A Concept for Optimal Warehouse Allocation Using Contextual Multi-Arm Bandits

This paper presents and demonstrates a conceptual approach for applying the Linear Upper Confidence Bound algorithm, a contextual Multi-arm Bandit agent, for optimal warehouse storage allocation. To minimize the cost of picking customer orders, an agent is trained to identify optimal storage locations for incoming products based on information about remaining storage capacity, product type and packaging, turnover frequency, and product synergy. To facilitate the decision-making of the agent for large-scale warehouses, the action selection is performed for a low-dimensional, spatially-clustered representation of the warehouse. The capability of the agent to suggest storage locations for incoming products is demonstrated for an exemplary warehouse with 4, 650 storage locations and 30 product types. In the case study considered, the performance of the agent matches that of a conventional ABC-analysis-based allocation strategy, while outperforming it in regards to exploiting inter-categorical product synergies.