A Data-Centric Machine Learning Approach for Controlling Exploration in Estimation of Distribution Algorithms

Exploration plays a key role in the performance of metaheuristics. An algorithm should perform more exploration when reaching the "ideal search scale"; this happens when solutions are regularly sampled from different attraction basins. The moment this search scale is reached depends on the topological features of the objective function and the inherent randomness of the heuristic optimization process. Previous work on adjusting exploration have mostly used fixed rules based on fitness improvement, in this paper, we model it as a supervised machine learning problem. We apply a data-centric approach to understand whether variations in the data are more relevant than variations in the classification models. For our study we use the Estimation Multivariate Normal Algorithm with Thresheld Convergence, which provides an ideal framework as it allows us to directly control exploration through the. parameter. Optimization results show that the machine learning hybrid significantly outperforms the baseline algorithm.