Offline Data-Driven Mixed-Variable Optimization Algorithm Using a Step-Wise Strategy

Yiteng Xu, Handing Wang

Abstract: Some real-world engineering problems are offline data-driven mixed-variable optimization problems, which involve optimizing both continuous and discrete variables using only historical experimental data. The main challenges are handling mixed variables and utilizing surrogate models effectively. We propose a novel algorithm that uses a step-wise strategy to optimize the discrete and continuous variables in two stages. In the first stage, we use different radial basis function networks models as surrogates and a voting method to select a promising subspace of discrete variable values. In the second stage, we fix the discrete variable values and use a selective ensemble strategy to optimize the continuous variables. We test our algorithm on 30 test problems and compare it with two representative algorithms. The results show that our algorithm is superior and more stable on most problems, especially on complex multimodal problems. Our algorithm is an effective and flexible framework for handling mixed variables and improving search efficiency and quality.