Parallel implementation of genetic algorithm for searching optimal parameters of artificial neural networks

Genetic algorithm is a population-based evolutionary optimization approach, and this has been used to train artificial neural networks successfully. When a large number of individuals make up the population, however, the running time of the algorithm often becomes very long. Parallel computation is a technique that can potentially be used to address this issue. This research investigates the adoption of parallel genetic algorithm for searching optimal parameters of artificial neural networks (multilayer perceptrons). Paralellization is conducted using Message Passing Interface, where sub-populations (and their fitness values) are exchanged between processors while selection, crossover, and mutation processes, which are necessary to generate new sub-populations, are performed independently between processors. Experiments of this research show that parallelization has indeed reduced the running time of genetic algorithm. The proposed method produces an average running time of 2.066, 2.656, and 47.788 seconds/generation for the game of Tic-Tac-Toe, car evaluation, and game of chess (King-Rook vs. King-Pawn) datasets, respectively. For comparison, the average running time of the serial version of the algorithm is 14.397, 12.961, and 350.963 seconds/generation for the same three datasets, respectively.