Replication of the Niche Radius Problem using Clustering Genetic Algorithms [Abstract]

Abstract

Niche Genetic Algorithms (NGA) are a specialized type of Genetic Algorithm (GA) that attempts to locate multiple optima. Many NGAs use a radius parameter. The success of the algorithm is dependent upon the selection of a “good” radius, which is normally half of the distance between optima. Since the purpose of a GA is to locate the optima, this is normally not known in advance. If the optima is known, it negates the need for running the GA. If the radius is set incorrectly, not all of the optima are located. This problem is known as the Niche Radius Problem (NRP). This research replicates the NRP using a simple Clustering NGA. It compares a traditional GA to a Clustering NGA with the radius set too small, too large and correctly. Twenty trials of each were created for each of the four cases. All other parameter values, with the exception of the radius, were consistent throughout the trials. Statistical tests were performed on the results. This research concludes that traditional GAs can locate only one optimum in multiple optima problems. Setting the radius too small correctly identifies all of the optima, but decreases the average fitness of the generations. Setting the radius too large will locate one of the optimum, but not all of them. Finally, setting the radius to the correct distance locates all of the optima with high precision.