AI trading agents can quietly learn to fix prices with each other, even without ever exchanging a message. New research proposes a fix.
The problem: when reinforcement learning agents repeatedly compete, such as in pricing algorithms, they can converge on supra-competitive outcomes that look like collusion, even though no one programmed them to cooperate. Researchers formalized this behavior using Simple Penal Codes, a game-theory framework for punishment-based cooperation, and showed that any meaningful version of it leaves a measurable fingerprint: a detectable difference in how an agent acts after cooperation versus after defection. They built on that signal to create CURB (Collusion Unwinding via Reward shaping and Belief injection), which penalizes that fingerprint during training. The method is proven to eliminate the punishment threats that let collusive equilibria hold.
Tested in Bertrand and Cournot repeated-game simulations, two standard economic models for price and quantity competition, CURB substantially cut collusion by Q-learning agents. The researchers also extended CURB to deep Q-network agents in Bertrand competition specifically, suggesting the approach isn't limited to simple tabular learning setups.
This matters because algorithmic collusion has mostly been treated as a narrow problem for specific markets, like ad auctions or ride-hailing platforms. A general-purpose fix that works across repeated-game settings is a meaningfully different proposition: it's a tool regulators or platform operators could plausibly bolt onto pricing algorithms before deployment, rather than patching each new case after the fact.
Worth noting: this is simulation work, not evidence pulled from real trading or pricing systems, and deep-network results so far cover only one of the two tested market structures.