Online Surrogate-Based Multi-Objective Design Optimization using Generative Adversarial Networks with Constraint Assistance

Abstract

Multi-objective design optimization problems can be computationally expensive, such is the case with many engineering optimization problems, due to the original objective and/or constraint functions of the problem being costly to evaluate. A method established in current scientific literature to reduce the computational cost for such optimization problems involves the implementation of a surrogate or a lower-cost model to be used in the optimization process in place of the computationally expensive objective/constraint functions. The approach developed in this thesis uses an online surrogate-based optimization method in which the surrogate is developed and iteratively updated as the optimizer converges to a solution.

The primary contribution of this work is the proposal of a new approach for online surrogate-based multi-objective design optimization using generative adversarial networks. A constraint boundary-informed support vector machine facilitates the approach to predict whether the generated solutions are feasible or infeasible. The performance of the proposed approach is evaluated and compared to two other methods from the literature. The comparison of these methods is carried out using several quality metrics and using numerical and engineering test problems. The engineering test problem is based on the optimization of the operating conditions of an unmanned surface vessel. The results from these test problems indicate that the proposed approach is able to outperform the other approaches for most of the quality metrics and test problems.