Noah's ark strategy for avoidance of excess convergence by a parallel genetic algorithm with an object-shared space

Abstract

In a genetic algorithm (GA), the undesirable phenomenon of excess convergence can often occur. Excess convergence is the phenomenon where the diversity of a group is lost. This phenomenon occurs because homogeneous individuals are increased rapidly in the group while evolving or searching. Therefore, crossover loses its function. Once the excess convergence occurs, the search by the GA becomes meaningless. Therefore, it is important to avoid excess convergence and maintain diversity. First, we show an implementation of a parallel GA based on a multiple-group-type island model, that uses object-shared space. Next, as a simple, effective method for avoiding excess convergence, we propose a diversity maintenance technique based on selection of the homogeneous individuals called the Noah's ark strategy for parallel GAs, and demonstrate its effectiveness on a knapsack problem. Our proposed method is to replace individuals in sub-groups that have excessively converged with the new individuals coming from the search space. That is, we avoid excess convergence by expelling homogeneous individuals, with the exception of one "elite" individual (that we call for Noah). Thus, we limit a decrease in diversity of an entire group.