Federated Offline Contextual Bayesian Optimization
Published in arXiv preprint, 2026
Discovering optimal designs through sequential data collection is essential in many real-world applications. While Bayesian optimization has achieved strong performance in this setting, context-specific optimal design introduces additional challenges because the learner must estimate a mapping from contexts to optimal designs.
This work proposes CCBO, Collaborative Contextual Bayesian Optimization, a framework for multiple clients to jointly perform contextual Bayesian optimization with controllable contexts. The framework supports online collaboration, offline initialization from peer beliefs, and optional privacy-preserving communication. We establish sublinear regret guarantees and demonstrate improved performance through simulations and a real-world hot rolling application.
Recommended citation: C.-Y. Chang, Q. Chen, T. Gao, C. Okwudire, D. Fenning, N. Dasgupta, W. Lu and R. A. Kontar. "Federated Offline Contextual Bayesian Optimization." arXiv preprint, 2026.
Download Paper