Evolutionary optimization with active learning of surrogate models and fixed evaluation batch

Evolutionary optimization is often applied to problems, where simulations or experiments used as the fitness function are expensive to run. In such cases, surrogate models are used to reduce the number of fitness evaluations. Some of the problems also require a fixed size batch of solutions to be evaluated at a time. Traditional methods of selecting individuals for true evaluation to improve the surrogate model either require individual points to be evaluated, or couple the batch size with the EA generation size. We propose a queue based method for individual selection based on active learning of a kriging model. Individuals are selected using the confidence intervals predicted by the model, added to a queue and evaluated once the queue length reaches the batch size. The method was tested on several standard benchmark problems. Results show that the proposed algorithm is able to achieve a solution using significantly less evaluations of the true fitness function. The effect of the batch size as well as other parameters is discussed.