Upload
upender-dhull
View
218
Download
0
Embed Size (px)
Citation preview
8/7/2019 2.1-Genetic Algorithms
1/97
Genetic Al orithms
ENETIC LGORITHMS
8/7/2019 2.1-Genetic Algorithms
2/97
Introduction
Genetic algorithms are inspired by Darwin's theoryof evolution (i.e., survival of the fittest);
The new breeds of classes of living things come into existence
through the process of reproduction, crossover, and mutation
among existing organisms.
been translated into an algorithm to search for
solutions to problems in a more natural way.
8/7/2019 2.1-Genetic Algorithms
3/97
8/7/2019 2.1-Genetic Algorithms
4/97
In the process of improving performance as muchas possible via tuning and specializing the geneticalgorithm operators, new and important findingsregarding the generality, robustness, andapplicability of genetic algorithms becameavailable.
Following the last couple of years of furiousdevelo ment of enetic al orithms in the sciences,
engineering and the business world, thesealgorithms in various guises have now beensuccessfull a lied to o timization roblemsscheduling, data fitting and clustering, trend
spotting and path finding.
8/7/2019 2.1-Genetic Algorithms
5/97
Biological Background
Chromosome: are strings of DNA (genes) andserve as a model for the whole organism
ac gene enco es a par cu ar pro e n. as ca y,
it can be said that each gene encodes a trait, for
example color of eyes.
Possible settings for a trait (e.g. blue, brown) are
called alleles.
ac gene as ts own pos t on n t echromosome. This position is called locus.
8/7/2019 2.1-Genetic Algorithms
6/97
Complete set of genetic material (all chromosomes)is called genome.
Particular set of genes in genome is called
The genotype is the after birth base for the
organism'sphenotype, its physical and mentalcharacteristics, such as eye color, intelligence, etc.
8/7/2019 2.1-Genetic Algorithms
7/97
http://gslc.genetics.utah.edu/units/basics/tour/
8/7/2019 2.1-Genetic Algorithms
8/97
During reproduction, recombination(orcrossover)first occurs. Genes from parents combine to form a
The newly created offspring can then be mutated.u a on means a e e emen s o are a
changed. This changes are mainly caused by errorsin copying genes from parents.
The fitnessof an organism is measured by successof the or anism in its life survival .
8/7/2019 2.1-Genetic Algorithms
9/97
Biological Metaphor
Algorithm begins with a set of solutions (represented by
chromosomes) called population.
goodness (fitness) and used to form a new population,(reproduction). This is motivated by a hope, that the newpopulation will be better than the old one.
Solutions which are then selected to form new solutions.(Offspring) are selected according to their fitness - the
reproduce.
This is repeated until some condition (for examplenum er o popu a ons or mprovemen o e essolution) is satisfied.
8/7/2019 2.1-Genetic Algorithms
10/97
Flowchart
1. Generate random population ofNchromosomes (feasiblesolutions for the problem)
START
FITNESS
2. va ua e e ness x o eacchromosome x in the population
3. Create a new population by
GENERATE NEW
POPULATION
the new population is completeby means of selection andcrossover or mutation
REPLACE
4. Replace unfit individuals in oldpopulation by new off springs
5. If the end condition is satisfied,END
CONDITION
SATISFIED
NO
,in current population
else Go to step 2STOPYES
8/7/2019 2.1-Genetic Algorithms
11/97
An Illustrative Example
How a line may be fit through a given data set using agenetic algorithm?
ons er a nown mo e structure, a near t
y= C1x+ C2
Encode C and C in 6-bit strin each L = 6
E.g., 0 0 0 1 1 1 0 1 0 1 0 0 (genotype)
C1 = 7 C2 = 20
Initial population = 4
Assume the minimum ofC1 and C2 are Cmin = -2; and the
max mum o 1 an 2 are max = .
