Surrogate-Based Agents for Constrained Optimization
Diane C. Villanueva, Rodolphe Le Riche, Gauthier Picard, Raphael T. Haftka · 2012
Multi-agent systems have been used to solve complex problems by decomposing them into autonomous subtasks. Drawing inspiration from both multi-surrogate and multi-agent techniques, we define in this article optimization subtasks that employ different approxi-mations of the data in subregions through the choice of surrogate, which creates surrogate-based agents. We explore a method of design space partitioning that assigns agents to subregions of the design space, which drives the agents to locate optima through a mixture of optimization and exploration in the subregions. These methods are illustrated on two constrained optimization problems, one with uncertainty and another with small, discon-nected feasible regions. It is observed that using a system of surrogate-based optimization agents is more effective at locating the optimum compared to optimization with a single surrogate over the entire design space. Nomenclature c = centroid f = objective function F = objective function values associated with design of experiments in database g = constraint G = constraint values associated with design of experiments in database t = time x = design variables X = design of experiments in database I.