Constrained evolutionary optimization based on dynamic knowledge transfer

Abstract

While utilizing evolutionary algorithms (EAs) to solve constrained optimization problems (COPs), seeking feasible solutions that satisfy the constraints is the primary concern. In some cases, utilizing the knowledge of the objective may facilitate the exploration of feasible regions and the exploitation of global optima. To strike a balance between objective and constraints, we attempt to address two problems, namely whether transferring the knowledge of objective in the early evolutionary stage and how to dynamically transfer the knowledge of objective to constraints. First, whether transferring in the early stage is determined by the characteristics of the objective, which has been classified to simple and complex based on the variation of population distribution. Second, how to transfer is solved by designing suitable constraint handling techniques (CHTs). To deal with COPs with simple objective, an objective-oriented CHT is proposed, where an indicator called Knowledge Transfer Rate (KTR) representing the relationship between the objective and constraints is mined and used. For COPs with complex objective, a constraint-oriented CHT is proposed, which includes a constraint-driven strategy and a hybrid-driven strategy. The proposed method is executed on three benchmark test suites, namely IEEE CEC2006, CEC2010, and CEC2017, it achieves superior or at least competitive performance in comparison with other state-of-the-art methods. Furthermore, the proposed method is successfully implemented in a redundant robotic manipulator motion planning problem.