Efficient Simulation Optimization with Simulation Learning

Abstract

Simulation optimization has found great success in automation science and engineering, such as the optimization of manufacturing systems, thanks to its capability to fully account for the complexity and uncertainty in systems. However, it remains a challenge to use simulation optimization in applications where decision time window is very short because of computational efficiency challenge. In this paper, a new framework known as Sequential Allocation using Machine-learning Predictions as Light-weight Estimates (SAMPLE) is proposed to address this challenge. SAMPLE utilizes an offline simulation learning phase to train machine learning models using simulation data. When a decision needs to be made, SAMPLE utilizes machine learning predictions under a Bayesian framework to determine optimal allocation of simulation sampling budget. The proposed approach enables fast-time simulation-based decision making for automation systems. SAMPLE is able to work with lightweight machine learning models that may only provide crude approximations but still achieve considerable computational efficiency gain. Numerical experiments with both benchmark test functions and a case study demonstrate the viability of the proposed SAMPLE framework, with significant performance improvement over decision making using only machine learning predictions, or simulations alone.