An adaptive population-based candidate search algorithm with surrogates for global multi objective optimization of expensive functions

We propose a new algorithm, MOPLS, for efficient global Multi Objective (MO) optimization of expensive functions. MOPLS is an iterative population-based parallel surrogate algorithm that incorporates simultaneous local candidate search on surrogate models to select numerous evaluation points in each iteration. The novel iterative framework of MOPLS simultaneously selects new points for evaluation by using i) Local Radial Basis Function (RBF) approximation, ii) surrogate-assisted neighborhood candidate search, and iii) a Tabu mechanism for adaptively avoiding neighborhoods that do not improve the non-dominated solution set. MOPLS is more efficient than ParEGO, Borg, NSGA-II and NSGA-III with application to 11 test problems and a watershed calibration problem, on a budget of 600 functions evaluations.