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

Approach and decision mechanism

The study models the coupled decisions as a partially observable multi-agent system. Separate agents choose production and distributor allocations, using recurrent neural networks and a differentiable communication layer.

1. ProductionBalance output, inventory, and opportunity costs.
2. AllocationConsider distributor skills, costs, and equity.
3. CommunicationShare learned messages between the decision agents.

Communication lets each agent account for information relevant to the other decision.

Evidence and scope

Numerical experiments examine stable and volatile markets, including different distributor-weight configurations.

These results concern the modeled scenarios; applying the approach elsewhere requires checking local constraints and the meaning of equity.

Cite this paper

Deng, Y., Sheng, G., Chow, A. H. F., Zhou, Z., Bai, Q., & Su, Z. (2025). Collaborative production control and distributor selection via multi-agent reinforcement learning with differentiable communication. Expert Systems with Applications, 282, 127539. https://doi.org/10.1016/j.eswa.2025.127539

@article{deng2025collaborative,
  title = {Collaborative production control and distributor selection via multi-agent reinforcement learning with differentiable communication},
  author = {Yang Deng and Guojun Sheng and Andy H. F. Chow and Zhili Zhou and Qinyang Bai and Zicheng Su},
  journal = {Expert Systems with Applications},
  year = {2025},
  volume = {282},
  pages = {127539},
  doi = {10.1016/j.eswa.2025.127539},
  url = {https://doi.org/10.1016/j.eswa.2025.127539}
}

Related research

• Coordinating delivery assignments and courier service fees under uncertainty.
• Planning production when downstream retail capabilities are uncertain.
• Recovering value from canceled food orders through pricing and reassignment.