An autonomous pharmaceutical supply chain network for resilience and optimization: A multi-agent deep reinforcement learning framework
Year: 2026
Type: article
Abstract: The pharmaceutical supply chain must sustain high service levels under demand shocks, cold-chain constraints, and data-sharing limitations. However, many studies still separate forecasting from replenishment, rely on static policies, or evaluate resilience without a clearly specified control model. This paper develops an Autonomous Pharmaceutical Supply Chain Network (APSCN) that couples demand forecasting, synthetic time-series generation, and multi-agent deep reinforcement learning for resilient inventory control. The case study represents an anonymized regional United States vaccine distribution network with one central manufacturing hub and ten regional distribution centers operating over a 52-week horizon. A long short-term memory (LSTM) forecaster and an extreme gradient boosting (XGBoost) benchmark are trained on historical demand and time-series generative adversarial network (TimeGAN)-augmented disruption scenarios informed by a susceptible-infected-recovered (SIR) epidemic signal. Inventory decisions are then coordinated by multi-agent deep deterministic policy gradient (MADDPG) agents under explicit shelf-life, service-level, and cost constraints. The paper contributes a formal optimization and reward framework for cold-chain pharmaceutical logistics, a richer case-study description linking geography, network topology, and disruption design, and a comparative evaluation against autoregressive integrated moving average (ARIMA), Prophet, XGBoost, static rules, and single-agent reinforcement learning. In the simulated disruption scenario, the proposed LSTM + MADDPG configuration reduces mean absolute percentage error from 11.7% to 6.0%, lowers total cost from $1.25M to $0.85M, maintains a 99.1% service level, eliminates stockout incidents, and shortens recovery time from 28 to 4 days. The findings indicate that autonomous, decentralized control can materially improve both efficiency and resilience in pharmaceutical distribution networks.
Author: Divanshu Mittal
Institution: Dakota State University
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: diamond
OpenAlex ID: W7147490458