Evaluate the performance from the given data set Cut-off value = 0.8 (relative fitness)
8/7/2019 2.1-Genetic Algorithms
12/97
Mappingfrom genotype to phenotype Ci= Cmin,i+ b / (2
L-1) (Cmax,i- Cmin,i)
corresponding more copies in the new generation (i.e.,
strings with lower fitness values are eliminated) (i.e.,
Crossover: strings are able to mix and match theirdesirable qualities in a random fashion (i.e., one-point
crossover Mutation: helps to increase the searching space,
allowing a vital bit of information to vary (from 1 o 0 orfrom 0 to 1) (i.e., bit-flip mutation). Mutation takes place
very rarely (e.g., 0.005/bit/generation)
8/7/2019 2.1-Genetic Algorithms
13/97
8/7/2019 2.1-Genetic Algorithms
14/97
Another Illustrative Example
Given some functional mapping for a system, some membershipfunctions and their shapes are assumed for the various fuzzy
variables defined for a problem.
These membership functions are them coded as bit strings that are
then concatenated.
each set of membership functions.
Single-input (x) single-output (y) system:
y 1 4 9 16 25
Functional mapping in fuzzy logic for the system:
y S (small) VL (very large) [ ]250y
8/7/2019 2.1-Genetic Algorithms
15/97
The only parameter needed to described themembership functions (shape and position) is the
x
.S L
6-bit string (binary) to define the base 24-bit string
Base 1 Base 2 x
c romosome
Initial population = 4
8/7/2019 2.1-Genetic Algorithms
16/97
8/7/2019 2.1-Genetic Algorithms
17/97
Design Parameters
1. Design a representation2. Decide how to initialize a population
. es gn a mec an sm o map e p eno ype ogenotype and vice versa
4. Design a way of evaluating an individual5. Design suitable mutation operators
6. Design suitable crossover operators
. 8. Decide how to select individuals to be replaced
9. Decide when to sto the al orithm
10. Decide how to manage the population
8/7/2019 2.1-Genetic Algorithms
18/97
1. Designing a Representation
We have to come up with a method of representingan individual as a genotype.
The way we choose to do it must be relevantto the
.
When choosing a representation, we have to bear in
mind how the genotypes will be evaluated and whatthe genetic operators might be.
8/7/2019 2.1-Genetic Algorithms
19/97
Forn-city Traveling Salesman Problem,
how should we desi n a re resentation(phenotype vs. genotype)?
0.9
1
0 4
0.5
0.6
0.7
0.8
0 0.1 0.2 0. 3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.1
0.2
0.3
.
8/7/2019 2.1-Genetic Algorithms
20/97
Binary Valued Representation
A chromosome should in some way containinformation about solution that it represents. Themost used way of encoding is a binary string. Ac romosome en cou oo e s:
Chromosome 1 1101 1001 0011 0110 Chromosome 2 1101 1110 0001 1110
string. Each bit in the string can represent somecharacteristics of the solution. Another possibility isthat the whole strin can re resent a number
8/7/2019 2.1-Genetic Algorithms
21/97
Examples
PHENOTYPE
Integer
Real Number
8 bit GENOTYPE
.
.
ny ng
WHAT CAN IT REPRESENT ???
8/7/2019 2.1-Genetic Algorithms
22/97
Phenotype can be an integer number
=
0*22+1*21+1*20
= 128+32+2+1 = 163
8/7/2019 2.1-Genetic Algorithms
23/97
Phenotype can be areal number
e.g. a num er etween
2.5 and 20.5 using 8
binar di its
X= 2.5 + (163/256) (20.5-
2.5) = 13.9609
8/7/2019 2.1-Genetic Algorithms
24/97
Job Timestep
Phenotype can
be a schedule
1 2
2 1
e.g. 8 jobs , 2
time steps4 1
5 1
6 1
7 2
8/7/2019 2.1-Genetic Algorithms
25/97
Real Valued Representation
A very natural encoding if the solution we are lookingfor is a list of real-valued numbers, then encode as a
-
0s)
Lots of applications , e.g. parameter optimization
8/7/2019 2.1-Genetic Algorithms
26/97
Individuals are represented as a tuple ofn realvalued numbers:
The fitness function maps tuples of real numbers to
f:Rn R
8/7/2019 2.1-Genetic Algorithms
27/97
Binary vs. Real Representation
Original genetic algorithm is based on binary coding, due to
biological evidence. But it has some drawbacks when applied to
multidimensional, high-precision numerical problems.
