Researchers have built a reinforcement learning framework that teaches rival train operators to price tickets dynamically — no back-channel coordination required.
The system, detailed in a new paper on arXiv, tackles a specific problem in liberalized rail markets: operators can't legally communicate with competitors, yet pricing decisions are deeply interdependent. Standard multi-agent reinforcement learning typically feeds agents flat, unstructured data about their environment. This team instead models the market as a graph of operational units — stations, routes, competitors — and uses a relational graph convolutional network to let each agent reason about how it sits within that web. A learned attention mechanism then decides which relationships matter most when setting a fare. The code is publicly available on GitHub.
The practical stakes are real. Rail privatization has spread across Europe and elsewhere, and opaque algorithmic pricing is already a flashpoint with regulators who worry that competing AIs can tacitly collude even without explicit communication — arriving at inflated prices by independently learning that everyone benefits from not undercutting. This paper doesn't solve that concern, but it does show a framework that outperforms baselines on revenue and price stability, which is exactly the kind of result operators will cite when arguing their systems are well-behaved.
The irony is that a tool built to respect anti-collusion rules could, if deployed carelessly, make algorithmic tacit collusion harder for regulators to detect — a tension the paper doesn't address.