Abstract
This paper studies a new type of 3D bin packing problem (BPP), in which anumber of cuboid-shaped items must be put into a bin one by one orthogonally.The objective is to find a way to place these items that can minimize thesurface area of the bin. This problem is based on the fact that there is nofixed-sized bin in many real business scenarios and the cost of a bin isproportional to its surface area. Based on previous research on 3D BPP, thesurface area is determined by the sequence, spatial locations and orientationsof items. It is a new NP-hard combinatorial optimization problem onunfixed-sized bin packing, for which we propose a multi-task framework based onSelected Learning, generating the sequence and orientations of items packedinto the bin simultaneously. During training steps, Selected Learning choosesone of loss functions derived from Deep Reinforcement Learning and SupervisedLearning corresponding to the training procedure. Numerical results show thatthe method proposed significantly outperforms Lego baselines by a substantialgain of 7.52%. Moreover, we produce large scale 3D Bin Packing order data setfor studying bin packing problems and will release it to the researchcommunity.