Cost Performance of Using Surrogate Models in GA Optimization for Machine Designs

Large scale multi-objective optimizations are becoming indispensable tools for machine designs. They require a considerable amount of computation and that is becoming an obstacle. Machine learning based surrogate models are expected to reduce the computation amount. On the other hand, the optimization performance is affected by the surrogate model's accuracy, which is determined by the training. The training requires a certain computation. Therefore, it is important to understand the relationships between optimization performance, the accuracy of surrogate models, and the cost. This paper explains the relationships by comparing two different training approaches, “offline” and “online”. The results show that the online training has preferable characteristics which improve the cost performance of the optimization.