Hierarchical multi-agent reinforcement learning for joint pricing and inventory control in a manufacturer-led Stackelberg framework

The joint optimization of pricing and inventory decisions is fundamental to maximizing profitability within manufacturing enterprises. These decisions are often decomposed into a hierarchical Stackelberg game, where strategic pricing follows a long-term cycle, while operational inventory replenishment operates on a short-term frequency. Traditional bi-level programming methods struggle with the computational complexity of such industrial-scale stochastic environments, while existing Multi-Agent Reinforcement Learning (MARL) paradigms fail to address asynchronous time scales and the non-stationary nature of leader-follower interactions. To address these challenges, this paper presents a hierarchical MARL framework for intra-enterprise Stackelberg games. The framework introduces a hierarchical architecture that structurally enforces sequential decision-making, resolving the time-scale mismatch through a gradient-coupling training mechanism that enables the leader to implicitly learn the follower’s response patterns. Furthermore, a predictive guidance mechanism is implemented, allowing the leader to transmit demand forecasts derived from pricing strategies to the follower, thereby transforming inventory management from reactive to proactive control. Systematic experimental validation demonstrates that our method achieves up to 13.7% higher total reward and 9.3% improved order fulfillment rate, while consistently maintaining profitability and service level gains under different price elasticity conditions. • SGH-MAPPO embeds Stackelberg game logic into hierarchical MARL for joint pricing-inventory control. • Introduce gradient-coupled training enabling the leader to implicitly anticipate the follower’s response patterns. • Implement a predictive guidance module that shifts inventory management from reactive to proactive control. • SGH-MAPPO maintains consistent profit and service gains across varying price elasticity conditions.