between72 and 84 for the entire hybrids In view of this the APheads had to be deployed at a location where the rock isrelatively soft for maximum signal strength Each computersimulation incorporated all the 6 different cases of rockhardnesssoftness to produce the matrices Detailed locationanalysis or the scalability of the model in relation to survival
ISRN Artificial Intelligence 9
Table 4 Hybrid of SBF and CRBF with negative nonlinear weight on CRBF (HSCR-BFminus cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 231020 343498 0382 91650 00985AVG 23102 34350 0038 9165 00098MinMax 29540 45314 0024 80000 00043
Table 5 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BFminus sin)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 307320 488381 0376 114300 0135AVG 30732 48838 0038 11430 0013MinMax 30740 56926 0007 61500 0021
probability range robot location and different rock types isrecorded in Table 3 Figures 4ndash6 represent different scenariosof SBF and CRBF hybrids with the [plusmn cos(119877)] and [plusmn sin(119877)]
nonlinear weights while Figure 7 represents the SBF andGaussian hybrids for [+ cos(119877)] nonlinear weight The tophalf of each figure indicates the optimized error or the finalerror after the neural training and the bottom half revealsmodel survival probability
Figure 3 discusses the SBF and CRBF hybrids (HSCR-BF) with nonlinear weight Training in negative nonlinearweights (HSCR-BF
minus cos HSCR-BFminus sin) responded well The
HSCR-BFminus cos had a steady and compact routing path which
was consistent through all rock layers with initial survivalprobability between 877 and 100 in soft layers decliningto 735ndash874 at harder rock layers The HSCR-BF
minus sinperformed quite well but was more dispersed (879 to 985at the soft rock and 665 to 818 at the hard rock)
In addition the various parameters mean iterationstandard variance standard deviation and the convergenttime of HSCR-BF
minus cos were optimised This is demonstratedby finding the relationship between the maximum and theminimum values of the parameters and comparing with theaverage figures and the closer the difference is to the averagethe more consistent the data are in the dataset (Table 4) thusHSCR-BF
minus cos provided the best results among the proposedhybrids The HSCR-BF
minus sin on the RBF yielded strength inmean iteration and standard variance with error of 0013(Table 5)
In Figure 4 training with nonlinear weight of positivecosine and sine of HSCR-BF (HSCR-BF
+ cos and HSCR-BF+ sin) had some similarities The two models streamed
well at the initial stages from 896 to 992 and 908to 990 at soft layers of the rock respectively Howeverat the last stages the probability of transmitting effectivelybecame marginal from 795 to 852 and 821 to 854respectively Much as this hybrid could perform well in areaswhere routing conditions aremuch better in less dense rescuesituations this could hamper rescue mission in both casesdue to battery drain or collision from traffic conjunctionThedata formean iteration and convergent time forHSCR-BF
+ cos(Table 6) and mean Iteration standard variance the final
error HSCR-BF+ sin (Table 7) were consistent with the nega-
tive sine having better results than the positive cosineThe survival probability of the positive nonlinear weight
Gaussian hybrids (HSGR-BF+ cos) in Figure 5 is more com-
pact both at the initial (ie 873ndash996) and later (661ndash832) stages declining to harder rock layers with an errorof 001173 (Table 8)
At the initial stage particles are sensitive to inputs asthey moved quickly in the search space towards the targetand while particles peaked closer to the target it became lesssensitive to the input and began to align (Figure 6) At thelater stage more RBF are used to keep the error at minimumfor accuracy It was also noticed that instead of acceler-ating higher and becoming more sensitive to the inputswhile particles were far from the target the hybrids withlowast
[minus sin(119877)] and lowast[ + sin(119877)] were less sensitive to the inputsas lowast[minus sin(119877)] descended from [119909 = minus1 119910 = 05785] to[119909 = 1 119910 = 04215] while lowast[+ sin(119877)] descended from [119909 =
minus3 119910 = 00404] to [119909 = minus1 119910 = minus004062] before becomingconscious of the inputs In addition the hybrids with thelowast
[+ cos(119877)] lagged slightly between [119909 = minus4 119910 = 0006014]
and [119909 = minus2 119910 = 006288] before becoming sensitive toinputs However they all lined up for accuracy in terms of theminimised error
The nonlinearity of the hybrids is presented (Figure 7)the negative cosinesine weight is used to reduce nonlinearfor small input values and this lies between minus1 lt 119909 lt +1 andthe positive cosinesine weight is used to keep high nonlinearfor large inputs for the remaining region
63 CPU Time Efficiency The relationship between varioushybrids with respect to the central processing time (CPU)was profiled for different runs (Table 9) and expressed in asixth-order polynomial given as 119884
119903
= 1205737
1198836
1
+ 1205736
1198835
1
+ 1205735
1198834
1
+
1205734
1198833
1
+ 1205733
1198832
1
+ 1205732
1198831
+ 1205730
where 1198831
is time (seconds) and 120573
is the coefficient of the polynomial (Figure 8) The proposedhybrid has better usage of CPU time with 119877
2
= 09160followed byHSCR-BF
minus sin andHSCR-BF+ cos with119877
2 of 07551and 07244 respectively (Table 10) Applying the proposedalgorithm into Gaussian the CPU usage had 119877
2 of 07345
10 ISRN Artificial Intelligence
0 50 100 150 200
002
004
006
008
X 249Y 001103
PSO optimized error for hybrid SBF-CRBF with positive nonlinear weight
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 607
08
09
1
11 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(HSCR+cos )
(a) HSCR-BF+ cos
0 50 100 150 200
005
01
015
02
X 249Y 001247
PSO optimized error with nonlinear weight [+sine(R)]
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 6
070809
111 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(b) HSCR-BF+ sin
Figure 4 Hybrid of SBF and CRBF (HSCR-BF) with (a) HSCR-BF+ cos and (b) HSCR-BF
+ sin nonlinear weight
Table 6 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BF+ cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 350780 522257 0311 115650 0121AVG 35078 52226 0031 11565 0012MaxMin 54840 63413 0009 93500 0006
and indeedmarginally outperformedHSCR-BF+ cos Detailed
work on SBF CRBF and GRBF with regard to scalabilitymemory usage and the central processing time has beencarried by the authors [1]
7 Conclusion
In summary we made the following contributions First weused the mix of SBF and CRBF to present several hybridswith different nonlinear weights of cosine and sine func-tions on compact radial basis function Next we showedthe performance of the proposed nonlinear weight hybridsHSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sin and HSCR-BF+ sin optimised all the parameters with minimised error
of 00098 0012 0013 and 0013 respectively compared to00117 for Gaussian HSCR-BF
+ cos The analyzed CPU usagewith corresponding 119877
2 values of 09160 075 072440 06731and 07345 for HSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sinHSCR-BF
+ sin and HSCR-BF+ cos respectively demonstrated
that the algorithm is scalable There exist some evacuationmodels that is Goh and Mandic [11] which offered a choicefor travelers and several schemes for the decisionmakersmaynot be applicable in emergency undergroundmine situationsThe proposed nonlinear hybrid algorithm with particle swarmoptimisation has better capability of approximation to underly-ing functions with a fast learning speed and high robustness and
is competitive and more computationally efficient to Gaussianwith the same nonlinear weight The algorithm is new andthat makes it difficult to identify limitations and we intend toinvestigate other hybrids and comparewith genetic algorithm(GA) in the future
Nomenclature
119873120575min 119873120575max Minimum and maximum signalreach
HSCR-BFHSGR-BF Hybrid of SBF with CRBFHybrid ofSBF with GRBF
HSCR-BFminus cos Hybrid of SBF and CRBF minuscos
nonlinear weightHSCR-BF
+ cos Hybrid of SBF and CRBF +cosnonlinear weight
HSCR-BFminus sin Hybrid of SBF and CRBF minussine
nonlinear weightHSCR-BF
+ sin Hybrid of SBF and CRBF +sinenonlinear weight
HSGR-BF+ cos Hybrid of SBF and GRBF +cos
nonlinear weight
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
ISRN Artificial Intelligence 11
Table 7 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BF+ sin)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 260360 407109 03457 99800 0129AVG 26036 40711 00346 9980 0013MinMax 27840 56040 0008 106500 0014
Table 8 Hybrid of SBF and GRBF with nonlinear weight on GRBF (HSGR-BF+ cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 219840 349010 0439 112100 01173AVG 21984 34901 0044 11210 00117MinMax 27000 46580 0008 80000 00165
0 50 100 1500
005
01
015
02
X 144Y 0009595
GRBF-PSO optimized error with nonlinear weight [+cos(R)]
