Adaptive Landscape-aware Constraint Handling with Application to Binary Knapsack Problem.

Real-world optimization problems frequently have constraints that define feasible and infeasible combinations of decision variables. Evolutionary algorithms do not inherently work with constraints, so they must be modified to include a suitable constraint handling technique (CHT) before they can be used to solve such problems. A range of different approaches to handling constraints have been used effectively with evolutionary algorithms, such as penalty-based, repair, feasibility rules, and bi-objective approaches. In this study we investigate different CHTs with an evolutionary algorithm on the 0/1 knapsack problem. We present results to show that performance complementarity exists between different CHTs on the knapsack problem. Landscape analysis is then used to characterize the search space of a large number of knapsack instances, and decision tree induction is used to derive rules for selecting the most appropriate CHT and switching it adaptively based on landscape features. We finally implement a landscape-aware CHT approach and show that it out-performs the constituent CHT approaches.