Co-evolution of Fitness Maximizers and Fitness Predictors

We introduce an estimation of distribution algorithm (EDA) based on co-evolution of fitness maximizers and fitness predictors for improving the performance of evolutionary search when evaluations are prohibitively expensive. Fitness predictors are lightweight objects which, given an evolving individual, heuristically approximate the true fitness. The predictors are trained by their ability to correctly differentiate between good and bad solutions using reduced computation. We apply co-evolving fitness prediction to symbolic regression and measure its impact. Our results show that the additional computational investment in training the co-evolving fitness predictors can greatly enhance both speed and convergence of individual solutions while overall reducing the number of evaluations. In application to symbolic regression, the advantage of using fitness predictors grows as the complexity of models increases.