An Analysis of Machine Learning Online Training Approaches for Simulation Optimization

Abstract

Simulation optimization refers to an optimization problem with a stochastic and potentially computationally expensive objective function. Machine learning (surrogate modeling) techniques have significant potential for enabling efficient simulation optimization, but the resulting surrogate model must have sufficient global accuracy to locate the region of the optimum, and sufficient local accuracy to pinpoint it. Historical surrogate modeling procedures have used a priori experimental designs, while more model literature has explored the use of training utility functions that select the next experimental point based on the current surrogate model state. Analysis of the performance of two training utility functions in comparison to random experimental selection is performed in the context of an online trained surrogate model. The results indicate that for a statically sized surrogate model, both utility functions provide equivalent performance that significantly exceeds random selection. An additional analysis quantifies the benefits to surrogate model accuracy of using gradient estimates in model training.