Authors:
Mohamad Omar Nachawati
and
Alexander Brodsky
Affiliation:
Department of Computer Science, George Mason University, Fairfax, Virginia, U.S.A.
Keyword(s):
Constrained Optimization, Grey-Box Optimization, Interval Analysis, Mixed-Integer Programming, Non-convex Optimization, Nonlinear Programming, Simulation Optimization, Surrogate-based Optimization.
Abstract:
In this paper an algorithmic framework, called GreyOpt, is proposed for the heuristic global optimization of simulations over general constrained mixed-integer sets, where simulations are expressed as a grey-box, i.e. computations using a mix of (1) closed-form analytical expressions, and (2) evaluations of numerical black- box functions that may be non-differentiable and computationally expensive. GreyOpt leverages the partially analytical structure of such problems to dynamically construct differentiable surrogate problems for multiple regions of the search space. These surrogate problems are then used in conjunction with a derivative-based method to locally improve sample points in each region. GreyOpt extends Moore interval arithmetic for approximating the intervals of grey-box objective and constraint functions by fitting quadric surfaces that attempt to roughly underestimate and overestimate embedded black-box functions. This serves as the foundation of a recursive partitioning
technique that GreyOpt uses to refine the best points found in each region. An experimental study of GreyOpt’s performance is conducted on a set of grey-box optimization problems derived from MINLPLib, where the ratio of black-box function evaluations to analytical expressions is small. The results of the study show that GreyOpt significantly outperforms three derivative-free optimization algorithms on these problems.
(More)