10 genetic algorithms 10 1 the basic algorithm
play

10. Genetic Algorithms 10.1 The Basic Algorithm General-purpose - PowerPoint PPT Presentation

T79.4201 Search Problems and Algorithms T79.4201 Search Problems and Algorithms 10. Genetic Algorithms 10.1 The Basic Algorithm General-purpose black-box optimisation method We consider the so called simple genetic


  1. T–79.4201 Search Problems and Algorithms T–79.4201 Search Problems and Algorithms 10. Genetic Algorithms 10.1 The Basic Algorithm ◮ General-purpose “black-box” optimisation method ◮ We consider the so called “simple genetic algorithm”; also proposed by J. Holland (1975) and K. DeJong (1975). many other variations exist. ◮ Method has attracted lots of interest, but theory is still ◮ Assume we wish to maximise a cost function c defined on incomplete and the empirical results inconclusive. ◮ Advantages: general-purpose, parallelisable, adapts n -bit binary strings: incrementally to changing cost functions (“on-line c : { 0 , 1 } n → R . optimisation”). ◮ Disadvantages: typically very slow – should be used with Other types of domains must be encoded into binary moderation for simple serial optimisation of a stable, easily strings, which is a nontrivial problem. (Examples later.) evaluated cost function. ◮ View each of the candidate solutions s ∈ { 0 , 1 } n as an ◮ Some claim that GA’s typically require fewer function individual or chromosome . evaluations to reach comparable results as e.g. simulated ◮ At each stage ( generation ) t the algorithm maintains a annealing. Thus the method may be good when function population of individuals p t = ( s 1 ,... , s m ) . evaluations are expensive (e.g. require some acutal physical measurement). I.N. & P .O. Autumn 2006 I.N. & P .O. Autumn 2006 T–79.4201 Search Problems and Algorithms T–79.4201 Search Problems and Algorithms Three operations defined on populations: ◮ selection σ ( p ) (“survival of the fittest”) ◮ recombination ρ ( p ) (“mating”, “crossover”) ◮ mutation µ ( p ) Selection (1/3) Denote Ω = { 0 , 1 } n . The selection operator σ : Ω m → Ω m maps The Simple Genetic Algorithm : populations probabilistically: given an individual s ∈ p , the expected number of copies of s in σ ( p ) is proportional to the function SGA( σ , ρ , µ ): p ← random initial population; fitness of s in p . This is a function of the cost of s compared to the costs of other s ′ ∈ p . while p “not converged” do p ′ ← σ ( p ) ; p ′′ ← ρ ( p ′ ) ; p ← µ ( p ′′ ) end while ; return p (or “fittest individual” in p ). end . I.N. & P .O. Autumn 2006 I.N. & P .O. Autumn 2006

  2. T–79.4201 Search Problems and Algorithms T–79.4201 Search Problems and Algorithms Selection (3/3) Selection (2/3) Once the fitness of individuals has been evaluated, selection Some possible fitness functions: can be performed in different ways: ◮ Relative cost ( ⇒ “canonical GA”): ◮ Roulette-wheel selection (“stochastic sampling with c ( s ) � c ( s ) replacement”): f ( s ) = c . 1 ¯ m ∑ ◮ Assign to each individual s ∈ p a probability to be selected in c ( s ′ ) proportion to its fitness value f ( s ) . Select m individuals according s ′ ∈ p to this distribution. ◮ Pictorially: Divide a roulette wheel into m sectors of width ◮ Relative rank : proportional to f ( s 1 ) ,... , f ( s m ) . Spin the wheel m times. r ( s ) 2 ◮ Remainder stochastic sampling : f ( s ) = = m + 1 · r ( s ) , 1 m ∑ ◮ For each s ∈ p , select deterministically as many copies of s as r ( s ′ ) indicated by the integer part of f ( s ) . After this, perform stochastic s ′ ∈ p sampling on the fractional parts of the f ( s ) . ◮ Pictorially: Divide a fixed disk into m sectors of width proportional where r ( s ) is the rank of individual s in a worst-to-best ordering of all s ′ ∈ p . to f ( s 1 ) ,... , f ( s m ) . Place an outer wheel around the disk, with m equally-spaced pointers. Spin the outer wheel once. I.N. & P .O. Autumn 2006 I.N. & P .O. Autumn 2006 T–79.4201 Search Problems and Algorithms T–79.4201 Search Problems and Algorithms Recombination (2/2) Recombination (1/2) Possible crossover operators: ◮ Given a population p , choose two random individuals ◮ 1-point crossover: s , s ′ ∈ p . With probability p ρ , apply a crossover operator ρ ( s , s ′ ) to produce two new offspring individuals t , t ′ that 1 1 0 1 0 0 1 1 0 0 1 0 1 1 0 0 0 1 1 0 0 1 replace s , s ′ in the population. 0 1 1 0 1 0 1 1 0 1 1 1 1 0 1 1 0 1 1 0 1 1 ◮ 2-point crossover: ◮ Repeat the operation m / 2 times, so that on average each individual participates once. Denote the total effect on the 1 1 0 1 0 0 1 1 0 0 1 1 1 1 0 1 0 1 1 0 0 1 population as p ′ = ρ ( p ) . 0 11 0 1 0 1 1 0 1 1 0 1 0 1 0 0 1 1 0 1 1 p ρ ◮ Practical implementation: choose 2 · m random pairs from p and apply crossover deterministically. ◮ uniform crossover: ◮ Typically p ρ ≈ 0 . 7 ... 0 . 9 . 1 1 0 1 0 0 1 1 0 0 1 0 1 0 1 1 0 1 1 0 0 1 0 1 1 0 1 0 1 1 0 1 1 1 1 0 1 1 0 1 1 0 0 1 I.N. & P .O. Autumn 2006 I.N. & P .O. Autumn 2006