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 604
06
08
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
Figure 5 Hybrid of SBF and GRBF with nonlinear weight on GRBF(HSCR-BF
+ cos)
0 2 4 6 8 10
002040608
1121416
Weighed position of particles
Min
imise
d er
ror
Optimised error in nonlinear weighed SBF-CRBF hybridsParticles
Less sensitive to inputsmuch closer to target
more accurate
Less sensitive to inputsmuch closer more accurate
Neurons are more sensitive to inputsfast and far from target
minus02minus2minus4minus6minus8minus10
[x = minus1 y = 05785]
[x = 1 y = 04215]
[minus1 004062]
[x = minus2 y = 006288]
[x = minus4 y = 0006014]
[minus3 00404]
Hybrid+cosHybrid+sin
HybridminuscosHybridminussin
Figure 6 Particles weighed position
Acknowledgments
The authors would like to appreciate the immense contribu-tion of the mining companies where the study was under-taken The authors are grateful to these individuals for
0 1 2 3 4 5
0
05
1
15
2
X 37Y 0989
Particles position
Opt
imise
d er
ror
Nonlinear weight hybrids
Keep large inputvalues more
nonlinear
Keep small inputvalues lessnonlinear
Large input values
Small input values
Using positive cosine and sine
Using negative cosine and sineminus05
minus5 minus4 minus3 minus2 minus1
HSCR-BF+cosHSGR-BF+cos
HSCR-BF+sin
HSCR-BFminuscos
HSCR-BFminuscos
HSCR-BFminussin
HSCR-BF+cos HSCR-BF+sin HSCR-BFminussinand HSGR-BF+cos
Figure 7 Nonlinear weighed curves for optimized error withoutGaussian negative cosine curve
HSGR-BF+cos
HSGR-BF+cos
HSCR-BF+sinHSCR-BFminuscosHSCR-BFminussin
00020000400006000080000
100000120000
0 2 4 6 8 10 12
CPU
tim
e
Time (s)
Nonlinear hybrids
Figure 8 The trend of hybrids CPU time curves using sixth-OrderPolynomial
their immeasurable contributions to this work Fred Atta-kuma Patrick Addai Isaac Owusu Kwankye Thomas KwawAnnan Willet Agongo Nathaniel Quansah Francis OwusuMensah Clement Owusu-Asamoah Joseph Adu-MensahShadrack Aidoo Martin Anokye F T Oduro Ernest Ekow
12 ISRN Artificial Intelligence
Table 9 CPU time usage efficiency for the hybrids
Runs HSCR-BFminus cos HSCR-BF
minus sin HSCR-BF+ cos HSCR-BF
+ sin HSGR-BF+ cos
Total 255889 254350 422344 643962 438133Average 25589 25435 42234 64396 43813
Table 10 Analysis of CPU time on the various hybrids
1205736
1205735
1205734
1205733
1205732
1205731
1205730
1198772
HSCR-BFminus cos 00159119909 minus02575119909 minus0759119909 33494119909 minus19901119909 40345119909 25322 09160
HSCR-BF+ cos 00903119909 minus23158119909 19932119909 minus58027119909 minus31337119909 32808119909 14747 07244
HSCR-BFminus sin minus0033119909 11354119909 minus15811119909 11368119909 minus4337119909 78008119909 minus20794 07551
HSCR-BF+ sin 02571119909 minus14263119909 17173119909 minus10069119909 2993119909 minus42374119909 27879 06731
HSGR-BF+ cos minus00437119909 10911119909 minus82923119909 66218119909 15934119909 minus53939119909 8191 07345
Abano E Omari-Siaw and Qiang Jia This work was sup-ported by the National Natural Science Foundation of China(no 71271103) and by the Six Talents Peak Foundation ofJiangsu Province
References
[1] M O Ansong H X Yao and J S Huang ldquoRadial and sigmoidbasis functions neural networks in wireless sensor routingtopology control in underground mine rescue operation basedon particle swarm optimizationrdquo International Journal of Dis-tributed Sensor Networks-Hindawi vol 2013 Article ID 37693114 pages 2013
[2] R Neruda and P Kudova ldquoLearning methods for radial basisfunction networksrdquo Future Generation Computer Systems vol21 no 7 pp 1131ndash1142 2005
[3] C T Cheng J Y Lin and Y G C Sun ldquoLong-term predictionof discharges in Manwan hydropower using adaptive-network-based fuzzy inference systems modelsrdquo in Advances in NaturalComputation vol 3498 of Lecture Notes in Computer Sciencepp 1152ndash1161 2005
[4] D Broomhead and D Lowe ldquoMultivariate functional interpo-lation and adaptive networksrdquo Complex System vol 2 pp 321ndash355 1988
[5] Y Chen C Chuah and Q Zhao ldquoNetwork configuration foroptimal utilization efficiency of wireless sensor networksrdquo AdHoc Networks vol 6 no 1 pp 92ndash107 2008
[6] P S Rajpal K S Shishodia and G S Sekhon ldquoAn artificialneural network for modeling reliability availability and main-tainability of a repairable systemrdquo Reliability Engineering andSystem Safety vol 91 no 7 pp 809ndash819 2006
[7] W Jang W M Healy and S J Mirosław ldquoWireless sensornetworks as part of a web-based building environmental mon-itoring systemrdquo Automation in Construction vol 17 no 6 pp729ndash736 2008
[8] M S Pan C H Tsai and Y C Tseng ldquoEmergengy guiding andmonitoring application in indoor 3D environment by WirelessSensor Networkrdquo Internation Journal Os Sensor Netwourk vol1 pp 2ndash10 2006
[9] S Zarifzadeh A Nayyeri and N Yazdani ldquoEfficient construc-tion of network topology to conserve energy in wireless ad hocnetworksrdquo Computer Communications vol 31 no 1 pp 160ndash173 2008
[10] R Riaz A Naureen A Akram A H Akbar K Kim and HFarooq Ahmed ldquoA unified security framework with three keymanagement schemes for wireless sensor networksrdquo ComputerCommunications vol 31 no 18 pp 4269ndash4280 2008
[11] S L Goh and D P Mandic ldquoAn augmented CRTRL forcomplex-valued recurrent neural networksrdquo Neural Networksvol 20 no 10 pp 1061ndash1066 2007
[12] L A B Munoz and J J V Ramosy ldquoSimilarity-based heteroge-neous neural networksrdquo Engineering Letters vol 14 no 2 pp103ndash116 2007
[13] R Taormina K Chau and R Sethi ldquoArtificial neural networksimulation of hourly groundwater levels in a coastal aquifer sys-tem of the Venice lagoonrdquo Engineering Applications of ArtificialIntelligence vol 25 no 8 pp 1670ndash1676 2012
[14] K Leblebicioglu and U Halici ldquoInfinite dimensional radialbasis function neural networks for nonlinear transformationson function spacesrdquo Nonlinear Analysis Theory Methods andApplications vol 30 no 3 pp 1649ndash1654 1997
[15] F Fernandez-Navarro C Hervas-Martınez J Sanchez-Monedero and P A Gutierrez ldquoMELM-GRBF a modifiedversion of the extreme learning machine for generalized radialbasis function neural networksrdquo Neurocomputing vol 74 no16 pp 2502ndash2510 2011
[16] C L Wu KW Chau and Y S Li ldquoPredicting monthly stream-flow using data-drivenmodels coupledwith data-preprocessingtechniquesrdquoWater Resources Research vol 45 no 8 Article IDW08432 2009
[17] C EACheng ldquoLong-termprediction of discharges inManwanReservoir using artificial neural network modelsrdquo in Advancesin Neural NetworksmdashISNN 2005 vol 3498 of Lecture Notes inComputer Science pp 1040ndash1045 2005
[18] H H Zhang M Genton and P Liu ldquoCompactly supportedradial basis function kernelsrdquo 2004
[19] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquo inProceedings of the 1995 IEEE International Conference on NeuralNetworks pp 1942ndash1945 December 1995
[20] J Eberhart R Eberhart and Y Shi ldquoParticle swarm optimiza-tion developments applications and resourcesrdquo IEEE vol 1 pp81ndash86 2001
[21] R Malhotra and A Negi ldquoReliability modeling using ParticleSwarm optimisationrdquo International Journal of System AssuranceEngineering and Management vol 4 no 3 pp 275ndash283 2013
[22] D Gies and Y Rahmat-Samii ldquoParticle swarm optimization forreconfigurable phase-differentiated array designrdquo Microwaveand Optical Technology Letters vol 38 no 3 pp 168ndash175 2003
ISRN Artificial Intelligence 13
[23] F Valdez P Melin and C Oscar ldquoAn improved evolution-ary method with fuzzy logic for combining Particle Swarmoptimization and genetic algorithmsrdquo Applied Soft ComputingJournal vol 11 no 2 pp 2625ndash2632 2011
[24] J Zhang and K Chau ldquoMultilayer ensemble pruning vianovel multi-sub-swarm particle swarm optimizationrdquo Journalof Universal Computer Science vol 15 no 4 pp 840ndash858 2009
[25] F Valdez P Melin and O Castillo ldquoParallel Particle Swarmoptimization with parameters adaptation using fuzzy logicrdquoin Proceedings of the 11th Mexican International Conference onArtificial Intelligence (MICAI rsquo12) vol 2 pp 374ndash385 San LuisPotosi Mexico October 2012
[26] J Peng and C P Y Pan ldquoParticle swarm optimization RBFfor gas emission predictionrdquo Journal of Safety Science andTechnology vol 7 pp 77ndash85 2011
[27] G Ren Z Huang Y Cheng X Zhao and Y Zhang ldquoAnintegrated model for evacuation routing and traffic signaloptimization with background demand uncertaintyrdquo Journal ofAdvanced Transportation vol 47 pp 4ndash27 2013
[28] X Feng Z Xiao and X Cui ldquoImproved RSSI algorithm forwireless sensor networks in 3Drdquo Journal of ComputationalInformation Systems vol 7 no 16 pp 5866ndash5873 2011
[29] F Valdez and P Melin ldquoNeural network optimization witha hybrid evolutionary method that combines particle swarmand genetic algorithms with fuzzy rulesrdquo in Proceedings ofthe Annual Meeting of the North American Fuzzy InformationProcessing Society (NAFIPS rsquo08) NewYork NYUSAMay 2008
