230 likes | 249 Views
A comparative analysis of selection schemes used in genetic algorithms. David E. Goldberg Kalyanmoy Deb. What is the paper about?. Defines and compare four selection schemes Presents a technique for comparisons: Produce a difference/differential equation modeling the selection scheme
E N D
A comparative analysis of selection schemes used in genetic algorithms Genetic Algorithms David E. Goldberg Kalyanmoy Deb
What is the paper about? • Defines and compare four selection schemes • Presents a technique for comparisons: • Produce a difference/differential equation modeling the selection scheme • Test computer implementation against diff. equation model • Defines criteria for comparison: • Convergence time • Schema growth ratios • Conclusions: practical applications of analysis Genetic Algorithms
Where are we now? • Many papers claim the superiority of this or that selection scheme • But many of these claims are based on limited (and uncontrolled experiments). • Little analysis has been done • This paper attempts to provide the needed analysis Genetic Algorithms
What selection strategies? • Proportionate reproduction • Ranking selection • Tournament selection • Genitor (“steady state”) selection Genetic Algorithms
Birth, life, and death • m(i, t+1) = m(i, t) + m(i, t, b) – m(i,t,d) • Ex: in non-overlapping population models: • m(i,t+1) = m(i,t,b) ; m(i,t,d) = m(i,t) • We can also do proportions: • P(i,t+1) = P(i,t) + P(i,t,b) – P(i,t,d) Genetic Algorithms
Proportionate Reproduction • Probability of selection: • prob(i,t) = f(i)/∑m(j,t) f(j) • Various methods for implementation: • Roulette wheel • Stochastic remainder • Stochastic universal Genetic Algorithms
How many in next generation? • m(i,t+1) = m(i,t) * n * prob(i,t) • m(i,t+1) = m(i,t) * f(i)/f(avg,t) • Proportion(i,t+1) = Proportion(i,t) * f(i)/f(avg,t) Genetic Algorithms
Graph of Eqn, implementation Genetic Algorithms Convergence behavior
How many individuals between specified values of x in objective function f(x)? Let p0(x) be uniform, integral 1 Consider f(x) = xc and f(x) = ecx Limits x and x – 1/n Takeover time Genetic Algorithms
Behavior of f(x) = x^c Integrate with limits x & x – 1/n x = 1 is best, x = 0 is worst individual Compare theory and experiment for f(x) = x Genetic Algorithms
Takeover time for f(x) = x^c Solving for t and approximating Setting x = 1, we get proportion of best individual Setting P = n-1/n, we calculate when population contains n-1 best individuals Thus the takeover time for a polynomially distributed objective function is O(nlogn) Genetic Algorithms
Takeover time for f(x) = e^cx The takeover time for a polynomially and exponentially distributed objective function is O(nlogn) Genetic Algorithms
Time complexity of Proportionate Reproduction • Roulette Wheel • O(n2) or O(nlogn) with binary search • Stochastic remainder selection • floor(f(i)/favg) number of copies • Remainder = flip(fractional(f(i)/favg)) • O(n) without replacement or O(n2) with Genetic Algorithms
Ranking • Sort from best to worst • Create a transformation function called an assignment function that converts a rank to an equivalent “fitness” • assignFunction(rank) • Proportionate reproduction on assignFunction(rank) Genetic Algorithms
Tournament Selection • Binary • N-ary • Randomly choose N individuals from population • Select best for further processing Genetic Algorithms
Binary Tournament • Tournament size = 3 Genetic Algorithms
Tournaments Genetic Algorithms
Genitor • Choose an offspring based on ranking • Replace worst individual in population Genetic Algorithms
Genitor Genetic Algorithms
Growth Comparison Genetic Algorithms
Takeover time comparison Genetic Algorithms
Time complexity Genetic Algorithms
Conclusions • The paper provides a framework for comparing selection operators • Compares selection “pressure” for each type of selection • Introduces the concept of takeover time to help us understand the exploration/exploitation tradeoff • Provides takeover time estimates for different types of selection • Implications for genetic search • The models provide us with theory necessary to compare selection methods and • Balance growth ratio (quick convergence) with higher crossover/mutation (more exploration) Genetic Algorithms