  3. T–79.4201 Search Problems and Algorithms T–79.4201 Search Problems and Algorithms 10.2 Analysis of GA’s Mutation Hyperplane sampling (1/4) ◮ Given population p , consider each bit of each individual ◮ A heuristic view of how a genetic algorithm works. and flip it with some small probability p µ . Denote the total ◮ A hyperplane (actually subcube) is a subset of Ω = { 0 , 1 } n , effect on the population as p ′ = µ ( p ) . where the values of some bits are fixed and other are free ◮ Typically, p µ ≈ 0 . 001 ... 0 . 01 . Apparently good choice: to vary. A hyperplane may be represented by a schema p µ = 1 / n for n -bit strings. H ∈ { 0 , 1 , ∗} n . ◮ Theoretically mutation is disruptive. Recombination and ◮ E.g. schema ’ 0 ∗ 1 ∗∗ ’ represents the 3-dimensional selection should take care of optimisation; mutation is hyperplane (subcube) of { 0 , 1 } 5 where bit 1 is fixed to 0, bit needed only to (re)introduce “lost alleles”, alternative 3 is fixed to 1, and bits 2, 4, and 5 vary. values for bits that have the bits that have the same value ◮ Individual s ∈ { 0 , 1 } n samples hyperplane H , or matches the in all current individuals. corresponding schema if the fixed bits of H match the ◮ In practice mutation + selection = local search. Mutation, corresponding bits in s . (Denoted s ∈ H .) even with quite high values of p µ , can be efficient and is ◮ Note : given individual generally samples many hyperplanes often more important than recombination. simultaneously, e.g. individual ’ 101 ’ samples ’ 10 ∗ ’, ’ 1 ∗ 1 ’, etc. I.N. & P .O. Autumn 2006 I.N. & P .O. Autumn 2006 T–79.4201 Search Problems and Algorithms T–79.4201 Search Problems and Algorithms Hyperplane sampling (2/4) Hyperplane sampling (3/4) Consider e.g. the following cost function and partition of Ω into ◮ order of hyperplane H : hyperplanes (in this case, intervals) of order 3: o ( H ) = number of fixed bits in H = n − dim H c(s) ◮ average cost of hyperplane H : 1 2 n − o ( H ) ∑ c ( H ) = c ( s ) s ∈ H ◮ m ( H , p ) = number of individuals in population p that sample hyperplane H . 000** 001** 010** 011** 100** 101** 110** 111** Ω ◮ average fitness of hyperplane H in population p : Here the current population of 21 individuals samples the 1 m ( H , p ) ∑ hyperplanes so that e.g. ’ 000 ∗∗ ’ and ’ 010 ∗∗ ’ are sampled by f ( H , p ) = f ( s , p ) s ∈ H ∩ p three individuals each, and ’ 100 ∗∗ ’ and ’ 101 ∗∗ ’ by two individuals each. Hyperplane ’ 010 ∗∗ ’ has a rather low average Heuristic claim : selection drives the search towards hyperplanes fitness in this population, whereas ’ 111 ∗∗ ’ has a rather high of higher average cost (quality). average fitness. I.N. & P .O. Autumn 2006 I.N. & P .O. Autumn 2006

  4. T–79.4201 Search Problems and Algorithms T–79.4201 Search Problems and Algorithms Hyperplane sampling (4/4) The effect of crossover on schemata (1/2) Then the result of e.g. roulette wheel selection on this ◮ Consider a schema such as population might lead to elimination of some individuals and duplication of others: H = ∗∗ 11 ∗∗ 01 ∗ 1 ∗∗ � �� � ∆( H )= 7 c(s) and assume that it is represented in the current population by some s ∈ H . ◮ If s participates in a crossover operation and the crossover point is located between bit positions 3 and 10, then with large probability the offspring are no longer in H ( H is 000** 001** 010** 011** 100** 101** 110** 111** Ω disrupted ). ◮ On the other hand, if the crossover point is elsewhere, then Then, in terms of expected values, one can show that one of the offspring stays in H ( H is retained ). E [ m ( H , σ ( p ))] = m ( H , p ) · f ( H , p ) . I.N. & P .O. Autumn 2006 I.N. & P .O. Autumn 2006 T–79.4201 Search Problems and Algorithms T–79.4201 Search Problems and Algorithms The effect of crossover on schemata (2/2) The Schema “Theorem” (1/2) ◮ Generally, the probability that in 1-point crossover a Heuristic estimate of the changes in representation of a given schema H = { 0 , 1 , ∗} n is retained, is (ignoring the possibility schema H from one generation to the next. Proposed by J. of “lucky combinations”) Holland (1975). Pr ( retain H ) ≈ 1 − ∆( H ) Denote: n − 1 , m ( H , t ) = number of individuals in population at generation t where ∆( H ) is the defining length of H , i.e. the distance that sample H . between the first and last fixed bit in H . ◮ More precisely, if H has m ( H , p ) representatives in Then: population p of total size m : (i) Effect of selection: � � Pr ( retain H ) ≥ 1 − ∆( H ) 1 − m ( H , p ) . m ( H , t ′ ) ≈ m ( H , t ) · f ( H ) n − 1 m I.N. & P .O. Autumn 2006 I.N. & P .O. Autumn 2006

Download Presentation
Download Policy: The content available on the website is offered to you 'AS IS' for your personal information and use only. It cannot be commercialized, licensed, or distributed on other websites without prior consent from the author. To download a presentation, simply click this link. If you encounter any difficulties during the download process, it's possible that the publisher has removed the file from their server.

Recommend


More recommend