[30] F Fernandez-Navarro C Hervas-Martınez P A Gutierrez JM Pena-Barragan and F Lopez-Granados ldquoParameter estima-tion of q-Gaussian Radial Basis Functions Neural Networkswith a Hybrid Algorithm for binary classificationrdquo Neurocom-puting vol 75 pp 123ndash134 2012
[31] H Soh S Lim T Zhang et al ldquoWeighted complex networkanalysis of travel routes on the Singapore public transportationsystemrdquo Physica A vol 389 no 24 pp 5852ndash5863 2010
[32] C K S Kumar R Sukumar and M Nageswari ldquoSensorslifetime enhancement techniques in wireless sensor networksmdasha critical reviewrdquo International Journal of Computer Science andInformation Technology amp Security vol 3 pp 159ndash164 2013
[33] S Li and F Qin ldquoA dynamic neural network approach forsolving nonlinear inequalities defined on a graph and itsapplication to distributed routing-free range-free localizationof WSNsrdquo Neurocomputing vol 117 pp 72ndash80 2013
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ISRN Artificial Intelligence 9
Table 4 Hybrid of SBF and CRBF with negative nonlinear weight on CRBF (HSCR-BFminus cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 231020 343498 0382 91650 00985AVG 23102 34350 0038 9165 00098MinMax 29540 45314 0024 80000 00043
Table 5 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BFminus sin)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 307320 488381 0376 114300 0135AVG 30732 48838 0038 11430 0013MinMax 30740 56926 0007 61500 0021
probability range robot location and different rock types isrecorded in Table 3 Figures 4ndash6 represent different scenariosof SBF and CRBF hybrids with the [plusmn cos(119877)] and [plusmn sin(119877)]
nonlinear weights while Figure 7 represents the SBF andGaussian hybrids for [+ cos(119877)] nonlinear weight The tophalf of each figure indicates the optimized error or the finalerror after the neural training and the bottom half revealsmodel survival probability
Figure 3 discusses the SBF and CRBF hybrids (HSCR-BF) with nonlinear weight Training in negative nonlinearweights (HSCR-BF
minus cos HSCR-BFminus sin) responded well The
HSCR-BFminus cos had a steady and compact routing path which
was consistent through all rock layers with initial survivalprobability between 877 and 100 in soft layers decliningto 735ndash874 at harder rock layers The HSCR-BF
minus sinperformed quite well but was more dispersed (879 to 985at the soft rock and 665 to 818 at the hard rock)
In addition the various parameters mean iterationstandard variance standard deviation and the convergenttime of HSCR-BF
minus cos were optimised This is demonstratedby finding the relationship between the maximum and theminimum values of the parameters and comparing with theaverage figures and the closer the difference is to the averagethe more consistent the data are in the dataset (Table 4) thusHSCR-BF
minus cos provided the best results among the proposedhybrids The HSCR-BF
minus sin on the RBF yielded strength inmean iteration and standard variance with error of 0013(Table 5)
In Figure 4 training with nonlinear weight of positivecosine and sine of HSCR-BF (HSCR-BF
+ cos and HSCR-BF+ sin) had some similarities The two models streamed
well at the initial stages from 896 to 992 and 908to 990 at soft layers of the rock respectively Howeverat the last stages the probability of transmitting effectivelybecame marginal from 795 to 852 and 821 to 854respectively Much as this hybrid could perform well in areaswhere routing conditions aremuch better in less dense rescuesituations this could hamper rescue mission in both casesdue to battery drain or collision from traffic conjunctionThedata formean iteration and convergent time forHSCR-BF
+ cos(Table 6) and mean Iteration standard variance the final
error HSCR-BF+ sin (Table 7) were consistent with the nega-
tive sine having better results than the positive cosineThe survival probability of the positive nonlinear weight
Gaussian hybrids (HSGR-BF+ cos) in Figure 5 is more com-
pact both at the initial (ie 873ndash996) and later (661ndash832) stages declining to harder rock layers with an errorof 001173 (Table 8)
At the initial stage particles are sensitive to inputs asthey moved quickly in the search space towards the targetand while particles peaked closer to the target it became lesssensitive to the input and began to align (Figure 6) At thelater stage more RBF are used to keep the error at minimumfor accuracy It was also noticed that instead of acceler-ating higher and becoming more sensitive to the inputswhile particles were far from the target the hybrids withlowast
[minus sin(119877)] and lowast[ + sin(119877)] were less sensitive to the inputsas lowast[minus sin(119877)] descended from [119909 = minus1 119910 = 05785] to[119909 = 1 119910 = 04215] while lowast[+ sin(119877)] descended from [119909 =
minus3 119910 = 00404] to [119909 = minus1 119910 = minus004062] before becomingconscious of the inputs In addition the hybrids with thelowast
[+ cos(119877)] lagged slightly between [119909 = minus4 119910 = 0006014]
and [119909 = minus2 119910 = 006288] before becoming sensitive toinputs However they all lined up for accuracy in terms of theminimised error
The nonlinearity of the hybrids is presented (Figure 7)the negative cosinesine weight is used to reduce nonlinearfor small input values and this lies between minus1 lt 119909 lt +1 andthe positive cosinesine weight is used to keep high nonlinearfor large inputs for the remaining region
63 CPU Time Efficiency The relationship between varioushybrids with respect to the central processing time (CPU)was profiled for different runs (Table 9) and expressed in asixth-order polynomial given as 119884
119903
= 1205737
1198836
1
+ 1205736
1198835
1
+ 1205735
1198834
1
+
1205734
1198833
1
+ 1205733
1198832
1
+ 1205732
1198831
+ 1205730
where 1198831
is time (seconds) and 120573
is the coefficient of the polynomial (Figure 8) The proposedhybrid has better usage of CPU time with 119877
2
= 09160followed byHSCR-BF
minus sin andHSCR-BF+ cos with119877
2 of 07551and 07244 respectively (Table 10) Applying the proposedalgorithm into Gaussian the CPU usage had 119877
2 of 07345
10 ISRN Artificial Intelligence
0 50 100 150 200
002
004
006
008
X 249Y 001103
PSO optimized error for hybrid SBF-CRBF with positive nonlinear weight
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 607
08
09
1
11 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(HSCR+cos )
(a) HSCR-BF+ cos
0 50 100 150 200
005
01
015
02
X 249Y 001247
PSO optimized error with nonlinear weight [+sine(R)]
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 6
070809
111 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(b) HSCR-BF+ sin
Figure 4 Hybrid of SBF and CRBF (HSCR-BF) with (a) HSCR-BF+ cos and (b) HSCR-BF
+ sin nonlinear weight
Table 6 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BF+ cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 350780 522257 0311 115650 0121AVG 35078 52226 0031 11565 0012MaxMin 54840 63413 0009 93500 0006
and indeedmarginally outperformedHSCR-BF+ cos Detailed
work on SBF CRBF and GRBF with regard to scalabilitymemory usage and the central processing time has beencarried by the authors [1]
7 Conclusion
In summary we made the following contributions First weused the mix of SBF and CRBF to present several hybridswith different nonlinear weights of cosine and sine func-tions on compact radial basis function Next we showedthe performance of the proposed nonlinear weight hybridsHSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sin and HSCR-BF+ sin optimised all the parameters with minimised error
of 00098 0012 0013 and 0013 respectively compared to00117 for Gaussian HSCR-BF
+ cos The analyzed CPU usagewith corresponding 119877
2 values of 09160 075 072440 06731and 07345 for HSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sinHSCR-BF
+ sin and HSCR-BF+ cos respectively demonstrated
that the algorithm is scalable There exist some evacuationmodels that is Goh and Mandic [11] which offered a choicefor travelers and several schemes for the decisionmakersmaynot be applicable in emergency undergroundmine situationsThe proposed nonlinear hybrid algorithm with particle swarmoptimisation has better capability of approximation to underly-ing functions with a fast learning speed and high robustness and
is competitive and more computationally efficient to Gaussianwith the same nonlinear weight The algorithm is new andthat makes it difficult to identify limitations and we intend toinvestigate other hybrids and comparewith genetic algorithm(GA) in the future
Nomenclature
119873120575min 119873120575max Minimum and maximum signalreach
HSCR-BFHSGR-BF Hybrid of SBF with CRBFHybrid ofSBF with GRBF
HSCR-BFminus cos Hybrid of SBF and CRBF minuscos
nonlinear weightHSCR-BF
+ cos Hybrid of SBF and CRBF +cosnonlinear weight
HSCR-BFminus sin Hybrid of SBF and CRBF minussine
nonlinear weightHSCR-BF
+ sin Hybrid of SBF and CRBF +sinenonlinear weight
HSGR-BF+ cos Hybrid of SBF and GRBF +cos
nonlinear weight
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
