Multi-Objective Reinforcement Learning for Sustainable Supply Chain Optimization

Building sustainable supply chains (SC) has become a strategic focus for many organizations due to the increasing regulatory requirements and pressure from their conscientious customers. These SC focus on integrating social and environmental considerations into the design and management of the system, with reducing greenhouse gas (GHG) emissions being one of the primary goals. Reducing GHG emissions requires a holistic approach that tracks and considers emissions throughout the SC, from manufacturing to delivering the products to end customers. Toward this goal, we study a finite-horizon multi-echelon SC with emission considerations under carbon cap-and-trade legislation policy and formulate it as a multi-objective Markov decision process, where inventory replenishment and carbon management decisions are made at each period. One of the main technical challenges in creating sustainable multi-echelon SC is balancing environmental objectives with traditional economic metrics, such as cost and service level. Applying deep multi-objective reinforcement learning algorithms, we obtain a set of Pareto near-optimal policies that achieve a wide range of tradeoffs between cost and environmental objectives. We further analyze these policies and provide managerial insights into their decisions and performance. The performance of our derived sustainable inventory policies is also compared with that of a traditional inventory policy without emission considerations. In addition, we perform a sensitivity analysis to assess how changes in environmental factors, including the emission allowance and the carbon credit price, would impact the policies.