Hybrid Evolutionary Algorithm with an Optimal Sample Allocation Strategy for Multifidelity Simulation Optimization Problems

Simulation has been applied to evaluate system performance even when the target system does not exist in practice. Dealing with model fidelity is required to apply simulation to practice. A high-fidelity (HF) simulation model is generally more accurate and requires more computational resources than a low-fidelity (LF) one. A low-fidelity model may have less accuracy than a HF one, but it can rapidly evaluate a design alternative. Consequently, the performance accuracy of the constructed simulation model and its computational cost involves a tradeoff. In this research, the simulation optimization problem under a large design space, where a LF model may not be able to evaluate all design alternatives in the limited computational resource, is studied. We extended multifidelity (MF) optimization with ordinal transformation and optimal sampling (MO2TOS), which enables the use of LF models to search for a HF one efficiently, and proposed a combination of the genetic algorithm and MO2TOS. A novel optimal sample allocation strategy called MO2TOSAS was proposed to improve search efficiency. We applied the proposed methods to two experiments on MF function optimization and a simultaneous scheduling problem of machine and vehicles (SSPMV) in flexible manufacturing systems. In SSPMV, we developed three fidelity simulation models that capture important characteristics, including the preventive deadlock situation of vehicles and alternative machines. Simulation results show that the combination of more than one fidelity level of simulation models can improve search efficiency and reduce computational costs.