[Paper Review] A Multi-task Selected Learning Approach for Solving 3D Flexible Bin Packing Problem

Lu Duan, Haoyuan Hu+5 more|arXiv (Cornell University)|Apr 17, 2018

Optimization and Packing Problems26 references44 citations

TL;DR

The paper presents a multi-task selected learning framework (MTSL) that jointly learns item sequence and orientation to minimize surface area in 3D flexible bin packing, outperforming several baselines and showing online production gains.

ABSTRACT

A 3D flexible bin packing problem (3D-FBPP) arises from the process of warehouse packing in e-commerce. An online customer's order usually contains several items and needs to be packed as a whole before shipping. In particular, 5% of tens of millions of packages are using plastic wrapping as outer packaging every day, which brings pressure on the plastic surface minimization to save traditional logistics costs. Because of the huge practical significance, we focus on the issue of packing cuboid-shaped items orthogonally into a least-surface-area bin. The existing heuristic methods for classic 3D bin packing don't work well for this particular NP-hard problem and designing a good problem-specific heuristic is non-trivial. In this paper, rather than designing heuristics, we propose a novel multi-task framework based on Selected Learning to learn a heuristic-like policy that generates the sequence and orientations of items to be packed simultaneously. Through comprehensive experiments on a large scale real-world transaction order dataset and online AB tests, we show: 1) our selected learning method trades off the imbalance and correlation among the tasks and significantly outperforms the single task Pointer Network and the multi-task network without selected learning; 2) our method obtains an average 5.47% cost reduction than the well-designed greedy algorithm which is previously used in our online production system.

Motivation & Objective

• Formalize 3D Flexible Bin Packing Problem (3D-FBPP) as a real-world packing task in e-commerce logistics.
• Develop a neural policy that jointly learns item sequence and orientation to minimize bin surface area.
• Mitigate task imbalance and leverage task correlation via a Multi-task Selected Learning (MTSL) framework.
• Evaluate on a large real-world dataset and in online AB tests to demonstrate practical impact.

Proposed method

• Use an encoder-decoder neural network with two output streams for sequence (items) and orientation, sharing a common representation.
• Incorporate an intra-attention mechanism in the sequence task to handle repeating-item constraints.
• Apply a hill-climbing inspired orientation learning with ground-truth orientation from best-ever solutions.
• Train with a hybrid loss L_all = alpha*L_seq + (1-alpha)*L_ori and a Multi-task Selected Learning schedule that samples among L_seq, L_ori, and L_all.
• Decode sequences with a pointer-network style mechanism and PPO for policy optimization.
• Evaluate with beam search (size 5) and sampling (128 solutions) to select best packing.

Experimental results

Research questions

• RQ1 Can a multi-task learning framework effectively learn both packing sequence and orientation to minimize bin surface area in 3D-FBPP?
• RQ2 Does Selected Learning help balance imbalanced and correlated tasks (sequence vs orientation) and improve solution quality over single-task or non-selected multi-task baselines?
• RQ3 What is the practical performance gain of the proposed method on large-scale real-world order data and in online production settings?

Key findings

• MTSL achieves up to 9.66% surface area reduction over the LWSC heuristic on BIN10 and BIN12 in the offline experiments and 6.16% to 8.25% gains on BIN8–BIN12 across various comparisons.
• MTSL significantly outperforms both the single-task Pointer Network and a multi-task model without Selected Learning.
• Online AB tests demonstrate an average 5.47% cost reduction in packing costs.
• Intra-attention for the sequence task improves packing surface area over vanilla RL models.
• Sampling multiple solutions (128) further improves results, with MTSL-sample often outperforming alternatives.