查看文章

In real-world applications, most optimization problems contain constraints (ranging from physical, time, geometric, design etc.) which need to be satisfied while finding an optimal solution. However, the presence of constraints alters the shape of the search space making it difficult to solve. In the last few decades, stochastic search algorithms such as evolutionary algorithms have gained popularity due to their effectiveness in solving optimization problems. However, since evolutionary algorithms or most meta-heuristics naturally designed for unconstrained optimization problems require additional mechanisms to solve constrained optimization problems.

Initially, the effectiveness of different penalty functions (both in evolutionary algorithms and in mathematical programming) has been investigated for several decades. However, the penalty functions have, in general, several limitations. For instance, they are not suitable for optimization problems where the optimum is on the boundary connecting the feasible and the infeasible regions or when the feasible region is disjoint. In addition, penalty functions require intensive fine-tuning to identify the most appropriate penaltyfactors to be employed. In the last few decades, a variety approaches have been proposed in conjunction with evolutionary algorithms to handle constraints. Most popular among them are self-adaptive penalty, epsilon constraint handling, superiority of feasible and stochastic ranking. Recently, the idea of ensemble of different constraint handling methods was proposed where each constraint handling method is apt only for a group of problems [1].