Epsilon Optimal Sampling.

Epsilon Optimal Sampling (EOS) is a novel algorithm that seeks to reduce the computational complexity of selecting the best design using stochastic simulation. EOS is an Optimal Computing Budget Allocation (OCBA) type algorithm that reduces computational complexity by integrating machine learning (ML) models into the simulation optimization algorithm. EOS avoids the pitfall of trading computational overhead in simulation execution for computational overhead in ML model training by using a concept we call policy stability. In this paper, we present the concept of policy stability, how it can be used to improve dynamic sampling techniques, and how low-fidelity ML estimates can be integrated into the process. Numerical results are presented to provide evidence as to the improvement in computational efficiency that can be achieved when using EOS in conjunction with ML models over the standard OCBA algorithm.