Solving the strip packing problem with a decomposition framework and a generic solver: Implementation, tuning, and reinforcement-learning-based hybridization

In the strip packing problem, the objective is to pack a set of two-dimensional items into a strip of fixed width such that the total height of the packing is minimized. The current state-of-the-art exact approach for the problem uses a decomposition framework in which the main problem (MP) fixes the item abscissas and the strip height, whereas the subproblem (SP) determines whether a set of item ordinates resulting in a feasible packing exists. Even though this decomposition framework has already been used several times in the literature, implementation details were often obfuscated, limiting the outreach of the approach. We address this issue by thoroughly describing and testing various builds for this framework, investigating important features such as the way to forbid an infeasible solution in the MP (e.g., by rejecting them or through a no-good cut) and the techniques used to solve the MP and the SP. One of our findings is that a minor implementation tweak such as changing the random seed between two MP iterations can bring the same level of improvement as a more involved feature such as strengthening the no-good cuts. From our extensive experiments, we identify two versions of the framework that produce complementary results: one where the main problem is solved with integer linear programming and the other where it is solved with constraint programming. We then train a reinforcement learning agent to find the best hybridization of these two algorithms and show that the resulting approach obtains state-of-the-art results on benchmark instances.