Dynamic Partitioning for Balancing Exploitation and Exploration in Constrained Optimization: A Multi-Agent Approach

Abstract

In this article, we define optimization subtasks that employ different approximations of the data in subregions through the choice of surrogate, which creates surrogate-based agents. Through design space partitioning, which assigns agents to subregions of the design space, the agents solve the optimization problem in their respective subregions, and use the feasibility and objective function value to assess the value of the solutions in order to center the subregions at local and global optima. Further, we introduce methods to create agents at run-time which allows additional exploration by creating a finer partition of the design space. The end result of this dynamic partitioning is a multi-agent system that inherently balances exploitation and exploration in the design space. We illustrate this approach on a constrained optimization problem with small, disconnected feasible regions. It was observed that the agents were effective at locating global and local optima.