Exploring the Use of Machine Learning to Solve the 2D Cutting and Packing Problem: Systematic Literature Mapping

This article investigates the application of Machine Learning (ML) to the Two-Dimensional Cutting and Packing Problem (2DCPP) under the amortized inference paradigm. A Systematic Literature Mapping was conducted following the PRISMA guidelines and the PICOC framework, resulting in the selection of 12 studies out of 181 initial records. The results indicate the consolidation of Deep Reinforcement Learning (DRL) from 2022 onwards, with a predominant use of Transformers and Multilayer Perceptrons (MLP) for regular formats, and Convolutional Neural Networks for irregular geometries (Nesting). The prevalence of hybrid approaches integrated with heuristics indicates that the feasibility of solving the 2DCPP purely via neural networks remains an open challenge, present in only 17% of the models. Furthermore, it is concluded that the field lacks empirical maturity and reproducibility, given that 83% of the studies do not provide source code and only 50% performed validation using standard benchmarks from the literature.