ISRN Artificial Intelligence 11
Table 7 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BF+ sin)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 260360 407109 03457 99800 0129AVG 26036 40711 00346 9980 0013MinMax 27840 56040 0008 106500 0014
Table 8 Hybrid of SBF and GRBF with nonlinear weight on GRBF (HSGR-BF+ cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 219840 349010 0439 112100 01173AVG 21984 34901 0044 11210 00117MinMax 27000 46580 0008 80000 00165
0 50 100 1500
005
01
015
02
X 144Y 0009595
GRBF-PSO optimized error with nonlinear weight [+cos(R)]
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 604
06
08
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
Figure 5 Hybrid of SBF and GRBF with nonlinear weight on GRBF(HSCR-BF
+ cos)
0 2 4 6 8 10
002040608
1121416
Weighed position of particles
Min
imise
d er
ror
Optimised error in nonlinear weighed SBF-CRBF hybridsParticles
Less sensitive to inputsmuch closer to target
more accurate
Less sensitive to inputsmuch closer more accurate
Neurons are more sensitive to inputsfast and far from target
minus02minus2minus4minus6minus8minus10
[x = minus1 y = 05785]
[x = 1 y = 04215]
[minus1 004062]
[x = minus2 y = 006288]
[x = minus4 y = 0006014]
[minus3 00404]
Hybrid+cosHybrid+sin
HybridminuscosHybridminussin
Figure 6 Particles weighed position
Acknowledgments
The authors would like to appreciate the immense contribu-tion of the mining companies where the study was under-taken The authors are grateful to these individuals for
0 1 2 3 4 5
0
05
1
15
2
X 37Y 0989
Particles position
Opt
imise
d er
ror
Nonlinear weight hybrids
Keep large inputvalues more
nonlinear
Keep small inputvalues lessnonlinear
Large input values
Small input values
Using positive cosine and sine
Using negative cosine and sineminus05
minus5 minus4 minus3 minus2 minus1
HSCR-BF+cosHSGR-BF+cos
HSCR-BF+sin
HSCR-BFminuscos
HSCR-BFminuscos
HSCR-BFminussin
HSCR-BF+cos HSCR-BF+sin HSCR-BFminussinand HSGR-BF+cos
Figure 7 Nonlinear weighed curves for optimized error withoutGaussian negative cosine curve
HSGR-BF+cos
HSGR-BF+cos
HSCR-BF+sinHSCR-BFminuscosHSCR-BFminussin
00020000400006000080000
100000120000
0 2 4 6 8 10 12
CPU
tim
e
Time (s)
Nonlinear hybrids
Figure 8 The trend of hybrids CPU time curves using sixth-OrderPolynomial
their immeasurable contributions to this work Fred Atta-kuma Patrick Addai Isaac Owusu Kwankye Thomas KwawAnnan Willet Agongo Nathaniel Quansah Francis OwusuMensah Clement Owusu-Asamoah Joseph Adu-MensahShadrack Aidoo Martin Anokye F T Oduro Ernest Ekow
12 ISRN Artificial Intelligence
Table 9 CPU time usage efficiency for the hybrids
Runs HSCR-BFminus cos HSCR-BF
minus sin HSCR-BF+ cos HSCR-BF
+ sin HSGR-BF+ cos
Total 255889 254350 422344 643962 438133Average 25589 25435 42234 64396 43813
Table 10 Analysis of CPU time on the various hybrids
1205736
1205735
1205734
1205733
1205732
1205731
1205730
1198772
HSCR-BFminus cos 00159119909 minus02575119909 minus0759119909 33494119909 minus19901119909 40345119909 25322 09160
HSCR-BF+ cos 00903119909 minus23158119909 19932119909 minus58027119909 minus31337119909 32808119909 14747 07244
HSCR-BFminus sin minus0033119909 11354119909 minus15811119909 11368119909 minus4337119909 78008119909 minus20794 07551
HSCR-BF+ sin 02571119909 minus14263119909 17173119909 minus10069119909 2993119909 minus42374119909 27879 06731
HSGR-BF+ cos minus00437119909 10911119909 minus82923119909 66218119909 15934119909 minus53939119909 8191 07345
Abano E Omari-Siaw and Qiang Jia This work was sup-ported by the National Natural Science Foundation of China(no 71271103) and by the Six Talents Peak Foundation ofJiangsu Province
References
[1] M O Ansong H X Yao and J S Huang ldquoRadial and sigmoidbasis functions neural networks in wireless sensor routingtopology control in underground mine rescue operation basedon particle swarm optimizationrdquo International Journal of Dis-tributed Sensor Networks-Hindawi vol 2013 Article ID 37693114 pages 2013
[2] R Neruda and P Kudova ldquoLearning methods for radial basisfunction networksrdquo Future Generation Computer Systems vol21 no 7 pp 1131ndash1142 2005
[3] C T Cheng J Y Lin and Y G C Sun ldquoLong-term predictionof discharges in Manwan hydropower using adaptive-network-based fuzzy inference systems modelsrdquo in Advances in NaturalComputation vol 3498 of Lecture Notes in Computer Sciencepp 1152ndash1161 2005
[4] D Broomhead and D Lowe ldquoMultivariate functional interpo-lation and adaptive networksrdquo Complex System vol 2 pp 321ndash355 1988
[5] Y Chen C Chuah and Q Zhao ldquoNetwork configuration foroptimal utilization efficiency of wireless sensor networksrdquo AdHoc Networks vol 6 no 1 pp 92ndash107 2008
[6] P S Rajpal K S Shishodia and G S Sekhon ldquoAn artificialneural network for modeling reliability availability and main-tainability of a repairable systemrdquo Reliability Engineering andSystem Safety vol 91 no 7 pp 809ndash819 2006
[7] W Jang W M Healy and S J Mirosław ldquoWireless sensornetworks as part of a web-based building environmental mon-itoring systemrdquo Automation in Construction vol 17 no 6 pp729ndash736 2008
[8] M S Pan C H Tsai and Y C Tseng ldquoEmergengy guiding andmonitoring application in indoor 3D environment by WirelessSensor Networkrdquo Internation Journal Os Sensor Netwourk vol1 pp 2ndash10 2006
[9] S Zarifzadeh A Nayyeri and N Yazdani ldquoEfficient construc-tion of network topology to conserve energy in wireless ad hocnetworksrdquo Computer Communications vol 31 no 1 pp 160ndash173 2008
[10] R Riaz A Naureen A Akram A H Akbar K Kim and HFarooq Ahmed ldquoA unified security framework with three keymanagement schemes for wireless sensor networksrdquo ComputerCommunications vol 31 no 18 pp 4269ndash4280 2008
[11] S L Goh and D P Mandic ldquoAn augmented CRTRL forcomplex-valued recurrent neural networksrdquo Neural Networksvol 20 no 10 pp 1061ndash1066 2007
[12] L A B Munoz and J J V Ramosy ldquoSimilarity-based heteroge-neous neural networksrdquo Engineering Letters vol 14 no 2 pp103ndash116 2007
[13] R Taormina K Chau and R Sethi ldquoArtificial neural networksimulation of hourly groundwater levels in a coastal aquifer sys-tem of the Venice lagoonrdquo Engineering Applications of ArtificialIntelligence vol 25 no 8 pp 1670ndash1676 2012
[14] K Leblebicioglu and U Halici ldquoInfinite dimensional radialbasis function neural networks for nonlinear transformationson function spacesrdquo Nonlinear Analysis Theory Methods andApplications vol 30 no 3 pp 1649ndash1654 1997
[15] F Fernandez-Navarro C Hervas-Martınez J Sanchez-Monedero and P A Gutierrez ldquoMELM-GRBF a modifiedversion of the extreme learning machine for generalized radialbasis function neural networksrdquo Neurocomputing vol 74 no16 pp 2502ndash2510 2011
[16] C L Wu KW Chau and Y S Li ldquoPredicting monthly stream-flow using data-drivenmodels coupledwith data-preprocessingtechniquesrdquoWater Resources Research vol 45 no 8 Article IDW08432 2009
[17] C EACheng ldquoLong-termprediction of discharges inManwanReservoir using artificial neural network modelsrdquo in Advancesin Neural NetworksmdashISNN 2005 vol 3498 of Lecture Notes inComputer Science pp 1040ndash1045 2005
[18] H H Zhang M Genton and P Liu ldquoCompactly supportedradial basis function kernelsrdquo 2004
[19] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquo inProceedings of the 1995 IEEE International Conference on NeuralNetworks pp 1942ndash1945 December 1995
[20] J Eberhart R Eberhart and Y Shi ldquoParticle swarm optimiza-tion developments applications and resourcesrdquo IEEE vol 1 pp81ndash86 2001
[21] R Malhotra and A Negi ldquoReliability modeling using ParticleSwarm optimisationrdquo International Journal of System AssuranceEngineering and Management vol 4 no 3 pp 275ndash283 2013
[22] D Gies and Y Rahmat-Samii ldquoParticle swarm optimization forreconfigurable phase-differentiated array designrdquo Microwaveand Optical Technology Letters vol 38 no 3 pp 168ndash175 2003
ISRN Artificial Intelligence 13
[23] F Valdez P Melin and C Oscar ldquoAn improved evolution-ary method with fuzzy logic for combining Particle Swarmoptimization and genetic algorithmsrdquo Applied Soft ComputingJournal vol 11 no 2 pp 2625ndash2632 2011
[24] J Zhang and K Chau ldquoMultilayer ensemble pruning vianovel multi-sub-swarm particle swarm optimizationrdquo Journalof Universal Computer Science vol 15 no 4 pp 840ndash858 2009
[25] F Valdez P Melin and O Castillo ldquoParallel Particle Swarmoptimization with parameters adaptation using fuzzy logicrdquoin Proceedings of the 11th Mexican International Conference onArtificial Intelligence (MICAI rsquo12) vol 2 pp 374ndash385 San LuisPotosi Mexico October 2012
[26] J Peng and C P Y Pan ldquoParticle swarm optimization RBFfor gas emission predictionrdquo Journal of Safety Science andTechnology vol 7 pp 77ndash85 2011
[27] G Ren Z Huang Y Cheng X Zhao and Y Zhang ldquoAnintegrated model for evacuation routing and traffic signaloptimization with background demand uncertaintyrdquo Journal ofAdvanced Transportation vol 47 pp 4ndash27 2013
