Efficient Global Optimization for Dynamic Problems

Efficient Global Optimization (EGO) is a very important black-box optimization framework for solving expensive optimization problems, which appear in high-fidelity simulation-based optimization. EGO operates by employing a Gaussian process model as an approximation to the computationally expensive black-box function. As we observe an increasing trend towards dynamic optimization problems that incorporate live data, combined with the fact that EGO is designed to solve static optimization problems, a need for adaptation arises. This paper analyzes and compares five EGO extensions that should aid the algorithm in dealing with dynamic changes. Results indicate that such dynamification is challenging and while successful for problem instances with the correct difficulty and change severity, it can have adverse effects if older data is unrepresentative of new scenarios.