There are three advantages of real representation:
First, the real representation proves adequate precision so that good values
are presentable in the search space. If a parameter is coded in binary form,ere s a ways e anger a one s a owe enoug prec s on o represen
parameter values that produce the best solution values.
Second, with the same digits, the real representation has larger range and
does not have to be a power of two.
Third, the real-coded genes have the ability to exploit the gradualness ofcontinuous variables (the gradualness means that small changes in the
variables correspond to small changes in the function).
disadvantages of binary one, there are still some weak points like
inherent premature convergence and leak mountain climbing.
8/7/2019 2.1-Genetic Algorithms
28/97
Order Based Representation
Individuals are represented as permutations Used for ordering, sequencing problems
Famous example: traveling salesman problem where
every city gets assigned a unique number from 1 to
. , , , , .
Famous example: N-queen problem (orN2-queen)
Needs special operators to make sure the individuals
stay in validpermutations.
8/7/2019 2.1-Genetic Algorithms
29/97
NQueens Problem
In chess, a queen can move as far
as she pleases, horizontally,vertically, or diagonally. A chess
.The standard 8 by 8 Queen'sproblem asks how to place 8
queens on an ordinary chess board
other in one move.
One solution - the prettiest in myopinion - is given in the figure
.12 essentially distinct solutions.(Two solutions are not essentiallydistinct if you can obtain one fromano er y ro a ng your c essboard, or placing in in front of a
mirror, or combining these twooperations.)
8/7/2019 2.1-Genetic Algorithms
30/97
Constraint SatisfactionProblem
8/7/2019 2.1-Genetic Algorithms
31/97
n 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Unique
solution
1 0 0 1 2 1 6 12 46 92 341 1787 9233 45752 285053
Distinct
solution
1 0 0 2 10 4 40 92 352 724 2680 14200 73712 365596 2279184
* Distinct solutions are derived from unique solutions by rotations,
reflections and transpositions
8/7/2019 2.1-Genetic Algorithms
32/97
Variants: Super Queen orN2 Queens (3-D version) problem
e requ res p ac ng queens n an
cube so that no two queens threaten each other. There can
only be one queen in any row parallel to any axis, or in any
. , ,
cube, so that Sijcontains a height, k, then S is clearly a
Latin square. A solution to 3DTN2QP is also a pandiagonal
Latin s uare or Knut-Vik desi n.
Demonstration (N: 4-60)
http://www.apl.jhu.edu/~hall/NQueens.html
8/7/2019 2.1-Genetic Algorithms
33/97
Tree Based Representation
Individuals in the population are trees Any expression can be drawn as a tree of function
an erm na s
Functions : sine, cosine, add, sub, if-then-else
, , . , , , ,
8/7/2019 2.1-Genetic Algorithms
34/97
Tree Based: Closure and Sufficiency
In the tree based representation we need to specifya function set and a terminal set. It is very desirable
By closure we mean that each of the functions in theunc on se s a e o accep as s argumen s anyvalue and data type that may possibly be returned bysome other function or terminal.
By sufficiencywe mean that there should be asolution in the s ace of all ossible ro ramsconstructed from the specified function and terminal
sets.
8/7/2019 2.1-Genetic Algorithms
35/97
2. Initialization of a Population
Usually, at random
Uniformly on the search space . ,if possible
Binary representation: 0 and 1 with probability of 0.5
-
(OK for bounded values only)
Seed the population with previously known values orthose from heuristics. With care:
Possible loss of enetic diversit
Possible unrecoverable bias
8/7/2019 2.1-Genetic Algorithms
36/97
Forn-city Traveling Salesman Problem,
an effective strate to initialize thepopulation, given the way how the solution is
coded?
8/7/2019 2.1-Genetic Algorithms
37/97
Randomly initialized path
er space- ng curve
Fast closest ath
8/7/2019 2.1-Genetic Algorithms
38/97
3. Mapping a Genotype to Phenotype
Sometimes producingthe phenotype from the Genotype Problem
Dataprocess.
