Poisson-MNL Bandit: Nearly Optimal Dynamic Joint Assortment and Pricing with Decision-Dependent Customer Arrivals
New AI model accounts for how prices affect customer traffic, achieving near-optimal regret bounds.
Deep Dive
Researchers Junhui Cai, Ran Chen, and team developed the 'Poisson-MNL Bandit' algorithm. It combines a Poisson arrival model with a Multinomial Logit (MNL) choice model to dynamically optimize product assortment and pricing over time T. Their UCB-based algorithm achieves a near-optimal regret bound of √(T log T). It learns how pricing decisions influence both customer arrivals and purchases, outperforming models that assume fixed traffic.
Why It Matters
Enables e-commerce and retail platforms to dynamically set prices that maximize revenue by accurately modeling real-world customer behavior.