A Deep Reinforcement Learning-Based Dynamic Replenishment Approach for Multi-Echelon Inventory Considering Cost Optimization

In the fast-moving consumer goods (FMCG) industry, inventory management is a critical component of supply chain management because it directly impacts cost efficiency and customer satisfaction. For instance, effective inventory management can minimize overstocking and reduce replenishment delays, which are particularly important in multi-echelon supply chain systems characterized by high complexity and dynamic demand. This study proposes a method based on deep reinforcement learning (DRL) aimed at optimizing replenishment decisions in multi-echelon inventory systems for FMCG industries. We designed a Dynamic Replenishment FMCG Multi-Echelon Optimization (ME-DRFO) model and incorporated a Markov Decision Process (MDP) to model the multi-echelon inventory system. By applying an improved Soft Actor–Critic with an adaptive alpha and learning rate (SAC-AlphaLR) algorithm, which introduces adaptive temperature parameters and adaptive learning rate mechanisms, our approach not only dynamically adapts to environmental changes but also effectively balances exploration and exploitation, ultimately achieving global replenishment cost minimization while ensuring supply chain stability. Through numerical experiments, our method demonstrates excellent performance by reducing replenishment costs by 12.31% and decreasing inventory shortages to 2.21%, significantly outperforming traditional methods such as overstocking, Particle Swarm Optimization (PSO), and the standard Soft Actor–Critic (SAC). This research provides new theoretical insights into multi-echelon inventory optimization and practical solutions for effectively managing complex supply chains under uncertain and dynamic conditions.