Offline Data-Driven Optimization at Scale: A Cooperative Coevolutionary Approach

Data-driven evolutionary algorithms (DDEAs) have received increasing attention during the past decade, but most existing studies are dedicated to solving relatively small-scale problems. For large-scale optimization problems (LSOPs), special efforts must be made to both the surrogate and the evolutionary components to overcome the "curse of dimensionality," which remains a challenge in this research area. To address this research limitation, we propose a novel cooperative coevolution-based DDEA (CC-DDEA). First, a hierarchical surrogate-joint learning model is designed to provide fitness approximations at both global and subdivided spaces, thus being able to guide the evolutionary population searching at different granularities. Then, optimization is conducted on both the global level and local subspace level in the manner of cooperative coevolution. In the local-level search, we introduce a gradient-based operator to accelerate the convergence efficiency of subspaces, owing to the differentiable property of our surrogate model. Additionally, the entire framework is used in conjunction with a progressive and dynamic space division strategy, enabling local parallel-to-global unified search and facilitating the final convergence. Experiments on up to 1000-D problems and the comparisons with state-of-the-art DDEAs validate the powerfulness of the proposed algorithm.