A comparison of fitness scallng methods in evolutionary algorithms

Abstract

Proportional selection (PS), as a selection mechanism for mating (reproduction with emphasis), selects individuals according to their fitness. Consequently the probability of an individual to obtain a number of offspring is directly proportional to its fitness value. This can lead to a loss of selective pressure in the fmal stages of the evolutionary process degrading the search. This presentation discusses performance results on evolutionary algorithms optimizing two highly multimodal (Michalewicz's and Griewank's) functions and a hard unimodal (Easom' s) function. Experiments were addressed to contrast the behaviour of a simple genetic algorithm against three scaling methods: linear, sigma truncation and recency-weighted- running-average. Diverse measures of performance were used to establish quality of results and convergence speed.