Adaptive Selection Routine for Evolutionary Algorithms

Abstract

This paper presents an adaptive selection scheme for use in evolutionary algorithms (EAs). The proposed algorithm adjusts the stochastic noise level in the determination of the mating pool in order to regulate the selection pressure. This eliminates the fitness scaling problem and allows optimization of the selection pressure throughout the learning phase, overcoming the major pitfalls of most popular EA selection procedures. Experimental evidence is given to prove the superior performance of the proposed technique compared with conventional EA procedures. The results also highlight how the application of windowing techniques to the roulette wheel procedure can increase the likelihood of premature convergence.