[28] X Feng Z Xiao and X Cui ldquoImproved RSSI algorithm forwireless sensor networks in 3Drdquo Journal of ComputationalInformation Systems vol 7 no 16 pp 5866ndash5873 2011
[29] F Valdez and P Melin ldquoNeural network optimization witha hybrid evolutionary method that combines particle swarmand genetic algorithms with fuzzy rulesrdquo in Proceedings ofthe Annual Meeting of the North American Fuzzy InformationProcessing Society (NAFIPS rsquo08) NewYork NYUSAMay 2008
[30] F Fernandez-Navarro C Hervas-Martınez P A Gutierrez JM Pena-Barragan and F Lopez-Granados ldquoParameter estima-tion of q-Gaussian Radial Basis Functions Neural Networkswith a Hybrid Algorithm for binary classificationrdquo Neurocom-puting vol 75 pp 123ndash134 2012
[31] H Soh S Lim T Zhang et al ldquoWeighted complex networkanalysis of travel routes on the Singapore public transportationsystemrdquo Physica A vol 389 no 24 pp 5852ndash5863 2010
[32] C K S Kumar R Sukumar and M Nageswari ldquoSensorslifetime enhancement techniques in wireless sensor networksmdasha critical reviewrdquo International Journal of Computer Science andInformation Technology amp Security vol 3 pp 159ndash164 2013
[33] S Li and F Qin ldquoA dynamic neural network approach forsolving nonlinear inequalities defined on a graph and itsapplication to distributed routing-free range-free localizationof WSNsrdquo Neurocomputing vol 117 pp 72ndash80 2013
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
10 ISRN Artificial Intelligence
0 50 100 150 200
002
004
006
008
X 249Y 001103
PSO optimized error for hybrid SBF-CRBF with positive nonlinear weight
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 607
08
09
1
11 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(HSCR+cos )
(a) HSCR-BF+ cos
0 50 100 150 200
005
01
015
02
X 249Y 001247
PSO optimized error with nonlinear weight [+sine(R)]
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 6
070809
111 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
(b) HSCR-BF+ sin
Figure 4 Hybrid of SBF and CRBF (HSCR-BF) with (a) HSCR-BF+ cos and (b) HSCR-BF
+ sin nonlinear weight
Table 6 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BF+ cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 350780 522257 0311 115650 0121AVG 35078 52226 0031 11565 0012MaxMin 54840 63413 0009 93500 0006
and indeedmarginally outperformedHSCR-BF+ cos Detailed
work on SBF CRBF and GRBF with regard to scalabilitymemory usage and the central processing time has beencarried by the authors [1]
7 Conclusion
In summary we made the following contributions First weused the mix of SBF and CRBF to present several hybridswith different nonlinear weights of cosine and sine func-tions on compact radial basis function Next we showedthe performance of the proposed nonlinear weight hybridsHSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sin and HSCR-BF+ sin optimised all the parameters with minimised error
of 00098 0012 0013 and 0013 respectively compared to00117 for Gaussian HSCR-BF
+ cos The analyzed CPU usagewith corresponding 119877
2 values of 09160 075 072440 06731and 07345 for HSCR-BF
minus cos HSCR-BF+ cos HSCR-BF
minus sinHSCR-BF
+ sin and HSCR-BF+ cos respectively demonstrated
that the algorithm is scalable There exist some evacuationmodels that is Goh and Mandic [11] which offered a choicefor travelers and several schemes for the decisionmakersmaynot be applicable in emergency undergroundmine situationsThe proposed nonlinear hybrid algorithm with particle swarmoptimisation has better capability of approximation to underly-ing functions with a fast learning speed and high robustness and
is competitive and more computationally efficient to Gaussianwith the same nonlinear weight The algorithm is new andthat makes it difficult to identify limitations and we intend toinvestigate other hybrids and comparewith genetic algorithm(GA) in the future
Nomenclature
119873120575min 119873120575max Minimum and maximum signalreach
HSCR-BFHSGR-BF Hybrid of SBF with CRBFHybrid ofSBF with GRBF
HSCR-BFminus cos Hybrid of SBF and CRBF minuscos
nonlinear weightHSCR-BF
+ cos Hybrid of SBF and CRBF +cosnonlinear weight
HSCR-BFminus sin Hybrid of SBF and CRBF minussine
nonlinear weightHSCR-BF
+ sin Hybrid of SBF and CRBF +sinenonlinear weight
HSGR-BF+ cos Hybrid of SBF and GRBF +cos
nonlinear weight
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
ISRN Artificial Intelligence 11
Table 7 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BF+ sin)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 260360 407109 03457 99800 0129AVG 26036 40711 00346 9980 0013MinMax 27840 56040 0008 106500 0014
Table 8 Hybrid of SBF and GRBF with nonlinear weight on GRBF (HSGR-BF+ cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 219840 349010 0439 112100 01173AVG 21984 34901 0044 11210 00117MinMax 27000 46580 0008 80000 00165
0 50 100 1500
005
01
015
02
X 144Y 0009595
GRBF-PSO optimized error with nonlinear weight [+cos(R)]
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 604
06
08
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
Figure 5 Hybrid of SBF and GRBF with nonlinear weight on GRBF(HSCR-BF
+ cos)
0 2 4 6 8 10
002040608
1121416
Weighed position of particles
Min
imise
d er
ror
Optimised error in nonlinear weighed SBF-CRBF hybridsParticles
Less sensitive to inputsmuch closer to target
more accurate
Less sensitive to inputsmuch closer more accurate
Neurons are more sensitive to inputsfast and far from target
minus02minus2minus4minus6minus8minus10
[x = minus1 y = 05785]
[x = 1 y = 04215]
[minus1 004062]
[x = minus2 y = 006288]
[x = minus4 y = 0006014]
[minus3 00404]
Hybrid+cosHybrid+sin
HybridminuscosHybridminussin
Figure 6 Particles weighed position
Acknowledgments
The authors would like to appreciate the immense contribu-tion of the mining companies where the study was under-taken The authors are grateful to these individuals for
0 1 2 3 4 5
0
05
1
15
2
X 37Y 0989
Particles position
Opt
imise
d er
ror
Nonlinear weight hybrids
Keep large inputvalues more
nonlinear
Keep small inputvalues lessnonlinear
Large input values
Small input values
Using positive cosine and sine
Using negative cosine and sineminus05
minus5 minus4 minus3 minus2 minus1
HSCR-BF+cosHSGR-BF+cos
HSCR-BF+sin
HSCR-BFminuscos
HSCR-BFminuscos
HSCR-BFminussin
HSCR-BF+cos HSCR-BF+sin HSCR-BFminussinand HSGR-BF+cos
Figure 7 Nonlinear weighed curves for optimized error withoutGaussian negative cosine curve
HSGR-BF+cos
HSGR-BF+cos
HSCR-BF+sinHSCR-BFminuscosHSCR-BFminussin
00020000400006000080000
100000120000
0 2 4 6 8 10 12
CPU
tim
e
Time (s)
Nonlinear hybrids
Figure 8 The trend of hybrids CPU time curves using sixth-OrderPolynomial
their immeasurable contributions to this work Fred Atta-kuma Patrick Addai Isaac Owusu Kwankye Thomas KwawAnnan Willet Agongo Nathaniel Quansah Francis OwusuMensah Clement Owusu-Asamoah Joseph Adu-MensahShadrack Aidoo Martin Anokye F T Oduro Ernest Ekow
12 ISRN Artificial Intelligence
Table 9 CPU time usage efficiency for the hybrids
Runs HSCR-BFminus cos HSCR-BF
minus sin HSCR-BF+ cos HSCR-BF
+ sin HSGR-BF+ cos
Total 255889 254350 422344 643962 438133Average 25589 25435 42234 64396 43813
Table 10 Analysis of CPU time on the various hybrids
1205736
1205735
1205734
1205733
1205732
1205731
1205730
1198772
HSCR-BFminus cos 00159119909 minus02575119909 minus0759119909 33494119909 minus19901119909 40345119909 25322 09160
HSCR-BF+ cos 00903119909 minus23158119909 19932119909 minus58027119909 minus31337119909 32808119909 14747 07244
HSCR-BFminus sin minus0033119909 11354119909 minus15811119909 11368119909 minus4337119909 78008119909 minus20794 07551
HSCR-BF+ sin 02571119909 minus14263119909 17173119909 minus10069119909 2993119909 minus42374119909 27879 06731
HSGR-BF+ cos minus00437119909 10911119909 minus82923119909 66218119909 15934119909 minus53939119909 8191 07345
Abano E Omari-Siaw and Qiang Jia This work was sup-ported by the National Natural Science Foundation of China(no 71271103) and by the Six Talents Peak Foundation ofJiangsu Province
References
[1] M O Ansong H X Yao and J S Huang ldquoRadial and sigmoidbasis functions neural networks in wireless sensor routingtopology control in underground mine rescue operation basedon particle swarm optimizationrdquo International Journal of Dis-tributed Sensor Networks-Hindawi vol 2013 Article ID 37693114 pages 2013
[2] R Neruda and P Kudova ldquoLearning methods for radial basisfunction networksrdquo Future Generation Computer Systems vol21 no 7 pp 1131ndash1142 2005
[3] C T Cheng J Y Lin and Y G C Sun ldquoLong-term predictionof discharges in Manwan hydropower using adaptive-network-based fuzzy inference systems modelsrdquo in Advances in NaturalComputation vol 3498 of Lecture Notes in Computer Sciencepp 1152ndash1161 2005
[4] D Broomhead and D Lowe ldquoMultivariate functional interpo-lation and adaptive networksrdquo Complex System vol 2 pp 321ndash355 1988
