Abstract
The two-dimensional cutting stock problem (CSP) is critical in several industries. Reinforcement learning (RL) is a novel method to obtain a quality solution of two-dimensional CSP in a short computation time. In this research, we applied a model-free off-policy RL algorithm to an industrial example of exact guillotine two-dimensional CSP, and compared the results with mixed-integer programming (MIP), which is a common traditional mathematical method for optimization. The results showed that RL had a much lower computation time than MIP with a solution closed to optima, and the ability to make a trade-off between waste, inventory level, and back order.