A computational framework for solving analytically intractable high-dimensional evolutionary games and applications in production
Evolutionary game theory (EGT) is widely used in production research to model coordination, regulation, and quality governance in supply chains and manufacturing systems. The classical analytical pipeline, deriving replicator equations, enumerating equilibria, and testing stability through Jacobian analysis, becomes intractable as players, strategies, and parameters grow, so realistic production-governance models can be built but not solved. This paper proposes a hierarchical Multi-Agent Reinforcement Learning (MARL) framework that finds parameter configurations that induce reliable convergence to a target profile under replicator dynamics, i.e. that make the target a potential evolutionarily stable strategy (ESS), meaning a profile reached from every tested initial condition. An inner loop of independent Proximal Policy Optimization (PPO) agents performs stateless self-play as decentralised solvers, while an outer Soft Actor-Critic (SAC) meta-optimiser searches the parameter space for configurations yielding stable convergence. Every recovered configuration is then validated after the search, independently of it, by multi-start replicator simulation from many random initial conditions and, where closed-form stability conditions exist, against those conditions. On a three-player supply-chain quality-governance game, the framework recovers the analytical stability landscape exactly. By returning a reliable interior ensemble of admissible configurations (100 independent optimisation runs) drawn from within the stable region for each target outcome, it converts an abstract stability question into concrete managerial levers, a scalable decision aid for production systems where analytical methods fail.Practitioner Summary: Strategic problems in production and supply-chain systems often depend on decisions made by many independent actors, such as suppliers, manufacturers, platforms, and regulators, whose competing incentives determine which behaviours persist. These settings are routinely modelled as evolutionary games, but once a model includes several actors with several options each, the standard analysis cannot be solved by hand: the number of possible outcomes grows explosively and the conditions deciding which outcomes are stable have no exact solution. Practitioners can therefore build a realistic model but cannot solve it. This paper provides a computational decision aid for exactly these cases. Rather than solving the stability conditions symbolically, it trains learning agents to play the system out and searches for the parameter settings that make a desired outcome stable and self-sustaining, then verifies each candidate against the underlying dynamics. The practical value is that the method returns not one answer but a reliable ensemble of settings drawn from the interior of the stable region for a target outcome, turning an abstract question into concrete levers: robust settings that keep the desired outcome stable with a safety margin and, when the least sufficient intervention is needed, the minimal level of a chosen lever, found by running the same procedure in a boundary-seeking mode. We demonstrate the approach on a supply-chain governance problem where the exact answer is known and the method recovers it, and the same procedure applies to other production and supply-chain games expressible as parameterised evolutionary games.
Authors
- Ridwan Al Aziz (ORCID: https://orcid.org/0000-0002-5871-5629)
- Mahamudul Hassan Siddique (ORCID: https://orcid.org/0009-0009-3868-1861)
- Fahimul Haque (ORCID: https://orcid.org/0009-0001-5003-7720)
Institutions
- Bangladesh University of Engineering and Technology (BD)
Publication Details
- Journal
- International Journal of Production Research
- Published
- 2026-11-21
- DOI
- https://doi.org/10.1080/00207543.2026.2762315
- Primary Topic
- Evolutionary Game Theory and Cooperation
- Type
- article
- Field-Weighted Citation Impact
- 0.00