Other times the
of parameters to somealgorithm, which work
Growt
Function
produce the phenotype.
Phenot e
8/7/2019 2.1-Genetic Algorithms
39/97
Forn-city Traveling Salesman Problem,
it should be strai htforward for the ma infrom genotype to phenotype.
8/7/2019 2.1-Genetic Algorithms
40/97
4. Evaluating an Individual
By far the most costlystep for real applications Do not re-evaluate unmodified individuals
It mi ht be a subroutine a black-box simulator or anexternal process (e.g., robot experiment).
The effectiveness of the process depends on the
choice of the fitness function. Your could use approximate fitness, but not far too long
(fitness inheritance orfitness approximation).
Constraint handlin - what if the henot e breaks
some constraint of the problem: Penalize the fitness
Specific evolutionary method
Multiobjective evolutionary optimization gives a set of
compromise solutions
8/7/2019 2.1-Genetic Algorithms
41/97
Forn-city Traveling Salesman Problem,
what would be the cost function?
8/7/2019 2.1-Genetic Algorithms
42/97
5. Mutation Operators
The mutation operator should allow every part of the
search space to be reached.
The size of mutation is important and should be
.
Mutation should produce validchromosomes.
8/7/2019 2.1-Genetic Algorithms
43/97
Forn-city Traveling Salesman Problem,
how should we desi n the mutation o erator?
0.9
1
0.4
0.5
0.6
0.7
.
0 0. 1 0. 2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.1
0.2
0.3
8/7/2019 2.1-Genetic Algorithms
44/97
Binary Valued Representation
1 1 1 1 1 1 1e ore
Grey Elephant
mutated gene
Mutation probabilitypm, for each gene
8/7/2019 2.1-Genetic Algorithms
45/97
Real Valued Representation
Perturb values by adding some random noise
Often, a Gaussian/normal distribution N(0,) is used,w ere
is the standard deviation
and xi = xi + N(0,i) for each parameter
8/7/2019 2.1-Genetic Algorithms
46/97
Order Based Representation
Randomly select two different genes and swap them.
7 83 41 2 6 5
7 83 46 2 1 5
bit swapping bit reversal
a sequence of bits is reversedsomewhere in the chromosome
8/7/2019 2.1-Genetic Algorithms
47/97
Tree Based Representation
Single point mutation selects one node and replaces
it with a similar one.
* *
2 * *r r r r
8/7/2019 2.1-Genetic Algorithms
48/97
Annealing Procedure Conjecture
Two fundamental parameters to be controlled: mutation rate and
degrees of variations
In early stages, allow mutation to occur frequent and
Toward the later stages (convergence), limit
occurrence in mutation and with small jumps for
exp o tat on
8/7/2019 2.1-Genetic Algorithms
49/97
6. Recombination/Crossover Operators
The child should inherit something from both.
mutation operator.
The recombination/crossover operator should bedesigned in conjunction with the representation sothat recombination is not always catastrophic.
Recombination should produce validchromosomes.
8/7/2019 2.1-Genetic Algorithms
50/97
Forn-city Traveling Salesman Problem,
how should we desi n the crossover o erator?
8/7/2019 2.1-Genetic Algorithms
51/97
Binary Valued Representation
. . .Whole Po ulation:
1 1 1 1 1 1 1 0 0 0 0 0 0 0parents
offspring1 1 1 0 0 0 0 0 0 0 1 1 1 1
1-point crossover
8/7/2019 2.1-Genetic Algorithms
52/97
Real Valued Representation
Uniform crossover: given two parents one child iscreated as follows
a db fc e g hg
a c eH
8/7/2019 2.1-Genetic Algorithms
53/97
Intermediate recombination (arithmetic crossover):given two parents one child is created as follows
a c e
FD ECBA
(a+A)/2 (b+B)/2 (c+C)/2 (e+E)/2(d+D)/2 (f+F)/2
o1 =p1*bias + p2*(1-bias)
* *2 1 - 2
bias toward the centerbetween two parents
8/7/2019 2.1-Genetic Algorithms
54/97
Blend (BLX-) Crossover:
picking a random number between min(p1,p2)- * and
*1, 2 , 1 2 ,
positive number (typically 0.5), and is |p1 p2|.
