Distributed Bayesian Optimization in Multi-Agent Systems

Sammanfattning: A variety of engineering problems require extremely resource consuming system performance optimization with an inaccessible system model, some examples include tuning the hyper-parameters in a complex machine learning model, simulation-based aerodynamic design and power system optimization. Bayesian optimization is an approach to solve black-box optimization problems when sample efficiency is of high priority. By parallelizing evaluations, the presence of multiple computing agents can be utilized to solve the optimization problem more efficiently. This thesis address the multi-agent black-box evaluation-expensive optimization problem by enabling distributed Bayesian optimization. Parallel methods proposed in previous research are extended to the distributed setting and a novel approach called Diversity Regularization is developed. Furthermore, motivated by applications in robotics systems such as source seeking, the evaluation-transition trade-off is addressed through regularization. Finally, empirical regret analysis to compare the presented methods on benchmark functions is performed.