Informative Detection of Collusion Sustained by Non-Punishment Mechanisms

Determine whether CURB’s total-variation detection signal remains informative when algorithmic collusion is sustained through mechanisms other than punishment dynamics.

Background

CURB is designed around the observation that collusive Q-learning policies can implement Simple Penal Codes: cooperation is sustained because a unilateral deviation triggers retaliatory behavior. The method therefore identifies conditional dependence between an opponent’s prior defection and the agent’s subsequent action distribution, and penalizes that dependence.

The paper acknowledges that algorithmic collusion may arise through alternative mechanisms that do not rely on punishment threats. It leaves unresolved whether the total-variation signal used by CURB would still detect such forms of collusion, making this a distinct open question about the scope and robustness of the proposed detector.