This formulation allows for the BLX operator to choose a
two parents. This allows for better placement of children
than the Arithmetic Crossover; however, the children will still
be biased to lie on baselines between sets of arents.
Therefore, children could still lack diversity.
iiii dxxX = ),min(211
21xxd =
iiii dxxX += ),max(212
8/7/2019 2.1-Genetic Algorithms
55/97
Unimodal Normally Distributed (UNDX) Crossover: It first selects three parents at random from the population.
Next it finds the mid oint of the first two arents and calls it xp.Then, it finds the difference vector of the first two parents as d= x1
x2. The line containing the first two parents is called the primary
search line, and the value D is computed as the distance from the
.
combined to form a child xc:
=
++=1
1
n
i
ii
ceDdmx
i
space and and are standard deviations determinedempirically.
This crossover operator has the advantage of preserving the mean
vector and covariance matrix of the parent population thereby
maintaining a similar distribution to the parent population in child
generations.
8/7/2019 2.1-Genetic Algorithms
56/97
Order Based Representation
Choose an arbitrary part from the first parent and
copy this to the first child.
Copy the remaining genes that are not in the copied
part to the first child: starting right from the cut point of the copied part
using the order of genes from the second parent
wra in around at the end of the chromosome
Repeat this process with the parent roles reversed.
8/7/2019 2.1-Genetic Algorithms
57/97
7 83 41 2 6 5 78 16 5234
Parent 1 Parent 2
7, 3, 4, 6, 5
81 2 4, 3, 6, 7, 5
Child 1
7 85 41 2 3 6
8/7/2019 2.1-Genetic Algorithms
58/97
Tree Based Representation
*
*
*
+ 2 * (r * r )
r / * r + r
1 r
wo su - rees are se ec e
for swapping.
8/7/2019 2.1-Genetic Algorithms
59/97
+
*
*
*
r / r r*
1 r* 2 +
2 * r /
r r 1 r
expressions
8/7/2019 2.1-Genetic Algorithms
60/97
Crossover vs. Mutation
Crossover modifications depend on the whole population
decreasing effects with convergence
exploitation operator
Mutation mandatory to escape local optima
exploration operator
emp as ze crossover; w e anemphasize mutation
8/7/2019 2.1-Genetic Algorithms
61/97
Exploration vs. Exploitation
Exploration = sample unknown regions
Too much exporation = random search, no
convergence
- -individuals
Too much expoitation = local search only
convergence to a local optimum
8/7/2019 2.1-Genetic Algorithms
62/97
7. Selection Strategy
We want to have some way to ensure that better
fitted individuals have a better chance of being
.
This will give us selection pressure which will drive
the o ulation forward. On the other hand, we have to be careful to give less
good individuals at least some chance of being
-
material.
Risk of loss diversit
8/7/2019 2.1-Genetic Algorithms
63/97
Forn-city Traveling Salesman Problem,
how should we desi n the selection o erator?
8/7/2019 2.1-Genetic Algorithms
64/97
Roulette Wheel Selection
Parents are selected according to their
fitness. The better the chromosomes are,the more chances they are to be selected.
The size of the section in the roulette
wheel is proportional to the value of thefitness function of every chromosome - thebigger the value is, the larger the sectionis.
Better (fitter) individuals have: more space more chances to be selected
8/7/2019 2.1-Genetic Algorithms
65/97
Fitness Proportionate Selection
Disadvantages:
Danger of premature convergence becauseoutstanding individuals take over the entire
o ulation ver uickl .
Low selection pressure when fitness values are near.
Behave differentl on trans osed versions of thesame function.
