Multi-Source Bayesian Optimization Improves Constrained Design Space Exploration.

Optimizing systems under unknown constraints, particularly when the viable solution space is very small, poses a significant challenge for traditional Bayesian Optimization (BO) methods. These methods typically rely solely on evaluations of the true objective and constraints, making efficient exploration difficult. However, many real-world scenarios offer additional data sources, such as simplified simulations or surrogate models, which could aid early exploration. Researchers have developed a new general multi-source framework designed to integrate these auxiliary data sources into constrained BO. This framework extends the concept of constrained Max-value Entropy Search. It effectively models the correlations between different information sources while carefully balancing the cost of evaluating each source against the potential information gain. Experiments on both synthetic and physics-based benchmarks demonstrate that this new approach consistently and efficiently identifies optimal and feasible solutions, particularly excelling in the initial phases of exploration, even when auxiliary data sources are only weakly correlated with the true system.