EURO 2024 Copenhagen
Abstract Submission

EURO-Online login

1327. Solving a 3D-Bin-Packing-Problem with Deep Reinforcement Learning

Invited abstract in session MC-7: Cutting and Packing 3 - 3D loading, stream Cutting and Packing (ESICUP).

Monday, 12:30-14:00
Room: 1019 (building: 202)

Authors (first author is the speaker)

1. Hendrik Jon Schmidt
SEE, Deutsches Forschungszentrum für Künstliche Intelligenz
2. Dennis Maecker
German Research Center for Artificial Intelligence

Abstract

The 3D Bin-Packing-Problem (3D-BPP) is an NP-hard problem that is particularly important in Operations Research whenever a load must be placed on a carrier. This problem has been studied for a long time from a heuristic perspective in the area of container and pallet loading. Recently, Deep Reinforcement Learning (DRL) methods have been applied to solve 3D-BPPs. These methods show improved performance in decreasing the inference time for generating solutions while their generated solutions can outperform those generated by heuristics. However, the research has paid little attention to generating loading patterns subject to stability constraints, which are relevant for making the generated solutions applicable in real-world settings. This paper develops a DRL algorithm based on Proximal Policy Optimisation to explore whether DRL methods can be used to generate stable loading patterns for placing parcels on a Euro-pallet. Only the offline 3D-BPP is considered. The DRL algorithm is combined with a physical simulation of the generated loading pattern to ensure the load's stability. The algorithm is evaluated based on the recently introduced BED-Benchmark dataset of realistic orders. The evaluation shows that the developed DRL algorithm, on its own, struggles to find suitable solutions for stably placing parcels on a pallet. One reason for this might be the large action space for placing any parcel. This issue could be addressed by combining the algorithm with heuristic solvers.

Keywords

Status: accepted

Back to the list of papers