A Comparison of Cohort Genetic Algorithms with Canonical Serial and Island-Model Distributed GA’s Huafeng Pei Erik Goodman Genetic Algorithms Research and Applications Group (GARAGe) Case Center, Michigan State University, 2857 W. Jolly Road, Okemos, MI 48864 Email: goodman@egr.msu.edu Abstract This work considers the cohort genetic algorithm, a new type of genetic algorithm introduced by Holland. The cohort GA differs in several ways from the traditional canonical serial GA and island-model distributed GA. A key motivation for its development was to reduce “hitchhiking” – premature convergence of currently low-significance loci located near loci at which good building blocks are found early in the search process. This work compares one version of the cohort GA with canonical serial and island-model distributed GA’s on the basis of their abilities to reduce hitchhiking. The comparison is done using two types of test functions: the “royal road with potholes” function and hyperplane-defined functions (“HDF’s”). It is experimentally shown that even though theoretically the cohort GA can reduce hitchhiking, the particular version of the cohort GA tested is prone to another form of premature convergence, and it performed worse than the other GA’s. It is also shown that a small change in the placement of offspring among cohorts in the cohort GA may dramatically improve its performance. This suggests that further work on the cohort GA may well be fruitful. 1 INTRODUCTION The genetic algorithm (GA) is a family of search methods introduced by Holland [1975]. Much research has been done in order to understand how the GA works and how to improve its performance. The cohort GA is a new type of GA designed more recently by Holland [1998] [2000]. It is aimed at reducing the “hitchhiking” effect that occurs in the process of a GA’s search. Hitchhiking is a form of premature convergence that can hinder the GA or even make it unlikely for the GA to find a good solution for a given problem. Hitchhiking is most severe when the maximum reproduction rate is relatively high – for example, if the expected number of offspring of the best individual in the population is on the order of two or more. Hitchhiking is reduced when fitness is scaled so that the expected number of copies of the most fit individual produced in the next generation is 1.2 or fewer, but then, as Holland points out, other problems arise: 1) exploitation of the fitness difference is slowed, and 2) there is higher variance in the sampling of the fitness distribution (many times, individual with better-than-average fitness will be lost from the population). This higher variance occurs because GA’s typically use any fractional fitness excess above 1.0 as a probability of creating a second copy of an individual in the next generation. Thus the best individual’s gain becomes uncertain. The cohort GA is designed to allow a reduction in reproduction rates without introducing this stochastic sampling problem. Holland conjectured that the cohort GA’s mechanism, with relatively low maximum scaled fitnesses, will reduce hitchhiking, thus improving the performance of GA on classes of problems in which hierarchical assembly of building blocks is important to the solution trajectory. In this work, we tested the hypothesis that a cohort GA can reduce the hitchhiking effect and therefore improve the performance of GA’s by comparing it with a canonical serial GA and an island-model distributed GA on two types of seemingly appropriate test functions. We used a version of Holland’s cohort GA provided by Belding [Holland, 1998], a student of Holland. We note that Holland’s most recent publication on the cohort GA [Holland, 2000] includes some new mechanisms he has introduced to fight the convergence issues we (later) found in our work with his earlier version of the cohort GA; we have not yet experimented with his newer formulation. Section 1 introduces the hitchhiking effect and the cohort GA. Section 2 presents the experimental design. Results are given in Section 3, and conclusions in Section 4.