Coactive Preference-Guided Multi-Objective Bayesian Optimization: An Application to Policy Learning in Personalized Plasma Medicine

The design of advanced learning- and optimization-based controllers requires selecting parameters that balance performance objectives and constraints. Bayesian optimization (BO) has proven effective for resource-efficient calibration of such controllers. Preference-guided BO incorporates user preferences to prioritize areas of interest, but it lacks a mechanism for users to specify desired outcomes directly. This letter introduces a user-centric framework for preference-guided BO, leveraging a novel knowledge-gradient based coactive acquisition function that allows users not only to select preferred outcomes but also also propose alternatives to guide exploration. To enable efficient implementation, we approximate the acquisition function, avoiding costly bilevel optimization. The approach is validated for control policy adaptation in personalized plasma medicine, where it outperforms standard preference-guided BO by effectively integrating user feedback to personalize treatment protocol.