A population-based cross-entropy method with dynamic sample allocation

This paper generalizes the cross-entropy (CE) method to a population-based setting, where a population of probabilistic models is maintained/updated and subsequently propagated from generation to generation. One of the key questions in the proposed approach is how to efficiently distribute a given sample budget among different models in a population to maximize algorithm performance. We formulate this problem as a Markov decision process (MDP) model and derive an efficient dynamic sample allocation scheme to adaptively allocate computational resources. We carry out numerical studies to illustrate the method and compare its performance with existing procedures.