Adaptive cost-aware Bayesian optimization

Abstract

Cost-aware optimization is a common and important problem in real-world optimizations. Since real-world optimization problems are costly and have no specific mathematical formula, Bayesian optimization (BO) is frequently used to optimize these black-box expensive functions. Typically, a total budget is assigned for BO to find the optimal solution, but how to efficiently use the given budget has not been carefully investigated. In this paper, we propose a single-objective cost-aware BO framework to efficiently optimize an expensive black-box function with regard to the budget. Our proposed method utilizes a multi-armed bandit algorithm to quickly figure out a suitable strategy to deal with the cost of the optimization problem. It is flexible in adapting to different types of optimum-cost relations, extendable to multiple strategies, and simple to implement. We conduct a comprehensive set of experiments on both synthetic and real-world optimization problems to demonstrate the advantages of our method. Experimental results show that our proposed method outperforms other cost-aware BO methods.