An RBF Network Based Crossover for Pseudo-Boolean Optimization

Recombination is an important operator in population metaheuristics. Recently, there has been interest in developing efficient recombination operators that explore characteristics of the optimization problem. We propose using a radial basis function network (RBFN) to generate recombination masks based on information contained in the parents and on past successful recombinations. The RBFN is trained online and the training set is built by adding examples of successful recombinations obtained during the optimization process. The time complexity for training the RBFN is O(N
²
), while the time complexity for applying the proposed crossover operator is O(N), the same time complexity of traditional crossover operators. Experiments with NK landscapes indicated that a genetic algorithm (GA) with the proposed crossover was capable of generating better results for successful recombination and improvement rates, when compared to a GA with traditional crossover. However, this did not imply in better results for the best fitness.