Reinforcement Learning and Heuristic Based Approach for Solving the Cutting Stock Problem: A Comparative Analysis

The Cutting Stock Problem (CSP) is a popular optimization problem in operations research that deals with cutting large stock materials into smaller pieces to satisfy demands while minimizing waste. This study provides a comprehensive comparative analysis of algorithmic strategies ranging from classical heuristics (Greedy, First Fit Decreasing) to advanced techniques like Genetic Algorithms, Column Generation, and Reinforcement Learning (RL). We propose a novel Q-learning-based RL framework tailored to spatially-aware, shape-diverse CSP instances, incorporating constraints from geometric shapes such as rectangles, circles, triangles, and half-circles. Unlike conventional approaches that rely on discrete pattern selection, our method models CSP as a sequential decision-making problem in a continuous two-dimensional environment. The RL agent learns to optimize cutting patterns via rewards based on area efficiency and trim loss. Experiments conducted across multiple simulated instances demonstrate that RL significantly outperforms baseline algorithms in all key metrics—trim loss, shape formation efficiency, and area fitness efficiency.