Abstract
The purpose of this project is to implement, and analyze the performance of, a variation of a Terrain-Based Memetic Algorithm for training a Self-Splitting Neural Network. The algorithm implementation is incorporated into the source code of a Visualization Toolkit to allow for more detailed analysis of the training process and resultant modular network. The primary concern is network generalization, evaluated by the number of needed networks in the solution, as well as the time and computational effort needed to discover a solution. The algorithm is tested by training a Self-Splitting Neural Network to solve the Two-Spiral Problem. For comparison the Terrain-Based Memetic Algorithm is tested alongside a Terrain-Based Genetic Algorithm and a standard Genetic Algorithm. Analysis of averaged testing results shows the Terrain-Based Memetic Algorithm to perform equal to or better than the Terrain-Based Genetic Algorithm and the standard Genetic Algorithm, with regard to number of total networks needed in the solution as well as time to train. While the Terrain-Based Memetic Algorithm performed significantly better than a standard Genetic Algorithm its performance improvements over the Terrain-Based Genetic Algorithm were marginal, especially when the algorithms were given significantly long times to train. Network topology played a large role in training success; however, the Terrain-Based Memetic Algorithm maintained its relative performance advantage over the comparison algorithms with all topologies tested.