220 likes | 370 Views
A Memory Enhanced Genetic Algorithm with Periodic Extinction and Dynamic Population Size. Arash Emami and Wei Chen. Motivation. Genetic Algorithms parallel biological genetics Mutation, selection, crossover – biological equivalents External factors also influence biology
E N D
A Memory Enhanced Genetic Algorithm with Periodic Extinction and Dynamic Population Size Arash Emami and Wei Chen
Motivation • Genetic Algorithms parallel biological genetics • Mutation, selection, crossover – biological equivalents • External factors also influence biology • Environment steers evolution
Punctuated Equilibrium • Evolution is not constant • Periods of evolution followed by periods of stasis • Natural disasters, disease, climate change • Periods of stasis = population convergence
Cataclysmic event Change Punctuated Equilibrium
Punctuated Equilibrium • Can we improve a classical GA by adding PE to it?
Cataclysmic Events • Natural disasters, etc. • Spurs the population to a new direction • Evolution proceeds in a different direction • In a GA, can get the population past local minima/maxima
GA Features • Selection • Tournament • Elite • One-point Crossover • Mutation
Three Main Features • Periodic Extinction • Population Memory • Dynamic Population Size
Population Memory • The best strings are stored in the Population Memory • Add the best individual of each generation for the first M generations • For generations M+1, compare best member of population against worst in memory
Periodic Extinction • Every N generations, the population is wiped • Population re-started • Strings in Population Memory • Random strings
Dynamic Population Size • Population size varies during the course of the GA run • Becomes smaller after an Extinction event • Grows in size as new random strings are added • Eventually reaches its original size
Three Phases • Normal Phase • Static Population Size • Extinction Phase • Population = 0 • Recovery Phase • Dynamic Population Size
The Algorithm • Implemented in Java Initialize population; Calculate individual fitness; //based on fitness function While (evolving) { Checkpoint; //pick up best individual & do population controlling If (extinction condition) { Do mass extinction; } else { Select parent individuals; //elite or tournament 1-point crossover; Mutation; //per bit } Calculate individual fitness; //based on fitness function }
Experiment • Test Task: 6 numerical functions and 1 “real” problem • 5 Well known DeJong functions • Dynamic problem • 3-bit deceptive problem (Moving peaks) • Test Configuration: • Extinction period: 50 • Memory Size: 10 or 100 • Initialize Population Size: 10 • Mutation Rate for all runs: 0.05
Experiment (cont.) DeJong Functions: F1: F2: F3: 3-bit Deceptive Problem: F4: F5: Dynamic Problem:
Population Size (Elite) 2000 Size 1500 1000 500 0 1 19 37 55 73 91 109 127 145 163 181 199 217 235 253 271 289 Generation Results (3-bit Deceptive Problem)
Future Work • Run different tests • De Jong functions inefficient • Add all random strings at once • Change mutation rate after extinction
Conclusion • We presented an algorithm that combined Mass Extinction, Dynamic Population Size, and Population Memory • Need to do more tests • De Jong functions were no good, too easy • Results of the 3-bit deceptive problem were promising
References • J. P. Cohoon, S. U. Hegde, W. N. Martin, and D. Richards, “Punctuated Equilibria: A Parallel Genetic Algorithm”, Proceedings of the Second International Conference on Genetic Algorithms, pages 148-154, • T. Krink, R. Thomson, “Self Organized Criticality and Mass Extinction in Evolutionary Algorithms”, Proceedings of the 2001 Congress on Evolutionary Computation CEC2001 • J. Costa, R. Tavares, A. Rosa, “An Experimental Study on Dynamic Random Variation of Population Size” • J. Branke, “Memory Enhanced Evolutionary Algorithms for Changing Optimization Problems” • DeJong, K. 1975. “An analysis of the behaviour of a class of genetic adaptive systems”. Ph.D. diss., Dept. of Computer Science, University of Michigan • Raich, A. M., and Ghaboussi, J., 1997. “Implicit Representation in Genetic Algorithms Using Redundancy”. Evolutionary Computation 5(3): 277-302. • http://www.aifb.uni-karlsruhe.de/~jbr/MovPeaks/
Thank you. Any Questions?