Intelligent Replenishment and Allocation Algorithms in Alibaba Health's Pharmaceutical E‑commerce Supply Chain

The presentation, delivered by senior algorithm expert Long Guangbin, introduces Alibaba Health's pharmaceutical e‑commerce supply‑chain intelligent replenishment and allocation algorithms.

It outlines the end‑to‑end supply‑chain framework, which includes modules such as demand forecasting, stock‑placement, intelligent replenishment, intelligent allocation, warehouse‑network routing, courier assignment, and pricing, all driven by rich product, sales and label data.

The intelligent replenishment model addresses the "hold‑stock vs. order‑uncertainty" problem, focusing on turnover and stock‑out metrics, and explains how the algorithm determines which items to replenish, when to replenish, and how much to order. Model evolution is traced from traditional safety‑stock (ss) methods to operations‑research optimization and finally to deep‑learning/reinforcement‑learning approaches.

The safety‑stock (ss) model is described in detail, covering deterministic EOQ, stochastic safety‑stock with demand and lead‑time uncertainty, the formula for reorder points, and extensions such as dynamic DTIA modeling and promotion‑factor adjustments.

The reinforcement‑learning (RL) replenishment model is formulated with state, action, and reward components, employing an actor‑critic twin‑tower network with embedding layers and carefully shaped reward functions; offline simulations show stable turnover and a reduction of stock‑out days.

A dedicated simulation system is introduced to evaluate replenishment and allocation decisions by serially modeling market dynamics, supplier behavior, and agent actions, enabling the calculation of turnover, stock‑out, and next‑day‑delivery indicators.

The intelligent allocation model is presented next, with objectives to maximize next‑day delivery, minimize inter‑warehouse allocation cost, and reduce stock‑out days. Constraints include allocation limits, demand satisfaction, minimum order quantities, and box‑size requirements, and decision variables comprise the allocation‑quantity matrix and feasibility matrix.

Allocation is solved as a multi‑objective linear/non‑linear programming problem, using weighted‑sum or Pareto‑frontier methods, with solution techniques ranging from integer programming to evolutionary algorithms (EA, ALNS). Reported gains include a 5‑point reduction in stock‑out rate and a 10‑point increase in next‑day delivery.

The Q&A section addresses practical concerns such as turnover‑stock‑out trade‑offs, simulation validation, model attribution, automation levels, balancing replenishment with allocation, handling supplier uncertainty, long‑tail item strategies, reward design, solving time, and selection of Pareto‑optimal solutions.

The session concludes with thanks to the audience.