8/7/2019 2.1-Genetic Algorithms
66/97
Fitness Scaling
At the start, a few extraordinary individuals will dominate the
evolution process premature convergence Later on, the population average fitness may be closed to the
population best fitness random walk among the mediocre
a cure for the problem issue
Start with the raw fitness function
Standardize to ensure
Lower fitness is better fitness
Optimal fitness equals to 0
Fitness ranges from 0 to 1 Normalize to ensure
The sum of the fitness values equals to 1
Linear scaling: f = af+ b
a and b are chosen such that favg = favg
8/7/2019 2.1-Genetic Algorithms
67/97
Tournament Selection
Select k random individuals, without replacement
a e t e est
k is called the size of the tournament
8/7/2019 2.1-Genetic Algorithms
68/97
8/7/2019 2.1-Genetic Algorithms
69/97
Ranked Based Selection
Individuals are sorted on their fitness value from best
to worse. e p ace n t s sorte st s ca e ran .
,rank is used by a function to select individuals from
this sorted list. The function is biased towards
individuals with a high rank (= good fitness).
8/7/2019 2.1-Genetic Algorithms
70/97
Fitness: f(A) = 5, f(B) = 2, f(C) = 19
Rank: r(A) = 2, r(B) = 3, r(C) = 1
h(x) = min + (max min)* (r(x) 1)/(n 1)
unc on: = , = , =
If applied, proportion on the roulette wheel:= = =. , . , .
8/7/2019 2.1-Genetic Algorithms
71/97
8. Replacement Strategy
The selection pressure is also affected by the way in
which we decide which members of the population to
individuals.
e can use e s oc as c se ec on me o s nreverse, or there are some deterministic replacementstrategies.
We can decide never to replace the best in theo ulation: elitism.
8/7/2019 2.1-Genetic Algorithms
72/97
Forn-city Traveling Salesman Problem,
how should we desi n the re lacementstrategy?
8/7/2019 2.1-Genetic Algorithms
73/97
Elitism
Should fitness constantly improves?
Re-introduce in the population previous best-so-far (elitism)or
Keep best-so-far in a safe place (preservation)
Theory GA: preserve mandatory ES: no elitism sometimes is better
Application: avoid users frustration
8/7/2019 2.1-Genetic Algorithms
74/97
Bottomline is survival of the fittest
Generational Replacement Genetic Algorithm (GRGA).
approach of Holland. If the original population size is N, then Noffspring
are generated, and they are combined with the parents to get 2N
individuals which are then sorted and the best Nindividuals are accepted
for the next generation. This approach is rather simple, and places the
entire burden of selection and extinction upon the reproductive
operations.
ea y a e ene c gor m
uses a fixed population wherein a finite fraction of the population isextinguished at every generation. This approach separates natural
,
replacement strategy. Whereas there is no choice of replacement
strategy for the GRGA (because everyone is replaced), there are choices
for replacement strategy in the SSGA.
8/7/2019 2.1-Genetic Algorithms
75/97
Replace Worstreplaces the worst existing chromosome in the
population with probability equal to one by the newly created child
chromosome in the population by the newly created child to
preserve diversity
Replace Random replaces a random existing chromosome in the
population.
,
of parents at random and replace the worst, or pick two parents atrandom and replace the worst of the two, with or without some
attached robabilit check. This can also be combined with roulette
wheel selection where the two candidates are chosen by roulette
wheel probabilities based upon absolute fitness, and then proceed
with the tournament. This is known as a Stochastic Tournament.
8/7/2019 2.1-Genetic Algorithms
76/97
Population Decimation All chromosomes below the cut-off
threshold do not survive to the next generation. Their replacementsare formed by mating the remaining individual chromosomes.
Mass Extinction Extinction allows the repopulation of niches and
gives room for new adaptations. In mass extinction models, major
parts of the population are occasionally replaced, as opposed to agradual substitution of single individuals that one might see in a
classical GA. Most of the time there are no extinctions, sometimes
a few individuals die, and very rarely a large proportion of the
population is affected. However, the current best individual (an elite
class of size one) always survives extinction events, and is oftenused to repopulate the extinguished individuals in the extinction
zones y mu a e e ana ogy n na ure s e re-co on za on o a
niche by fit individuals that quickly explore the environment and
adapt to it.
8/7/2019 2.1-Genetic Algorithms
77/97
9. Stopping Criteria
The optimum is reached !!!!!
m on resources
Maximum number of evolution generations
Maximum number of fitness evaluations
m t on t e user s pat ence
After some generations without improvement
8/7/2019 2.1-Genetic Algorithms
78/97
Forn-city Traveling Salesman Problem,
what would be a reasonable sto in criterion?
