Abstract
Many real-world applications involve black-box optimization of multipleobjectives using continuous function approximations that trade-off accuracy andresource cost of evaluation. For example, in rocket launching research, we needto find designs that trade-off return-time and angular distance usingcontinuous-fidelity simulators (e.g., varying tolerance parameter to trade-offsimulation time and accuracy) for design evaluations. The goal is toapproximate the optimal Pareto set by minimizing the cost for evaluations. Inthis paper, we propose a novel approach referred to as information-TheoreticMulti-Objective Bayesian Optimization with Continuous Approximations (iMOCA)}to solve this problem. The key idea is to select the sequence of input andfunction approximations for multiple objectives which maximize the informationgain per unit cost for the optimal Pareto front. Our experiments on diversesynthetic and real-world benchmarks show that iMOCA significantly improves overexisting single-fidelity methods.