AN ACTIVE-SET EVOLUTION STRATEGY FOR MIXED-INTEGER BLACK-BOX OPTIMIZATION WITH EXPLICIT CONSTRAINTS

Abstract

Many real-world applications involve the optimization of both continuous and discrete variables simultaneously. Evolution Strategies are stochastic black-box optimization techniques that are most commonly used for continuous optimization. Recently, the Active-Set Evolution Strategy has been developed for constrained continuous optimization problems, which assumes that the objective function is a black box, but the constraint functions are explicit and computationally inexpensive to evaluate. This assumption allows the algorithm to evaluate the feasibility of constraints at multiple points before spending an objective function evaluation. In this thesis, we propose a Mixed-Integer Active-Set Evolution Strategy for solving black-box mixed-integer optimization problems with explicit constraints. A heuristic is employed in place of the bounding mechanism to select the node to propagate forward, rather than solving subproblems to optimality. Then, we analyze its behaviour on linearly constrained sphere problems and conduct computational experiments to compare its performance against several algorithms designed for constrained optimization.