Upload
zao
View
220
Download
5
Embed Size (px)
Citation preview
A New Method of Dynamic Decision in Knowledge System
Yibo Yang 1 1 Department of Computer Information,
Hubei Institute of Economics and Management WuHan, 430079, China
e-mail: [email protected]
Tao Lv 2 2 School of Computer Science & Engineering,
WuHan Institute of Technology WuHan, 430073, China
e-mail: [email protected]
Abstract—In most knowledge system, the decision mechanism of knowledge system can not update according to the environment. A new method of dynamic decision in knowledge system is proposed. The standardized attributes knowledge and scheduling rules knowledge are acquired by the system from bids and experience of experts.
Keywords-knowledge system,decision,dynamic
I. INTRODUCTION
The knowledge system should response with the changing market environment rapidly by internal flexibility[1]. The frequency of uncertain incidents is distinctly higher in knowledge system environment than in traditional one. As a result, the knowledge system scheduling mostly behaves as dynamic scheduling. Through reuse of the acquired knowledge, the knowledge-based scheduling reduces the load of manpower and achieves the agility of decision. Nevertheless, there are some constraints during the application of most systems, such as that the knowledge sources have their own subjective tendentiousness[2], scheduling decision making relies on partial attributes[3-4], and the knowledge can not be updated with the dynamic environment. All of these are opposite with the agile or flexible system.
Direct to these problems, this paper presents a dynamic decision method. Agents in multi agent systems are highly autonomous, intelligent, interactive and adaptive. The reinforcement learning principle is used to update the knowledge, and then it can update itself as long as scheduling.
II. ATTRIBUTE KNOWLEDGE According to the information of tasks and the capacity
of resources, condition attributes may be classified into these kinds[2]: (1) Time, (2) Cost, (3) Quality, (4) Load, (5) Priority. They are defined as below: (1) Time:
① ),( jie TAT ≤ ),( jpe TAT ),(1 ji TATk− , it means iA
can accomplish jT ahead of schedule.
② ),( jpe TAT ),(1 ji TATk− ≺ ),( jie TAT ),( jpe TAT≤ ,
it means iA can accomplish jT on schedule. k1 is a
constant, and 10 1 << k .
(2)Cost: ① ),( ji TAC ≤ ),( jp TAC , it means iA can accomplish
jT under the standard of cost.
② ),( ji TAC ),( jp TAC , it means iA can accomplish
jT beyond the standard of cost.
(3)Quality: ①
31),(),(
kTAQTAQ
jp
ji −≤ , it means iA can accomplish jT
on a high quality level. ②
21),(),(
kTAQTAQ
jp
ji −≤ , it means iA can accomplish jT
on an equal quality level to requirement. ③ 1
),(),(
=jp
ji
TAQTAQ , it means iA can accomplish jT on a
low quality level. k2,k3 are constants, and 10 32 <<< kk .
(4)Load: ① ),((),( 4 jpejpsic TATkTATLT +≤+ )),( jps TAT− , it
means iA is insufficiently loaded.
②
),((),( 4 jpejpsic TATkTATLT ++ )),( jps TAT− , it
means iA is fully loaded. k4 is a constant, and 10 4 << k .
(5)Priority: ① 1)( ≤jTρ , it means jT is a crucial task.
② 1)( jTρ , it means jT is not crucial task.
2009 Pacific-Asia Conference on Circuits,Communications and System
978-0-7695-3614-9/09 $25.00 © 2009 IEEE
DOI 10.1109/PACCS.2009.37
575
The information of Resource Agent capacity, which is shown in bid, is transformed into knowledge of standardized condition attributes according to above definitions. They are used for antecedent of scheduling rules.
Moreover, the evaluation to Resource Agent’s bid is regarded as decision attribute. It is used for consequent of rule and also assorted into three levels:
① Precedence, means it is perfect ② Feasible, means it is feasible. ③ Defective, means there is some defection in
somewhere.
III. RULE KNOWLEDGE
(1) Acquisition of rules. After the standardization of experience and scheduling
decisions selected from some experts, some scheduling decision tables are generated: mB,...,B,B 21 , which contain
condition attributes a、b、c、d、e and decision attribute f . During the selection, as more as possible experts and combinations of condition attributes in each table should be included to achieve the credibility and integrality of knowledge rules.
(2)Fusion of rules. F is defined as the set of decision attribute’s value, so
f3}f2,{f1,F = .To a combination: )( iiiii ,e,d,c,ba , the
decision attribute value varies in m decision tables. For this, its decision attribute if should be selected from F . In other words, the rule iiiiii fedcba ⇒),,,,( should be
decided. The method of ballot is adopted, and the process of fusion is shown as follows:
∈∀ kf F , )f( kADi is defined as the support degree
of decision attribute kf :
i
ii N
nkkAD =)f( .