[5] Y Chen C Chuah and Q Zhao ldquoNetwork configuration foroptimal utilization efficiency of wireless sensor networksrdquo AdHoc Networks vol 6 no 1 pp 92ndash107 2008
[6] P S Rajpal K S Shishodia and G S Sekhon ldquoAn artificialneural network for modeling reliability availability and main-tainability of a repairable systemrdquo Reliability Engineering andSystem Safety vol 91 no 7 pp 809ndash819 2006
[7] W Jang W M Healy and S J Mirosław ldquoWireless sensornetworks as part of a web-based building environmental mon-itoring systemrdquo Automation in Construction vol 17 no 6 pp729ndash736 2008
[8] M S Pan C H Tsai and Y C Tseng ldquoEmergengy guiding andmonitoring application in indoor 3D environment by WirelessSensor Networkrdquo Internation Journal Os Sensor Netwourk vol1 pp 2ndash10 2006
[9] S Zarifzadeh A Nayyeri and N Yazdani ldquoEfficient construc-tion of network topology to conserve energy in wireless ad hocnetworksrdquo Computer Communications vol 31 no 1 pp 160ndash173 2008
[10] R Riaz A Naureen A Akram A H Akbar K Kim and HFarooq Ahmed ldquoA unified security framework with three keymanagement schemes for wireless sensor networksrdquo ComputerCommunications vol 31 no 18 pp 4269ndash4280 2008
[11] S L Goh and D P Mandic ldquoAn augmented CRTRL forcomplex-valued recurrent neural networksrdquo Neural Networksvol 20 no 10 pp 1061ndash1066 2007
[12] L A B Munoz and J J V Ramosy ldquoSimilarity-based heteroge-neous neural networksrdquo Engineering Letters vol 14 no 2 pp103ndash116 2007
[13] R Taormina K Chau and R Sethi ldquoArtificial neural networksimulation of hourly groundwater levels in a coastal aquifer sys-tem of the Venice lagoonrdquo Engineering Applications of ArtificialIntelligence vol 25 no 8 pp 1670ndash1676 2012
[14] K Leblebicioglu and U Halici ldquoInfinite dimensional radialbasis function neural networks for nonlinear transformationson function spacesrdquo Nonlinear Analysis Theory Methods andApplications vol 30 no 3 pp 1649ndash1654 1997
[15] F Fernandez-Navarro C Hervas-Martınez J Sanchez-Monedero and P A Gutierrez ldquoMELM-GRBF a modifiedversion of the extreme learning machine for generalized radialbasis function neural networksrdquo Neurocomputing vol 74 no16 pp 2502ndash2510 2011
[16] C L Wu KW Chau and Y S Li ldquoPredicting monthly stream-flow using data-drivenmodels coupledwith data-preprocessingtechniquesrdquoWater Resources Research vol 45 no 8 Article IDW08432 2009
[17] C EACheng ldquoLong-termprediction of discharges inManwanReservoir using artificial neural network modelsrdquo in Advancesin Neural NetworksmdashISNN 2005 vol 3498 of Lecture Notes inComputer Science pp 1040ndash1045 2005
[18] H H Zhang M Genton and P Liu ldquoCompactly supportedradial basis function kernelsrdquo 2004
[19] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquo inProceedings of the 1995 IEEE International Conference on NeuralNetworks pp 1942ndash1945 December 1995
[20] J Eberhart R Eberhart and Y Shi ldquoParticle swarm optimiza-tion developments applications and resourcesrdquo IEEE vol 1 pp81ndash86 2001
[21] R Malhotra and A Negi ldquoReliability modeling using ParticleSwarm optimisationrdquo International Journal of System AssuranceEngineering and Management vol 4 no 3 pp 275ndash283 2013
[22] D Gies and Y Rahmat-Samii ldquoParticle swarm optimization forreconfigurable phase-differentiated array designrdquo Microwaveand Optical Technology Letters vol 38 no 3 pp 168ndash175 2003
ISRN Artificial Intelligence 13
[23] F Valdez P Melin and C Oscar ldquoAn improved evolution-ary method with fuzzy logic for combining Particle Swarmoptimization and genetic algorithmsrdquo Applied Soft ComputingJournal vol 11 no 2 pp 2625ndash2632 2011
[24] J Zhang and K Chau ldquoMultilayer ensemble pruning vianovel multi-sub-swarm particle swarm optimizationrdquo Journalof Universal Computer Science vol 15 no 4 pp 840ndash858 2009
[25] F Valdez P Melin and O Castillo ldquoParallel Particle Swarmoptimization with parameters adaptation using fuzzy logicrdquoin Proceedings of the 11th Mexican International Conference onArtificial Intelligence (MICAI rsquo12) vol 2 pp 374ndash385 San LuisPotosi Mexico October 2012
[26] J Peng and C P Y Pan ldquoParticle swarm optimization RBFfor gas emission predictionrdquo Journal of Safety Science andTechnology vol 7 pp 77ndash85 2011
[27] G Ren Z Huang Y Cheng X Zhao and Y Zhang ldquoAnintegrated model for evacuation routing and traffic signaloptimization with background demand uncertaintyrdquo Journal ofAdvanced Transportation vol 47 pp 4ndash27 2013
[28] X Feng Z Xiao and X Cui ldquoImproved RSSI algorithm forwireless sensor networks in 3Drdquo Journal of ComputationalInformation Systems vol 7 no 16 pp 5866ndash5873 2011
[29] F Valdez and P Melin ldquoNeural network optimization witha hybrid evolutionary method that combines particle swarmand genetic algorithms with fuzzy rulesrdquo in Proceedings ofthe Annual Meeting of the North American Fuzzy InformationProcessing Society (NAFIPS rsquo08) NewYork NYUSAMay 2008
[30] F Fernandez-Navarro C Hervas-Martınez P A Gutierrez JM Pena-Barragan and F Lopez-Granados ldquoParameter estima-tion of q-Gaussian Radial Basis Functions Neural Networkswith a Hybrid Algorithm for binary classificationrdquo Neurocom-puting vol 75 pp 123ndash134 2012
[31] H Soh S Lim T Zhang et al ldquoWeighted complex networkanalysis of travel routes on the Singapore public transportationsystemrdquo Physica A vol 389 no 24 pp 5852ndash5863 2010
[32] C K S Kumar R Sukumar and M Nageswari ldquoSensorslifetime enhancement techniques in wireless sensor networksmdasha critical reviewrdquo International Journal of Computer Science andInformation Technology amp Security vol 3 pp 159ndash164 2013
[33] S Li and F Qin ldquoA dynamic neural network approach forsolving nonlinear inequalities defined on a graph and itsapplication to distributed routing-free range-free localizationof WSNsrdquo Neurocomputing vol 117 pp 72ndash80 2013
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ISRN Artificial Intelligence 11
Table 7 Hybrid of SBF and CRBF with nonlinear weight on CRBF (HSCR-BF+ sin)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 260360 407109 03457 99800 0129AVG 26036 40711 00346 9980 0013MinMax 27840 56040 0008 106500 0014
Table 8 Hybrid of SBF and GRBF with nonlinear weight on GRBF (HSGR-BF+ cos)
Runs Mean iteration Std variation Std deviation Convergent time Final errorTotal 219840 349010 0439 112100 01173AVG 21984 34901 0044 11210 00117MinMax 27000 46580 0008 80000 00165
0 50 100 1500
005
01
015
02
X 144Y 0009595
GRBF-PSO optimized error with nonlinear weight [+cos(R)]
Number of iterations
Fitn
ess v
alue
1 15 2 25 3 35 4 45 5 55 604
06
08
1 Output
Rock hardness case (07ndash09)
Surv
ival
pro
babi
lity
Figure 5 Hybrid of SBF and GRBF with nonlinear weight on GRBF(HSCR-BF
+ cos)
0 2 4 6 8 10
002040608
1121416
Weighed position of particles
Min
imise
d er
ror
Optimised error in nonlinear weighed SBF-CRBF hybridsParticles
Less sensitive to inputsmuch closer to target
more accurate
Less sensitive to inputsmuch closer more accurate
Neurons are more sensitive to inputsfast and far from target
minus02minus2minus4minus6minus8minus10
[x = minus1 y = 05785]
[x = 1 y = 04215]
[minus1 004062]
[x = minus2 y = 006288]
[x = minus4 y = 0006014]
[minus3 00404]
Hybrid+cosHybrid+sin
HybridminuscosHybridminussin
Figure 6 Particles weighed position
Acknowledgments
The authors would like to appreciate the immense contribu-tion of the mining companies where the study was under-taken The authors are grateful to these individuals for
0 1 2 3 4 5
0
05
1
15
2
X 37Y 0989
Particles position
Opt
imise
d er
ror
Nonlinear weight hybrids
Keep large inputvalues more
nonlinear
Keep small inputvalues lessnonlinear
Large input values
Small input values
Using positive cosine and sine
Using negative cosine and sineminus05
minus5 minus4 minus3 minus2 minus1
HSCR-BF+cosHSGR-BF+cos
HSCR-BF+sin
HSCR-BFminuscos
HSCR-BFminuscos
HSCR-BFminussin
HSCR-BF+cos HSCR-BF+sin HSCR-BFminussinand HSGR-BF+cos
Figure 7 Nonlinear weighed curves for optimized error withoutGaussian negative cosine curve
HSGR-BF+cos
HSGR-BF+cos
HSCR-BF+sinHSCR-BFminuscosHSCR-BFminussin
00020000400006000080000
100000120000
0 2 4 6 8 10 12
CPU
tim
e
Time (s)
Nonlinear hybrids
Figure 8 The trend of hybrids CPU time curves using sixth-OrderPolynomial
their immeasurable contributions to this work Fred Atta-kuma Patrick Addai Isaac Owusu Kwankye Thomas KwawAnnan Willet Agongo Nathaniel Quansah Francis OwusuMensah Clement Owusu-Asamoah Joseph Adu-MensahShadrack Aidoo Martin Anokye F T Oduro Ernest Ekow
12 ISRN Artificial Intelligence
Table 9 CPU time usage efficiency for the hybrids
Runs HSCR-BFminus cos HSCR-BF
minus sin HSCR-BF+ cos HSCR-BF
+ sin HSGR-BF+ cos
Total 255889 254350 422344 643962 438133Average 25589 25435 42234 64396 43813
Table 10 Analysis of CPU time on the various hybrids
1205736
1205735
1205734
1205733
1205732
1205731
1205730
1198772
HSCR-BFminus cos 00159119909 minus02575119909 minus0759119909 33494119909 minus19901119909 40345119909 25322 09160
