Integrating AI with Bayesian Tracking in Supply Chain Management

Abstract

We study online replenishment under non-stationary demand, where decisions must be made sequentially from noisy demand-related signals while the underlying process can drift. We propose Artificial Intelligence Bayesian Tracking Supply Chain Management (ABT-SCM), which separates state estimation from ordering by combining Bayesian tracking of a latent demand driver (state-space model) with an optimistic exploration-exploitation rule over a discrete order grid. Each period, ABT-SCM updates a belief via Kalman filtering and chooses an order quantity by maximizing an upper-confidence objective computed from a learned reward surrogate. We compare ABT-SCM with Thompson Sampling, Sliding-Window Upper Confidence Bound, Discounted UCB, Sliding-Window Thompson Sampling, EXP3-IX, and a Random policy using cumulative reward and cumulative cost-regret relative to a discrete oracle. Over 30 replications, ABT-SCM delivers statistically significant gains on synthetic non-stationary data and remains robust to stronger drift, higher observation noise, heavy-tailed shocks, asymmetric holding or backorder costs, and alternative reward mappings. In multi-node supply-chain networks with chain, star, and Erdos-Renyi topologies, it consistently improves reward and reduces regret, suggesting robustness under stylized network scaling. On a real dataset evaluated via block bootstrap, ABT-SCM achieves the lowest mean cumulative cost-regret, with statistically significant regret improvement against EXP3-IX and directionally favorable but not consistently significant differences against the other baselines.