Multi-Objective Optimization of Supply Chain Network Design Using Hybrid NSGA-III and Deep Reinforcement Learning

Modern supply chains operate in a highly volatile environment, demanding network designs that are not only cost-efficient but also responsive and resilient to disruptions. Traditional optimization models often fail to capture the dynamic nature of operations and the complex trade-offs between multiple conflicting objectives. This paper proposes a novel two-level hybrid framework that integrates a multi-objective evolutionary algorithm (NSGA-III) with a Deep Reinforcement Learning (DRL) agent for designing robust supply chain networks. At the strategic level, the Non-dominated Sorting Genetic Algorithm III (NSGA-III) is employed to explore the solution space of potential network configurations (e.g., facility locations) to optimize three key objectives: minimizing total cost, minimizing average delivery time (enhancing responsiveness), and maximizing resilience against disruptions. At the operational level, for each candidate network design evaluated by NSGA-III, a Deep Q-Network (DQN) agent is trained to learn optimal, dynamic inventory and distribution policies in the face of stochastic demand and link disruptions. The performance of the DRL-derived policy provides the fitness values for the NSGA-III objectives. Through extensive simulations on a benchmark case study, the proposed hybrid NSGA-III-DRL approach is shown to generate a well-distributed Pareto front of optimal solutions, offering decision-makers a diverse set of network designs that effectively balance cost, responsiveness, and resilience. Comparative analysis demonstrates the superiority of our hybrid model over traditional static optimization methods and single-level learning approaches.

Keywords:

Supply Chain Network Design , Multi-Objective Optimization , NSGA-III , Deep Reinforcement Learning , Resilience , Hybrid Algorithm