HSCR-BF+ cos 00903119909 minus23158119909 19932119909 minus58027119909 minus31337119909 32808119909 14747 07244
HSCR-BFminus sin minus0033119909 11354119909 minus15811119909 11368119909 minus4337119909 78008119909 minus20794 07551
HSCR-BF+ sin 02571119909 minus14263119909 17173119909 minus10069119909 2993119909 minus42374119909 27879 06731
HSGR-BF+ cos minus00437119909 10911119909 minus82923119909 66218119909 15934119909 minus53939119909 8191 07345
Abano E Omari-Siaw and Qiang Jia This work was sup-ported by the National Natural Science Foundation of China(no 71271103) and by the Six Talents Peak Foundation ofJiangsu Province
References
[1] M O Ansong H X Yao and J S Huang ldquoRadial and sigmoidbasis functions neural networks in wireless sensor routingtopology control in underground mine rescue operation basedon particle swarm optimizationrdquo International Journal of Dis-tributed Sensor Networks-Hindawi vol 2013 Article ID 37693114 pages 2013
[2] R Neruda and P Kudova ldquoLearning methods for radial basisfunction networksrdquo Future Generation Computer Systems vol21 no 7 pp 1131ndash1142 2005
[3] C T Cheng J Y Lin and Y G C Sun ldquoLong-term predictionof discharges in Manwan hydropower using adaptive-network-based fuzzy inference systems modelsrdquo in Advances in NaturalComputation vol 3498 of Lecture Notes in Computer Sciencepp 1152ndash1161 2005
[4] D Broomhead and D Lowe ldquoMultivariate functional interpo-lation and adaptive networksrdquo Complex System vol 2 pp 321ndash355 1988
[5] Y Chen C Chuah and Q Zhao ldquoNetwork configuration foroptimal utilization efficiency of wireless sensor networksrdquo AdHoc Networks vol 6 no 1 pp 92ndash107 2008
[6] P S Rajpal K S Shishodia and G S Sekhon ldquoAn artificialneural network for modeling reliability availability and main-tainability of a repairable systemrdquo Reliability Engineering andSystem Safety vol 91 no 7 pp 809ndash819 2006
[7] W Jang W M Healy and S J Mirosław ldquoWireless sensornetworks as part of a web-based building environmental mon-itoring systemrdquo Automation in Construction vol 17 no 6 pp729ndash736 2008
[8] M S Pan C H Tsai and Y C Tseng ldquoEmergengy guiding andmonitoring application in indoor 3D environment by WirelessSensor Networkrdquo Internation Journal Os Sensor Netwourk vol1 pp 2ndash10 2006
[9] S Zarifzadeh A Nayyeri and N Yazdani ldquoEfficient construc-tion of network topology to conserve energy in wireless ad hocnetworksrdquo Computer Communications vol 31 no 1 pp 160ndash173 2008
[10] R Riaz A Naureen A Akram A H Akbar K Kim and HFarooq Ahmed ldquoA unified security framework with three keymanagement schemes for wireless sensor networksrdquo ComputerCommunications vol 31 no 18 pp 4269ndash4280 2008
[11] S L Goh and D P Mandic ldquoAn augmented CRTRL forcomplex-valued recurrent neural networksrdquo Neural Networksvol 20 no 10 pp 1061ndash1066 2007
[12] L A B Munoz and J J V Ramosy ldquoSimilarity-based heteroge-neous neural networksrdquo Engineering Letters vol 14 no 2 pp103ndash116 2007
[13] R Taormina K Chau and R Sethi ldquoArtificial neural networksimulation of hourly groundwater levels in a coastal aquifer sys-tem of the Venice lagoonrdquo Engineering Applications of ArtificialIntelligence vol 25 no 8 pp 1670ndash1676 2012
[14] K Leblebicioglu and U Halici ldquoInfinite dimensional radialbasis function neural networks for nonlinear transformationson function spacesrdquo Nonlinear Analysis Theory Methods andApplications vol 30 no 3 pp 1649ndash1654 1997
[15] F Fernandez-Navarro C Hervas-Martınez J Sanchez-Monedero and P A Gutierrez ldquoMELM-GRBF a modifiedversion of the extreme learning machine for generalized radialbasis function neural networksrdquo Neurocomputing vol 74 no16 pp 2502ndash2510 2011
[16] C L Wu KW Chau and Y S Li ldquoPredicting monthly stream-flow using data-drivenmodels coupledwith data-preprocessingtechniquesrdquoWater Resources Research vol 45 no 8 Article IDW08432 2009
[17] C EACheng ldquoLong-termprediction of discharges inManwanReservoir using artificial neural network modelsrdquo in Advancesin Neural NetworksmdashISNN 2005 vol 3498 of Lecture Notes inComputer Science pp 1040ndash1045 2005
[18] H H Zhang M Genton and P Liu ldquoCompactly supportedradial basis function kernelsrdquo 2004
[19] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquo inProceedings of the 1995 IEEE International Conference on NeuralNetworks pp 1942ndash1945 December 1995
[20] J Eberhart R Eberhart and Y Shi ldquoParticle swarm optimiza-tion developments applications and resourcesrdquo IEEE vol 1 pp81ndash86 2001
[21] R Malhotra and A Negi ldquoReliability modeling using ParticleSwarm optimisationrdquo International Journal of System AssuranceEngineering and Management vol 4 no 3 pp 275ndash283 2013
[22] D Gies and Y Rahmat-Samii ldquoParticle swarm optimization forreconfigurable phase-differentiated array designrdquo Microwaveand Optical Technology Letters vol 38 no 3 pp 168ndash175 2003
ISRN Artificial Intelligence 13
[23] F Valdez P Melin and C Oscar ldquoAn improved evolution-ary method with fuzzy logic for combining Particle Swarmoptimization and genetic algorithmsrdquo Applied Soft ComputingJournal vol 11 no 2 pp 2625ndash2632 2011
[24] J Zhang and K Chau ldquoMultilayer ensemble pruning vianovel multi-sub-swarm particle swarm optimizationrdquo Journalof Universal Computer Science vol 15 no 4 pp 840ndash858 2009
[25] F Valdez P Melin and O Castillo ldquoParallel Particle Swarmoptimization with parameters adaptation using fuzzy logicrdquoin Proceedings of the 11th Mexican International Conference onArtificial Intelligence (MICAI rsquo12) vol 2 pp 374ndash385 San LuisPotosi Mexico October 2012
[26] J Peng and C P Y Pan ldquoParticle swarm optimization RBFfor gas emission predictionrdquo Journal of Safety Science andTechnology vol 7 pp 77ndash85 2011
[27] G Ren Z Huang Y Cheng X Zhao and Y Zhang ldquoAnintegrated model for evacuation routing and traffic signaloptimization with background demand uncertaintyrdquo Journal ofAdvanced Transportation vol 47 pp 4ndash27 2013
[28] X Feng Z Xiao and X Cui ldquoImproved RSSI algorithm forwireless sensor networks in 3Drdquo Journal of ComputationalInformation Systems vol 7 no 16 pp 5866ndash5873 2011
[29] F Valdez and P Melin ldquoNeural network optimization witha hybrid evolutionary method that combines particle swarmand genetic algorithms with fuzzy rulesrdquo in Proceedings ofthe Annual Meeting of the North American Fuzzy InformationProcessing Society (NAFIPS rsquo08) NewYork NYUSAMay 2008
[30] F Fernandez-Navarro C Hervas-Martınez P A Gutierrez JM Pena-Barragan and F Lopez-Granados ldquoParameter estima-tion of q-Gaussian Radial Basis Functions Neural Networkswith a Hybrid Algorithm for binary classificationrdquo Neurocom-puting vol 75 pp 123ndash134 2012
[31] H Soh S Lim T Zhang et al ldquoWeighted complex networkanalysis of travel routes on the Singapore public transportationsystemrdquo Physica A vol 389 no 24 pp 5852ndash5863 2010
[32] C K S Kumar R Sukumar and M Nageswari ldquoSensorslifetime enhancement techniques in wireless sensor networksmdasha critical reviewrdquo International Journal of Computer Science andInformation Technology amp Security vol 3 pp 159ndash164 2013
[33] S Li and F Qin ldquoA dynamic neural network approach forsolving nonlinear inequalities defined on a graph and itsapplication to distributed routing-free range-free localizationof WSNsrdquo Neurocomputing vol 117 pp 72ndash80 2013
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
12 ISRN Artificial Intelligence
Table 9 CPU time usage efficiency for the hybrids
Runs HSCR-BFminus cos HSCR-BF
minus sin HSCR-BF+ cos HSCR-BF
+ sin HSGR-BF+ cos
Total 255889 254350 422344 643962 438133Average 25589 25435 42234 64396 43813
Table 10 Analysis of CPU time on the various hybrids
1205736
1205735
1205734
1205733
1205732
1205731
1205730
1198772
HSCR-BFminus cos 00159119909 minus02575119909 minus0759119909 33494119909 minus19901119909 40345119909 25322 09160
HSCR-BF+ cos 00903119909 minus23158119909 19932119909 minus58027119909 minus31337119909 32808119909 14747 07244
HSCR-BFminus sin minus0033119909 11354119909 minus15811119909 11368119909 minus4337119909 78008119909 minus20794 07551
HSCR-BF+ sin 02571119909 minus14263119909 17173119909 minus10069119909 2993119909 minus42374119909 27879 06731
HSGR-BF+ cos minus00437119909 10911119909 minus82923119909 66218119909 15934119909 minus53939119909 8191 07345
Abano E Omari-Siaw and Qiang Jia This work was sup-ported by the National Natural Science Foundation of China(no 71271103) and by the Six Talents Peak Foundation ofJiangsu Province
References
[1] M O Ansong H X Yao and J S Huang ldquoRadial and sigmoidbasis functions neural networks in wireless sensor routingtopology control in underground mine rescue operation basedon particle swarm optimizationrdquo International Journal of Dis-tributed Sensor Networks-Hindawi vol 2013 Article ID 37693114 pages 2013
[2] R Neruda and P Kudova ldquoLearning methods for radial basisfunction networksrdquo Future Generation Computer Systems vol21 no 7 pp 1131ndash1142 2005
