An Artificial Intelligence Based Method For Optimized Warehouse Storage Allocation

Order picking is a major driver of warehouse operation costs. With the objective of minimizing the time and cost required for picking customer orders in large warehouses, this paper presents an artificial intelligence (AI)-based algorithm for optimized warehouse storage allocation. Specifically, the linear upper confident bound (linUCB) algorithm, a contextual multi-arm bandit algorithm, is used to select storage locations for incoming products, that are optimal with regard to the expected stock removal. To facilitate the perception and decision making of the agent, dimensionality reduction by means of clustering is employed, enabling the linUCB agent to interact with a low dimensional representation of the warehouse. For training, we present a reward function that evaluates the agent's decision making based on the actual cost of picking a product from the warehouse. Because the calculation of the reward metric is exclusively based on actually incurred picking distances, human bias in the design of the reward function is minimized. In a practical case study, the suggested method is applied to a real warehouse layout with 4,650 storage locations, two picking areas and 30 different product categories. For the training and evaluation of the method a warehouse simulation is used. The performance of the linUCB agent is benchmarked against a conventional ABC allocation strategy. A comparison shows that the artificial intelligence-based storage allocation outperforms the ABC-method.