Permutations improve performance in three-dimensional bin packing problem

Xinyuan Tian, pengzhan qu, Hongsheng Qi

Keywords: permutation;deep reinforcement learning; 3D packing;

Abstract: In recent years, with the advent of deep learning and reinforcement learning, researchers have begun to explore the use of deep reinforcement learning to solve the three-dimensional bin packing problem. However, current innovations in the 3D bin packing problem primarily involve modifications to the network architecture or the incorporation of heuristic rules. Efforts to improve performance from the perspective of function approximation are relatively scarce. As is well known, one of the crucial theoretical foundations of deep learning is the ability of neural networks to approximate many functions. As such, we propose a method based on approximation theory that uses permutations to better approximate policy functions, which we refer to as Permutation Packing. Nonetheless, due to the high memory requirements when the number of permutations is large, we also propose a memory-efficient variation of Permutation Packing, which we call Limited-Memory Permutation Packing. Both methods can be efficiently integrated with existing models. We demonstrate the effectiveness of both Permutation Packing and Limited-Memory Permutation Packing from both theoretical and experimental perspectives. Furthermore, based on our theoretical and experimental results, we find that our methods can effectively improve performance even without retraining the model.

Primary Area: general machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4889