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

20 Pages Posted: 26 Jun 2023

See all articles by Weipeng Zhang

Weipeng Zhang

Georgetown University; Georgetown University - Department of Economics

Date Written: June 23, 2023

Abstract

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.

Keywords: Repeated games, algorithmic collusion, experience replay, deep reinforcement learning, deep Q-learning, deep recurrent Q-learning

JEL Classification: C72, C73, D43, K21, L1, L13, L4, L51

Suggested Citation

Zhang, Weipeng, Pricing via Artificial Intelligence: The Impact of Neural Network Architecture on Algorithmic Collusion (June 23, 2023). Available at SSRN: https://ssrn.com/abstract=4489688 or http://dx.doi.org/10.2139/ssrn.4489688

Weipeng Zhang (Contact Author)

Georgetown University ( email )

Washington, DC 20057
United States

Georgetown University - Department of Economics ( email )

Washington, DC 20057
United States

Do you have negative results from your research you’d like to share?

Paper statistics

Downloads
69
Abstract Views
233
Rank
604,425
PlumX Metrics