[3] C T Cheng J Y Lin and Y G C Sun ldquoLong-term predictionof discharges in Manwan hydropower using adaptive-network-based fuzzy inference systems modelsrdquo in Advances in NaturalComputation vol 3498 of Lecture Notes in Computer Sciencepp 1152ndash1161 2005
[4] D Broomhead and D Lowe ldquoMultivariate functional interpo-lation and adaptive networksrdquo Complex System vol 2 pp 321ndash355 1988
[5] Y Chen C Chuah and Q Zhao ldquoNetwork configuration foroptimal utilization efficiency of wireless sensor networksrdquo AdHoc Networks vol 6 no 1 pp 92ndash107 2008
[6] P S Rajpal K S Shishodia and G S Sekhon ldquoAn artificialneural network for modeling reliability availability and main-tainability of a repairable systemrdquo Reliability Engineering andSystem Safety vol 91 no 7 pp 809ndash819 2006
[7] W Jang W M Healy and S J Mirosław ldquoWireless sensornetworks as part of a web-based building environmental mon-itoring systemrdquo Automation in Construction vol 17 no 6 pp729ndash736 2008
[8] M S Pan C H Tsai and Y C Tseng ldquoEmergengy guiding andmonitoring application in indoor 3D environment by WirelessSensor Networkrdquo Internation Journal Os Sensor Netwourk vol1 pp 2ndash10 2006
[9] S Zarifzadeh A Nayyeri and N Yazdani ldquoEfficient construc-tion of network topology to conserve energy in wireless ad hocnetworksrdquo Computer Communications vol 31 no 1 pp 160ndash173 2008
[10] R Riaz A Naureen A Akram A H Akbar K Kim and HFarooq Ahmed ldquoA unified security framework with three keymanagement schemes for wireless sensor networksrdquo ComputerCommunications vol 31 no 18 pp 4269ndash4280 2008
[11] S L Goh and D P Mandic ldquoAn augmented CRTRL forcomplex-valued recurrent neural networksrdquo Neural Networksvol 20 no 10 pp 1061ndash1066 2007
[12] L A B Munoz and J J V Ramosy ldquoSimilarity-based heteroge-neous neural networksrdquo Engineering Letters vol 14 no 2 pp103ndash116 2007
[13] R Taormina K Chau and R Sethi ldquoArtificial neural networksimulation of hourly groundwater levels in a coastal aquifer sys-tem of the Venice lagoonrdquo Engineering Applications of ArtificialIntelligence vol 25 no 8 pp 1670ndash1676 2012
[14] K Leblebicioglu and U Halici ldquoInfinite dimensional radialbasis function neural networks for nonlinear transformationson function spacesrdquo Nonlinear Analysis Theory Methods andApplications vol 30 no 3 pp 1649ndash1654 1997
[15] F Fernandez-Navarro C Hervas-Martınez J Sanchez-Monedero and P A Gutierrez ldquoMELM-GRBF a modifiedversion of the extreme learning machine for generalized radialbasis function neural networksrdquo Neurocomputing vol 74 no16 pp 2502ndash2510 2011
[16] C L Wu KW Chau and Y S Li ldquoPredicting monthly stream-flow using data-drivenmodels coupledwith data-preprocessingtechniquesrdquoWater Resources Research vol 45 no 8 Article IDW08432 2009
[17] C EACheng ldquoLong-termprediction of discharges inManwanReservoir using artificial neural network modelsrdquo in Advancesin Neural NetworksmdashISNN 2005 vol 3498 of Lecture Notes inComputer Science pp 1040ndash1045 2005
[18] H H Zhang M Genton and P Liu ldquoCompactly supportedradial basis function kernelsrdquo 2004
[19] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquo inProceedings of the 1995 IEEE International Conference on NeuralNetworks pp 1942ndash1945 December 1995
[20] J Eberhart R Eberhart and Y Shi ldquoParticle swarm optimiza-tion developments applications and resourcesrdquo IEEE vol 1 pp81ndash86 2001
[21] R Malhotra and A Negi ldquoReliability modeling using ParticleSwarm optimisationrdquo International Journal of System AssuranceEngineering and Management vol 4 no 3 pp 275ndash283 2013
[22] D Gies and Y Rahmat-Samii ldquoParticle swarm optimization forreconfigurable phase-differentiated array designrdquo Microwaveand Optical Technology Letters vol 38 no 3 pp 168ndash175 2003
ISRN Artificial Intelligence 13
[23] F Valdez P Melin and C Oscar ldquoAn improved evolution-ary method with fuzzy logic for combining Particle Swarmoptimization and genetic algorithmsrdquo Applied Soft ComputingJournal vol 11 no 2 pp 2625ndash2632 2011
[24] J Zhang and K Chau ldquoMultilayer ensemble pruning vianovel multi-sub-swarm particle swarm optimizationrdquo Journalof Universal Computer Science vol 15 no 4 pp 840ndash858 2009
[25] F Valdez P Melin and O Castillo ldquoParallel Particle Swarmoptimization with parameters adaptation using fuzzy logicrdquoin Proceedings of the 11th Mexican International Conference onArtificial Intelligence (MICAI rsquo12) vol 2 pp 374ndash385 San LuisPotosi Mexico October 2012
[26] J Peng and C P Y Pan ldquoParticle swarm optimization RBFfor gas emission predictionrdquo Journal of Safety Science andTechnology vol 7 pp 77ndash85 2011
[27] G Ren Z Huang Y Cheng X Zhao and Y Zhang ldquoAnintegrated model for evacuation routing and traffic signaloptimization with background demand uncertaintyrdquo Journal ofAdvanced Transportation vol 47 pp 4ndash27 2013
[28] X Feng Z Xiao and X Cui ldquoImproved RSSI algorithm forwireless sensor networks in 3Drdquo Journal of ComputationalInformation Systems vol 7 no 16 pp 5866ndash5873 2011
[29] F Valdez and P Melin ldquoNeural network optimization witha hybrid evolutionary method that combines particle swarmand genetic algorithms with fuzzy rulesrdquo in Proceedings ofthe Annual Meeting of the North American Fuzzy InformationProcessing Society (NAFIPS rsquo08) NewYork NYUSAMay 2008
[30] F Fernandez-Navarro C Hervas-Martınez P A Gutierrez JM Pena-Barragan and F Lopez-Granados ldquoParameter estima-tion of q-Gaussian Radial Basis Functions Neural Networkswith a Hybrid Algorithm for binary classificationrdquo Neurocom-puting vol 75 pp 123ndash134 2012
[31] H Soh S Lim T Zhang et al ldquoWeighted complex networkanalysis of travel routes on the Singapore public transportationsystemrdquo Physica A vol 389 no 24 pp 5852ndash5863 2010
[32] C K S Kumar R Sukumar and M Nageswari ldquoSensorslifetime enhancement techniques in wireless sensor networksmdasha critical reviewrdquo International Journal of Computer Science andInformation Technology amp Security vol 3 pp 159ndash164 2013
[33] S Li and F Qin ldquoA dynamic neural network approach forsolving nonlinear inequalities defined on a graph and itsapplication to distributed routing-free range-free localizationof WSNsrdquo Neurocomputing vol 117 pp 72ndash80 2013
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ISRN Artificial Intelligence 13
[23] F Valdez P Melin and C Oscar ldquoAn improved evolution-ary method with fuzzy logic for combining Particle Swarmoptimization and genetic algorithmsrdquo Applied Soft ComputingJournal vol 11 no 2 pp 2625ndash2632 2011
[24] J Zhang and K Chau ldquoMultilayer ensemble pruning vianovel multi-sub-swarm particle swarm optimizationrdquo Journalof Universal Computer Science vol 15 no 4 pp 840ndash858 2009
[25] F Valdez P Melin and O Castillo ldquoParallel Particle Swarmoptimization with parameters adaptation using fuzzy logicrdquoin Proceedings of the 11th Mexican International Conference onArtificial Intelligence (MICAI rsquo12) vol 2 pp 374ndash385 San LuisPotosi Mexico October 2012
[26] J Peng and C P Y Pan ldquoParticle swarm optimization RBFfor gas emission predictionrdquo Journal of Safety Science andTechnology vol 7 pp 77ndash85 2011
[27] G Ren Z Huang Y Cheng X Zhao and Y Zhang ldquoAnintegrated model for evacuation routing and traffic signaloptimization with background demand uncertaintyrdquo Journal ofAdvanced Transportation vol 47 pp 4ndash27 2013
[28] X Feng Z Xiao and X Cui ldquoImproved RSSI algorithm forwireless sensor networks in 3Drdquo Journal of ComputationalInformation Systems vol 7 no 16 pp 5866ndash5873 2011
[29] F Valdez and P Melin ldquoNeural network optimization witha hybrid evolutionary method that combines particle swarmand genetic algorithms with fuzzy rulesrdquo in Proceedings ofthe Annual Meeting of the North American Fuzzy InformationProcessing Society (NAFIPS rsquo08) NewYork NYUSAMay 2008
[30] F Fernandez-Navarro C Hervas-Martınez P A Gutierrez JM Pena-Barragan and F Lopez-Granados ldquoParameter estima-tion of q-Gaussian Radial Basis Functions Neural Networkswith a Hybrid Algorithm for binary classificationrdquo Neurocom-puting vol 75 pp 123ndash134 2012
[31] H Soh S Lim T Zhang et al ldquoWeighted complex networkanalysis of travel routes on the Singapore public transportationsystemrdquo Physica A vol 389 no 24 pp 5852ndash5863 2010
[32] C K S Kumar R Sukumar and M Nageswari ldquoSensorslifetime enhancement techniques in wireless sensor networksmdasha critical reviewrdquo International Journal of Computer Science andInformation Technology amp Security vol 3 pp 159ndash164 2013
[33] S Li and F Qin ldquoA dynamic neural network approach forsolving nonlinear inequalities defined on a graph and itsapplication to distributed routing-free range-free localizationof WSNsrdquo Neurocomputing vol 117 pp 72ndash80 2013
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014