Mitigating Dimensionality in 2D Rectangle Packing Problem under Reinforcement Learning Schema

TL;DR

This paper introduces a reinforcement learning approach with reduced state-action representations for 2D rectangle packing, achieving comparable results to heuristics and showing potential for broader applications.

Contribution

It presents a novel RL-based method using UNet and PPO with reduced representations, improving generalizability to complex packing problems.

Findings

Achieved packing performance comparable to MaxRect heuristic.

Demonstrated potential for generalization to nonrectangular packing.

Proposed a scalable RL framework with reduced state-action spaces.

Abstract

This paper explores the application of Reinforcement Learning (RL) to the two-dimensional rectangular packing problem. We propose a reduced representation of the state and action spaces that allow us for high granularity. Leveraging UNet architecture and Proximal Policy Optimization (PPO), we achieved a model that is comparable to the MaxRect heuristic. However, our approach has great potential to be generalized to nonrectangular packing problems and complex constraints.