8/7/2019 2.1-Genetic Algorithms
79/97
In general, generation number, evolution time,
fitness threshold, fitness convergence are indicators.
Population convergence The population is deemed
as conver ed when the avera e fitness across thecurrent population is less than a user-specified
percentage away from the best fitness of the current
.
Gene convergence This is a method that stops theevolution when a user-specified percentage of the
genes are deemed converged.
8/7/2019 2.1-Genetic Algorithms
80/97
10. Managing the Population
In general, fix population is always used for simplicity.
If the predetermined population size is too small, there will not be-
solution. If the population size is too large, it may require
unnecessarily large computational resources and result into an
extremely long running time. Optimal population size?
For MOP, a fixed population size will have great difficulty in
obtainin a Pareto front with a desired resolution because the
size and shape of the true Pareto front is unknown in apriorifor
most of the MOPs
A d namic o ulation size will be more reasonable for MOEA if
the computational effort can be adaptively adjusted based on the
complexity of the problem
8/7/2019 2.1-Genetic Algorithms
81/97
Performance
Neverdraw any conclusion from a single run
Use statistical measure (average, median) (Box plot)
rom a su c en num er o n epen en runs - m n mum
From the application point of view
Design perspective
Find a very goodsolution at least once
Production ers ective
Find a goodsolution at almost every run
,
performance on toy data and expect it to work with real
data
8/7/2019 2.1-Genetic Algorithms
82/97
Box Plot
In descriptive statistics, a boxplot (alsoknown as a box-and-whisker diagram or
plot) is a convenient way of graphicallydepicting groups of numerical data through
e r ve-num er summar es e sma esobservation, lower quartile (Q1), median(Q2), upper quartile (Q3), and largestobservation). A boxplot may also indicate
, ,considered outliers. The boxplot wasinvented in 1977 by the Americanstatistician John Tukey.
Box lots can be useful to dis ladifferences between populations without
making any assumptions of the underlyingstatistical distribution. The spacingsbetween the different parts of the box helpn ca e e egree o spers on spreaand skewness in the data, and identifyoutliers. Boxplots can be drawn either
horizontally or vertically.
8/7/2019 2.1-Genetic Algorithms
83/97
For this data set:
" "- lower (first) quartile (Q1,x.25) = 7
median(second quartile) (Med,x.5) = 8.5 u er third uartile 3 x = 9 .
largest non-outlier observation = 10 (right "whisker")
the value 3.5 is a "mild" outlier, between
the value 0.5 is an "extreme" outlier,
the data are skewed to the left (negatively skewed)
the mean value of the data can also be labeled with a point.
8/7/2019 2.1-Genetic Algorithms
84/97
Flowchart
1. Generate random population of
Nchromosomes (feasiblesolutions for the problem)
START
FITNESS
2. va ua e e ness x o eacchromosome x in the population
3. Create a new population byGENERATE NEW
POPULATION
the new population is completeby means of selection andcrossover or mutation
REPLACE
4. Replace unfit individuals in old
population by new off springs5. If the end condition is satisfied,
END
CONDITION
SATISFIED
NO
,in current population
else Go to step 2STOP
YES
8/7/2019 2.1-Genetic Algorithms
85/97
Genetic Algorithm for TSP
Chromosome representation: (order-based)
Population size & initialization: (50, at random)
Fitness function: (distance traveled)
Mutation: (5%)
ecom na on: (order-based)
Selection: (rank-based)
Stopping criteria: (number of generations- 200)
8/7/2019 2.1-Genetic Algorithms
86/97
Demonstration
C S d 101 i TSP bl ( il101)
8/7/2019 2.1-Genetic Algorithms
87/97
Case Study: 101-city TSP problem (eil101)
Total Distance:678.854309.
