Discovering Action Regions for Solving the Bin Packing Problem through Hyper-heuristics

Hyper-heuristics represent an innovative method for solving hard combinatorial problems such as the Bin Packing Problem. In this work, we propose a solution model that incorporates insights from unsupervised learning to produce solvers that base their decisions on maximizing a reward. The solution model takes the form of a hyper-heuristic. As the search progresses, this strategy chooses among different heuristics, adapting the solution process to the current problem state under exploration. The proposed model relies on the k-means clustering algorithm to identify the centroids of what we define as "action regions" (regions where we can recommend one particular heuristic). To recommend a heuristic in such action regions, we use a simple yet useful reward-based approach that analyzes the performance of individual heuristics. Then, the hyper-heuristic (a collection of action regions) decides which heuristic is the most suitable one to apply given one specific problem state. The experimental setup was carried out on a total of 580 bin packing instances wills promising results.