Bayesian Forward-Inverse Transfer for Multiobjective Optimization

Tingyang Wei, Jiao Liu, Abhishek Gupta, Puay Siew Tan, Yew-Soon Ong

Abstract: We present an evolutionary optimizer incorporating knowledge transfer through forward and inverse surrogate models for solving multiobjective problems, within a stringent computational budget. Forward knowledge transfer is employed to fully exploit solution-evaluation datasets from related tasks by building Bayesian forward multitask surrogate models that map points from decision to objective space. Inverse knowledge transfer via Bayesian inverse multitask models makes possible the creation of high-quality solution populations in decision space by mapping back from preferred points in objective space. In contrast to prior work, the proposed method can improve the overall convergence performance to multiple Pareto sets by fully exploiting information available for diverse multiobjective problems. Empirical studies conducted on benchmark and real-world multitask multiobjective optimization problems demonstrate the faster convergence rate and enhanced inverse modeling accuracy of our algorithm compared to state-of-the-art algorithms.