An Analysis of Safety Guarantees in Multi-Task Bayesian Optimization

In many practical scenarios of black box optimization, the objective function is subject to constraints that must be satisfied to avoid undesirable outcomes. Such constraints are typically unknown and must be learned during optimization. Safe Bayesian optimization aims to find the global optimum while ensuring that the constraints are satisfied with high probability. However, it is often sample-inefficient due to the small initial feasible set, which requires expansion by evaluating the objective or constraint functions, limiting its applicability to low-dimensional or inexpensive problems. To enhance sample efficiency, additional information from cheap simulations can be leveraged, albeit at the cost of safeness guarantees. This paper introduces a novel safe multi-task Bayesian optimization algorithm that integrates multiple tasks while maintaining high-probability safety. We derive robust uniform error bounds for the multi-task case and demonstrate the effectiveness of the approach on benchmark functions and a control problem. Our results show a significant improvement in sample efficiency, making the proposed method well-suited for expensive-to-evaluate functions.