Collaborative production control and distributor selection via multi-agent reinforcement learning with differentiable communication

Collaborative production control and distributor selection are essential for resource allocation and meeting the core of Industry 5.0’s human-centric vision. However, traditional approaches typically handle these decisions independently, failing to adequately address fluctuating market conditions, demand uncertainty, and varying distributor competencies. This paper integrates production control and distributor selection as a Partially Observable Markov Decision Process (POMDP) in a multi-agent system. Specifically, a production control agent optimizes outputs by balancing inventory levels and opportunity costs, while a distributor selection agent dynamically adjusts allocations considering workforce skill diversity, cost efficiency , and equity. The formulated POMDP is solved using a multi-agent reinforcement learning (MARL) framework featuring a differentiable communication layer and GRU-based recurrent neural networks . Numerical experiments conducted under both stable and highly volatile market conditions demonstrate the proposed system’s enhanced adaptability and responsiveness. In particular, inter-agent messaging communication leading to improved welfare metrics and robust performance under diverse distributor-weight configurations. Notably, the resulting system promotes equitable distributor involvement, aligning with Industry 5.0’s emphasis on sustainable, people-centric supply chain operations. • We formulate the collaborative production and distributor selection problem as a partially observable Markov decision process. • The uncertainties in demand, supply, and the stochastic capabilities of distributors are rigorously considered. • We proposed a reinforcement learning approach with a differentiable communication layer to facilitate inter-agent coordination. • Empirical results show improved efficiency, market responsiveness, and equal opportunities for distributors.