Regression versus Classification for Predicting Feasibility in Offline Data- Driven Evolutionary Constrained Optimization

There are some real-world optimization problems that only historical experimental data are available for optimization. Surrogate models are built based on these offline data for the objective function and constraints to evaluate the quality and feasibility of solutions during the optimization process, which is called offline data-driven optimization. For solving offline data-driven constrained optimization problems, building reliable surrogate models for constraints and designing suitable constraint-handling technique are the key issues. In this paper, we summarize a general framework of offline DDEAs using regression models or classification models as surrogate models for constraints, and conduct comparative experiments on these algorithms. The experiment results on the test problems show that the algorithms using regression models can obtain the solutions of better objective value than that of classification models, while that of using classification models can obtain feasible solutions on most of problems but the solutions have poor quality.