MULTI-AGENT BAYESIAN OPTIMIZATION FOR UNKNOWN DESIGN SPACE EXPLORATION

This paper proposes a multi-agent Bayesian optimization (MABO) framework as a reference model for rational design teams to study the effects of information exchange on a team's search performance in finding global optimum of complex objective functions with many local optima. The core idea of the framework has three main steps. First, the design space is divided into regions based on the number of agents involved in the search. In each region, only one agent works on the part of the objective function. Second, a global-local communication strategy is developed to allow agents in local searches to share their sampled design points with a global evaluator. The global evaluator computes the posterior mean and variance based on all sampled points from local agents and evaluates the acquisition function (e.g., the expected improvement) to recommend the next sampling decisions for local agents. Third, when making the decision about where to sample next, each local agent only has access to the expected improvement evaluated in its local region and chooses the design that yields the largest value locally. To evaluate how the information exchange between agents and between local and global impact the search results, our framework is compared with a multi-agent model that does not allow information sharing and global-local interaction. Furthermore, we evaluated the performance of the model based on benchmark functions with varying complexities and also investigated the impact of the number of agents on search performance. We observe that when information sharing is allowed and global-local interaction is enabled in all scenarios, there is a significant improvement in convergence speed as well as the success rate of convergence.