Pricing via Artificial Intelligence : The Impact of Neural Network Architecture on Algorithmic Collusion

Classic artificial intelligence (Q-learning) algorithms have been capable of consistently learning supra-competitive pricing strategies in infinitely repeated Nash-Bertrand pricing games without human communication. Such algorithms have been able to converge due to the temporal correlation of consecutive states and actions in the learning process, which restores stationarity in an otherwise highly non-stationary setting. It is difficult for more realistic AI algorithms to converge, as the necessary training processes breaks the aforementioned temporal correlation, rendering the algorithms ineffective in learning reward-punishment strategies that result in collusive market outcomes. We adapt several widely used neural network architectures to the framework of model-free reinforcement learning and experimentally explore how the structure of AI algorithms affects market outcomes in a workhorse oligopolistic model of repeated price competition. While it is possible to train advance AI algorithms to always best respond in environments where the rival exercises a fixed strategy, it is unlikely that such algorithms can learn to coordinate in setting supra-competitive prices due to the non-stationarity of multi-agent learning processes, suggesting that algorithmic collusion may not be an immediate concern for antitrust authorities