Solving an Order Batching and Sequencing Problem with Reinforcement Learning

Abstract

The purpose of this research is to determine whether a DRL solution would be a suitable solution for the OBSP problem and to compare it with traditional methods. For this purpose, models trained with the PPO algorithm were tested in a complex and realistic warehouse environment, and an attempt was made to measure whether a strategy was developed to decrease the number of orders being late. A heuristic method was also applied and the results were compared on the same environment and data. The results showed that DRL approach that combines heuristics with the PPO algorithm has a better performance than the heuristics in minimizing the tardy order percentage in all tested scenarios.