Multi Agent Rollout for Bayesian Optimization

Solving black-box global optimization problems efficiently across domains remains challenging especially for large scale optimization problems. Bayesian optimization has obtained important success as a black box optimization technique based on surrogates, but it still suffers when applied to large scale heterogeneous landscapes. Recent approaches have proposed non-myopic approximations and partitioning of the input domain into subregions to prioritize regions that capture important areas of the solution space. We propose a Multi Agent Rollout formulation of Bayesian optimization (MAroBO) that partitions the input domain among finite set of agents for distributed sampling. In addition to selecting candidate samples from their respective sub regions, these agents also influence each other in partitioning the sub regions. Consequently, a portion of the function is optimized by these agents which prioritize candidate samples that don't undermine exploration in favor of a single step greedy exploitation. The efficacy of MAroBO is demonstrated on synthetic test functions.