Rigorous Analyses of Simple Diversity Mechanisms Tobias Friedrich Nils Hebbinghaus Frank Neumann Max-Planck-Institut für Informatik Saarbrücken
Tobias Friedrich Diversity Important issue for designing successful EAs Prevents an EA from too large selection pressure Assumption: The right diversity mechanism may be crucial for the success of an algorithm Aim of this talk: Show this observed behavior by rigorous runtime analyses Frank Neumann
Tobias Friedrich Runtime Analysis Lot of progress in recent years Results for pseudo-Boolean functions Well-known combinatorial optimization problems Most results are for the (1+1) EA Some examine the choice of the “ right ” population size No analyses that consider the impact of diversity Question in this talk: What about diversity in populations? Frank Neumann
Tobias Friedrich Simple Diversity Mechanisms Diversify the population with respect to search points Diversify the population with respect to fitness values Show situations where the behavior of these strategies differs significantly Frank Neumann
Tobias Friedrich Search point diversifying (μ+1)-EA initial population (fitness = size): select random individual mutate this if already in population, goto 1. add new individual delete individual with lowest fitness current population (fitness = size): Frank Neumann
Tobias Friedrich Fitness diversifying (μ+1)-EA initial population (fitness = size): select random individual mutate this if individual with same fitness in population, replace this by new individual and goto 1 current population (fitness = size): Frank Neumann
Tobias Friedrich Fitness diversifying (μ+1)-EA select random individual mutate this if individual with same fitness in population, replace this by new individual and goto 1 add new individual delete individual with lowest fitness current population (fitness = size): Frank Neumann
Tobias Friedrich Plateaus Examine the choice of diversity on plateau functions Plateaus are regions in the search space where all search points have the same fitness Size and structure determines difficulty for evolutionary search Investigations for the (1+1) EA on pseudo-Boolean functions, maximum matchings, Eulerian cycles Frank Neumann
Tobias Friedrich Investigations Search point vs. Fitness diversifying ( μ +1)-EA Constant population size Search space {0,1} n, mutate each bit with 1/n Compare them on different plateau functions Runtime:= Number of fitness evaluations to reach an optimal search point Show advantage/disadvantage of the different diversity mechanisms Frank Neumann
Tobias Friedrich Theorem 1 Th eorem 1 : O n f ( x ) : = 8 < : j x j 0 : x 6 2 f 1 i 0 n ¡ i ; 0 < i · n g n + 1 : x 2 f 1 i 0 n ¡ i ; 0 < i < n g n + 2 : x = 1 n searc h po i n t d i vers i f y i ng ( ¹ + 1 ) - EAh as expec t e d run t i me O ( n 3 ), ¯ t ness d i vers i f y i ng ( ¹ + 1 ) - EAh as exponen t i a l run t i mew i t h overw h e l m i ngpro b a b i l i t y.
Tobias Friedrich Formal Proof of Theorem 1 Tonto Plateau (Grand Canyon)© by Prof. Ian Parker, Univ. of California f ( x ) : = 8 < : j x j 0 : x 6 2 f 1 i 0 n ¡ i ; 0 < i · n g n + 1 : x 2 f 1 i 0 n ¡ i ; 0 < i < n g n + 2 : x = 1 n Pl a t eau f unc t i on i n t h eory: Pl a t eau i n t h erea l wor ld :
Tobias Friedrich Proof of Theorem 1 Tonto Plateau (Grand Canyon)© by Prof. Ian Parker, Univ. of California 0 n 1 n o t h erw i se f 1 i 0 n ¡ i ; 0 < i < n g Pl a t eauw i t h¯ t ness n + 1 Pl a t eau f unc t i on f ( x ) : = 8 < : j x j 0 : x 6 2 f 1 i 0 n ¡ i ; 0 < i · n g n + 1 : x 2 f 1 i 0 n ¡ i ; 0 < i < n g n + 2 : x = 1 n O p t i mumw i t h ¯ t ness n + 2
Tobias Friedrich Proof of Theorem 1 0 n 1 n o t h erw i se f 1 i 0 n ¡ i ; 0 < i < n g Fitness diversifying ( μ +1)-EA
Tobias Friedrich Proof of Theorem 1 0 n 1 n o t h erw i se f 1 i 0 n ¡ i ; 0 < i < n g Mutation with probability Selection kills individual on plateau Mutation Selection ? ? 1 n
Tobias Friedrich Proof of Theorem 1 0 n 1 n o t h erw i se f 1 i 0 n ¡ i ; 0 < i < n g Individual on plateau cannot perform random walk exponential runtime with overwhelming probability
Tobias Friedrich Proof of Theorem 1 0 n 1 n o t h erw i se f 1 i 0 n ¡ i ; 0 < i < n g Search point diversifying ( μ +1)-EA expected polynomial runtime Optimum found!
Tobias Friedrich Theorem 2 searc h po i n t d i vers i f y i ng ( ¹ + 1 ) - EAh as exponen t i a l run t i mew i t h pro b a b i l i t y 1 = 2 ¡ o ( 1 ), ¯ t ness d i vers i f y i ng ( ¹ + 1 ) - EAh as expec t e d run t i me O ( n 3 ).
Tobias Friedrich Proof of Theorem 2 D ou bl e-p l a t eau f unc t i on:
Tobias Friedrich Proof of Theorem 2 f ( x ) : = 8 > > < > > : n + 1 : x 2 Pl a t eau 1 ( w i t h ou t O p t i mum ) n + 2 : x 2 Pl a t eau 2 n + 3 : x = O p t i mum ( on Pl a t eau 1 ) j x j 0 : o t h erw i se. D ou bl e-p l a t eau f unc t i on: D ou bl e-p l a t eau i n t h erea l wor ld :
Tobias Friedrich Proof of Theorem 2 f ( x ) : = 8 > > < > > : n + 1 : x 2 Pl a t eau 1 ( w i t h ou t O p t i mum ) n + 2 : x 2 Pl a t eau 2 n + 3 : x = O p t i mum ( on Pl a t eau 1 ) j x j 0 : o t h erw i se. D ou bl e-p l a t eau f unc t i on: D ou bl e-p l a t eau \ c l ose t o t h erea l wor ld" :
Tobias Friedrich Proof of Theorem 2 Plateau 2 Plateau 1 Optimum f ( x ) : = 8 > > < > > : n + 1 : x 2 Pl a t eau 1 ( w i t h ou t O p t i mum ) n + 2 : x 2 Pl a t eau 2 n + 3 : x = O p t i mum ( on Pl a t eau 1 ) j x j 0 : o t h erw i se. D ou bl e-p l a t eau f unc t i on:
Tobias Friedrich Proof of Theorem 2 Mutation 1 2 Search point diversifying ( μ +1)-EA (only avoiding duplicates) 1 2
Tobias Friedrich Proof of Theorem Optimum found! Search point diversifying ( μ +1)-EA reaches Optimum with prob. 1 2
Tobias Friedrich Proof of Theorem Search point diversifying ( μ +1)-EA expected exponential runtime
Tobias Friedrich Proof of Theorem 2 Mutation Fitness diversifying ( μ +1)-EA expected polynomial runtime Optimum found!
Tobias Friedrich Larger Populations
Tobias Friedrich Conclusions Ensuring diversity is important for successful EAs First rigorous runtime analysis on this topic Using the “ right ” strategy may have a great impact on the runtime Proven for some basic plateau functions Same effect can be observed in multi-objective optimization (upcoming CEC paper) Future work: Other measures for diversity, classical combinatorial optimization problems Thanks!