Constraint-Bound Agnostic Bayesian Optimization: One Model for All Thresholds

The paper proposes Constraint-Bound Agnostic Bayesian Optimization (CBA-BO), a framework that learns a parametric mapping from constraint thresholds to optimized solutions. Instead of solving a separate constrained Bayesian optimization problem for every threshold configuration, CBA-BO predicts solutions for previously unseen threshold queries and applies a one-step Bayesian optimization refinement. The authors report experiments on benchmark and engineering problems, plus an intent-guided mechanism for recommending constraint bounds according to user preferences over objective performance and feasibility.