Self-organized Space Partitioning for Multi-Agent Optimization

Abstract. In this paper we explore the use of multi-agent systems to tackle opti-mization problems in which each point is expensive to get and there are multiple local optima. The proposed strategy dynamically partitions the search space be-tween several agents that use different surrogates to approximate their subregion landscape. Agents coordinate by exchanging points to compute their surrogate and by modifying the boundaries of their subregions. Through a self-organized process of creation and deletion, agents adapt the partition as to exploit potential local optima and explore unknown regions. The overarching goal of this tech-nique is to all local optima rather than just the global one. The rationale behind this is to assign adequate surrogate to each subregion so that (i) optimization is cheaper, (ii) the overall optimization process is not only global in scope but also stabilizes on local optima and (iii) the final partitioning provides a better under-standing of the optimization problem.