Sustainable Material Cutting Optimization Using Deep Q-Networks: A Reinforcement Learning Approach for Resource Efficiency

The University of Edinburgh, School of Informatics, Scotland

^* Corresponding author: s2231967@ed.ac.uk

Abstract

This paper proposes an innovative approach to the Cutting Stock Problem (CSP) by integrating Graph Neural Networks (GNN) which effectively extract and process graph-structured data and Deep Reinforcement Learning (DRL) which utilizes the data generated by the GNN model to make sequential cutting decisions. The GNN model is embedded with Graph Convolutional Networks (GCN) layers, while the DRL model is structured with Deep Q-network (DQN). In my own study using KTH-TIPS dataset for model training, I have achieved promising experimental outcomes, decreasing loss functions and stabilizing total rewards, which demonstrates the model’s ability to progressively decrease prediction errors in stages and reach the best cutting patterns.