Transaction selection policy in tier-to-tier SBSRS by using Deep Q-Learning

This paper studies a Deep Q-Learning (DQL) method for transaction sequencing problems in an automated warehousing system, Shuttle-based Storage and Retrieval System (SBSRS), in which shuttles can move between tiers flexibly. Here, the system is referred to as tier-to-tier SBSRS (t-SBSRS), developed as an alternative design to tier-captive SBSRS (c-SBSRS). By the flexible travel of shuttles between tiers in t-SBSRS, the number of shuttles in the system may be reduced compared to its simulant c-SBSRS design. The flexible travel of shuttles makes the operation decisions more complex in that system, motivating us to explore whether integration of a machine learning approach would help to improve the system performance. We apply the DQL method for the transaction selection of shuttles in the system to attain process time advantage. The outcomes of the DQN are confronted with the well-applied heuristic approaches: first-come-first-serve (FIFO) and shortest process time (SPT) rules under different racking and numbers of shuttles scenarios. The results show that DQL outperforms the FIFO and SPT rules promising for the future of smart industry applications. Especially, compared to the well-applied SPT rule in industries, DQL improves the average cycle time per transaction by roughly 43% on average.