In this function, iN means the number of decision
table which contains the combination ),,,,( iiiii edcba ,
and ink is regarded as the number of decision table whose
decision attributes’ value is kf among above iN ones,
mNnk ii ≤≤≤0 .
Appoint fl as that: kl ff = , and )})3f(),2f(),1f(max{)f(( iiii ADADADkAD =
If lf is sole, lf i f= , and the rule is gotten:
ledcba iiiii f),,,,( ⇒ ;
If not, we suppose that: f3f2f1 >> , and choose the highest one from lf as if . As a result, the rule is generated as: iiiiii fedcba ⇒),,,,( .
iCF is defined as this rule’s credibility, and
)}3f(),2f(),1f(max{ iiii ADADADCF = . The above fusion operations are executed to each
combination.
IV. WEIGHT OF ATTRIBUTE
Many knowledge systems have some limitation, such as that the knowledge sources have their own subjective tendentiousness, and the scheduling decision knowledge can not be updated in time. Thus, the knowledge system must has knowledge update mechanism, and the system can apperceive the change of the environment, then the knowledge based system will be more reliable in scheduling decision. In this paper, we present a knowledge updating strategy which is based on reinforcement learning principle.
During the reasoning from condition attributes to decision attribute, each condition attribute plays a role with different weight, which can be reflected by attribute’s importance according to its definition in Rough Sets:
)F(/)F( },,,,{}}\{,,,,{ edcbaedcba POSPOS ββμ =
),,,,( edcba=β . In above function, )F(}}\{,,,,{ βedcbaPOS stands for the
number of determinate decisions which are made by combinations of condition attributes without β included;
)F(},,,,{ edcbaPOS stands for the number of determinate
decisions which are made by combinations of complete condition attributes.
Through standardizing condition attribute’s importance, each attribute’s weight is figured out asω :
∑=
=
},,,,{ edcbaiiμ
μω β
β, ),,,,( edcba=β
V. PROCESSING ARITHMETIC
(1) The scheduling agent Ap disintegrates the task into many subtasks, TS={T1,T2,…,Tn}.
(2) According to the technological requirements of the subtask, the scheduling agent Ap issues the tasks. All resource agents check their capability and refer the bidding document.
(3) The bid-set iBD is constructed according to the
bidding document. If iBD is null, adjust the tasks or
576
increase the resources, and then return step 1. (4) For every resource agent of iBD , the values of
conditional attributes are calculated and normalized into discrete values. Every resource agent in iBD has a
combination of ),,,,( jjjjjj edcbau = and ( )fkuv j ,
correspondingly. (5) If there is only one resource agent, it is bid-winner.
If there are many resource agents, first the resource agents’ ( ) ( ) ( )}3,,2,,1,max{ fuvfuvfuv jjj are selected. We
stipulate the PRI as f1>f2>f3. If fk is unique, it is bid-
winner. If not, we choose the resource agent which has the largest ( )fkuv j , as bid-winner.
VI. CONCLUSIONS
This dynamic decision method overcomes some limitation. The updating can be implemented itself. In the course of acquiring rule knowledge, conflicts are solved by ballot. The initial ( )fkuv i , are got according to the support
degree of decision attribute, and it can reflect the experts’ experience. The reinforcement learning principle is applied to update the scheduling knowledge. When giving reward, the bidirectional feedback method is used for the negative reward, and it can improve the perception ability with the changing environment. At the same time, the weight of conditional attribute is considered in reward, and then the knowledge updating will be more reasonable. Also there are many problems that should be studied more. For example, more estimate methods of scheduling result should be explored, and the sensitivity of the perception to environment should be improved further. So in our later study, we will pay more attention to the above problems.
REFERENCES
[1] Bao Zhenqiang, Wang Ningsheng, Cai Zongtan. Research on the model
of scheduling agent based on bid using Rough-Fuzzy sets[J]. Machine
Engineering of China. 2003, 14(22): 1943-1946.
[2] Miyashita, Kazuo. Knowledge-level Analysis for Eliciting Composable
Scheduling Knowledge[J]. Artificial Intelligence in Engineering. 1995,
9(4): 253-264.
[3] Dubois Didier, Fargier Helene, Fortemps Philippe. Fuzzy Scheduling:
Modeling Flexible Constraints vs. Coping with Incomplete Knowledge[J].
European Journal of Operational Research. 2003, 147(2): 231-252.
[4] Leslie Pack Kaelbling, Michael L. Littman. Reinforcement Learning: A
Survey[J]. Journal of Artificial Intelligence Research. 1996, 4:237-285
577