Improvement of Search Performance in Genetic Algorithms with Fitness Prediction

When Genetic Algorithm (GA) is applied to actual engineering problems, for example applied to optimization of design parameters, the searching time is usually huge because the fitness is calculated by repetitive simulation or analysis, which needs a large amount of calculation time. In order to shorten the searching time, a method called fitness prediction GA has been proposed. It reduces the calculation time by predicting fitness instead of actually calculating it, and eventually shortens the searching time. In this paper, we propose a new method, Dual Population GA (DPGA), which employs both real and virtual populations. Real populations have actual fitness value, and virtual ones have predicted one. DPGA can prevent the decline of performance caused by prediction errors, which may happen in fitness prediction GA, by appropriately migrating virtual populations into real ones and accelerating evolution. A fitness predictor based on a neural network is also proposed in this paper. Through computer simulations, DPGA is proved to be able to improve the searching performance of fitness prediction GA.