Event Details

This session from Sebastian Lautz and Aleksandar Kolev of Tesco's Data Science team will be delivered as both a Management Science seminar and also a guest lecture for students on the NATCOR Stochastic Modelling course.

Abstract: Tesco, the largest retailer in the UK with over 3,000 stores, faces a variety of optimisation challenges daily. These range from truck route planning and online advertising to in-store fraud detection and community support, as well as forecasting seasonal product demand and estimating minute-by-minute demand for items like tomatoes. This talk will focus on our efforts to price items nearing the end of their shelf life, a process known as markdown or reduced to clear. The most straightforward approach is 'predict-then-optimise,' which involves two components: predicting demand based on available features, followed by optimisation to determine the best pricing strategies given business constraints.

Predicting demand is more complex than a simple supervised learning problem, due to the limited diversity of the training dataset. Additionally, the optimisation step is a complex, slow, and resource-intensive task that must balance multiple objectives, such as profit, waste, infrastructure, display duration, capacity, and labour costs. The manual interventions by tens of thousands of colleagues and shoppers daily add significant noise to the data generation process. An alternative to 'predict-then-optimise' is to learn an optimal pricing policy directly using reinforcement learning (RL). RL is a powerful framework for sequential decision-making under uncertainty, with applications ranging from drug discovery to autonomous driving and computer vision. We will briefly outline the application of RL to markdown pricing, including problem formulation, potential solutions, and challenges.

This seminar will be delivered as both a Management Science seminar and also a guest lecture for students on the NATCOR Stochastic Modelling course.