An Imitation Learning Approach for Computing Anticipatory Picking Decisions in Retail Distribution Centres

This paper describes a machine learning approach for controlling the flow of products from their respective storage locations to the packing area in a retail distribution center (DC). The goal of the algorithm is to maximise earned revenue from expedited-delivery online orders, by extending the `booking deadline' for placing such orders on the retailer's website. Extension of the deadline by making the component processes more efficient has been studied extensively, with marginally reducing returns. Instead, we anticipate the type and quantity of products that will be ordered on a given day, and pre-emptively launch the delivery process before actual orders are received. The algorithm combines predictive analytics (for forecasting orders) and imitation learning (for computing optimal picking decisions under capacity constraints). We show that forecasting using long short-term memory cells and decision-making using imitation learning is more successful in maximising earned revenues on a public data set, than several alternative approaches using heuristics and traditional time-series methods.