A Machine Learning Application for Transaction Picking in a Tier-to-tier SBS/RS

With the recent growth of e-commerce, the order profiles have shifted towards smaller quantities with faster delivery time requests of customers. This change has led to companies seek for fast transaction processing automation technologies in operations of warehouses. Shuttle-based storage and retrieval system (SBS/RS) is an automated warehousing technology mostly utilized in large distribution centers because of its capability of processing high transaction rate. While the advantage of this system is its capability of processing high transaction rate by the excess numbers of shuttles in the system, a disadvantage is that the average utilization of shuttles is very low, compared to the lifting mechanisms in the system. Since a dedicated shuttle is assigned at each tier of an aisle, this system is also referred as tier-captive SBS/RS in literature. In an effort to balance the utilization levels of shuttles and lifts, a novel design referred as tier-to-tier SBS/RS is introduced. In that design, there is decreased number of shuttles in the system so that they are allowed to travel between tiers by using a separate lifting mechanism specifically dedicated for travel of them. This novel design not only balances the service lifts and shuttles, but also decreases the initial investment cost for the system by the decreased number of shuttles. However, those advantages cause a disadvantage, that is increased average cycle time per transaction performance metric in the system. In this thesis, in an effort to contribute on decreasing average cycle time per transaction performance metric, we apply a machine learning methodology for smart transaction processing in the system. Specifically, we apply Reinforcement Learning and Deep Reinforcement Learning methods for transaction selection of shuttles. The proposed approaches are compared with well-known First-in-First-out (FIFO) and Shortest Process Time (SPT) selection rules. The results show that the proposed approaches outperform both FIFO and SPT rules, significantly.