Efficient Multi-Objective Assembly Sequence Planning via Knowledge Transfer between Similar Assemblies

Alexander Cebulla, Tamim Asfour, Torsten Kröger

Abstract: Industry 4.0 aims to automate the production of customized product variants, which presents several challenges, particularly in the realm of assembly sequence planning (ASP). Manufacturers are often interested not only in a viable sequence but also in optimizing multiple additional objectives. However, because there are N! potential sequences for an assembly containing N parts, discovering such sequences can be time-consuming. To accelerate this process, we propose an approach that combines Monte Carlo Tree Search (MCTS) with deep learning to effectively transfer knowledge between similar assemblies. Specifically, we employ learnable state-action functions using graph neural networks for two common objectives: minimizing the number of direction changes and maximizing part accessibility. After pretraining these functions on similar assemblies, we could use them to efficiently guide an MCTS such that it found assembly sequences that optimized both objectives for two sets of 3D puzzles consisting of either 38 or 58 parts. In fact, for both sets, our approach outperformed both the unmodified MCTS and an MCTS that utilized state-action functions trained during the search.