Population:50
Generations:200
C St d 535 i t bl ( li535)
8/7/2019 2.1-Genetic Algorithms
88/97
Case Study: 535-airport problem (ali535)
Total Distance:
2433.995486
Population:50
Generations:
C St d N Q P bl
8/7/2019 2.1-Genetic Algorithms
89/97
Case Study: N Queens Problem
In chess, a queen can move as faras she pleases, horizontally,vertically, or diagonally. A chess
.The standard 8 by 8 Queen'sproblem asks how to place 8queens on an ordinary chess board
other in one move.
One solution - the prettiest in myopinion - is given in the figure
.12 essentially distinct solutions.(Two solutions are not essentiallydistinct if you can obtain one fromano er y ro a ng your c essboard, or placing in in front of amirror, or combining these two
operations.)
8/7/2019 2.1-Genetic Algorithms
90/97
Applications in
Traffic control
VLSI design
Deadlock prevention
Parallel memory storage schemes
Sensor deployment
.
Proble Iss es in GA
8/7/2019 2.1-Genetic Algorithms
91/97
Problem Issues in GA
Population size
Binary representation vs. real representation Population initialization
Noisy fitness function Stochastic fitness function (or dynamic environment) Fitness inheritance and fitness approximation
Selection/rankin strate Crossover/Recombination operator
Mutation operator
Replacement strategy
Elitism strategy Benchmark test functions
Ex loration vs. ex loitation dilemma
Constraint handling Diversity promotion
Population management
Homework #3 due 11/7/2009
8/7/2019 2.1-Genetic Algorithms
92/97
Homework #3- due 11/7/2009
Problem #1 (Combinatorial Optimization)
Develop a genericgenetic algorithm to solve the traveling
within a unit square in a 2-dimensional plane. The coordinates
of 20 cities are given below in a matrix:
=95,0.2867.7148,0.239,0.2184,09697,0.597,0.8274,0.740,0.50290.9500,0.6
67,0.8148.6091,0.877,0.9536,00398,0.136,0.2124,0.695,0.59060.6606,0.9cities
36,0.8668.1247,0.169,0.9464,07471,0.544,0.3429,0.041,0.02130.3876,0.7
Show the best route you find and the associated distance with
46,0.6768.4351,0.862,0.8191,08192,0.939,0.3025,0.296,0.16490.8200,0.3
attached computer coding (with documentation- show your
recipe). An example is given below for reference.
8/7/2019 2.1-Genetic Algorithms
93/97
0. 9
1
0. 7
0. 8
0. 4
0. 5
.
0. 1
0. 2
0. 3
0 0. 1 0. 2 0.3 0. 4 0 .5 0. 6 0.7 0. 8 0. 9 1
8/7/2019 2.1-Genetic Algorithms
94/97
Problem 2 (Scalability)
Extend your genetic algorithm to solve the benchmark 101-city
Eilson).The benchmark problem can be found from TSPLIB
archive at
http://www.iwr.uni-heidelberg.de/groups/comopt/software/TSPLIB95/
o nee o urn n e co es. n y e es rou e oun .e., s
configuration and distance in the form of above figure) is to beturned in. This problem is to test if your algorithm can scale up
.
8/7/2019 2.1-Genetic Algorithms
95/97
Problem 3 (Numerical Optimization)
Apply the genetic algorithm to De Jongs Rosenbrock's Valley
where global minimum is at the origin. This function is known
2
1
2
2
2
121 )1()(100),( xxxxxF +=
. .global optimum is at the tip of the sharp ridge, which is
parabolic shaped flat valley. The difficulty of this problem is to
conver e to the lobal o timum.
8/7/2019 2.1-Genetic Algorithms
96/97
Problem 4 (Multimodel Function)
Apply the genetic algorithm to De Jongs Shekels Foxholes
( ) =
+=25
1
5
1002.0
j jfF x ( )
=
+=2
1
6
,
i
jiij axjf
=
321601632321601632321601632321601632321601632a
This function, a multimodel function, which has many local
323232323216161616160000016161616163232323232
2,1,536.16536.16where = ixi
optimal (in this case 25). This function is a challenge to many
standard optimization algorithms because they can get stuck ina local optimal. Note that all 25 peaks have the same base
, , , , ... , - , -
32).
8/7/2019 2.1-Genetic Algorithms
97/97