Scholars Question the Explanation of Collusion Punishment Mechanisms in Q-Learning Pricing Agents

Kaue Lopes de Moraes and Rodrigo De-Losso have published research replicating Calvano et al.'s 2020 Bertrand duopoly Q-learning pricing simulation experiment, challenging the traditional punishment mechanism explanation for independently trained tabular Q-learning agents converging to supracompetitive prices. The study audited the original experimental conclusions under frozen policy states where learning was disabled. It found that in the original impulse response tests, single-period price cuts were unprofitable in 92% of pairs, but the optimal deviation calculated through dynamic programming was profitable on average by 17.5%; 914/1024 (ninety percent) of converged pairs had a single attractor covering the entire state space. When independently trained frozen agents were re-paired, coordination decreased by 80%, with no cross-pair strategy retention. The study concludes that the original impulse response patterns themselves cannot identify a collusion mechanism, and the associated payoffs are overdetermined rather than selective.

AI-generated from reports · 2 hr agoUpdated

Not enough consecutive observations yet; trend is not drawn.