Can Machines Learn Whether Machines Are Learning to Collude?
Online economic interactions generate data that artificial intelligence (AI), machine learning (ML) and deep learning (DL) can use: business predictive analytics, process optimisation and market power; consumer search and choice; and government gathering evidence and regulating harmful behaviour. In algorithmic collusion (AC), revenue management algorithms implement tacitly collusive behaviour. This paper summarises theoretical and empirical evidence and considers how ML methods affect AC and whether regulators’ algorithms can help. It examines links between Internet regulation and competition policy.
Early ML literature concerned programmes ‘learning’ their environments, e.g. predicting rivals’ prices to maximise profit by estimating prices/costs, identifying strategies or influencing learning. Here, ML is AI that self-programs to optimise specific objectives (data and model ‘layers’) and DL is many-layered ML. Increased depth makes behaviour an intricate convolution of data and programme history; invisible to programmers and inexplicable to others. ML by many firms may fail to converge or have unintended consequences.
Many models use simple ML algorithms to demonstrate behaviour consistent with collusion. It is not classically collusive without communication. Populations of simple AI can learn reward/punishment strategies that sustain profitable outcomes. This paper considers further variations taking into account strategic variations, finite-memory or dominance elimination and the impact of product characteristics and search. Simulation illustrates classic inefficiencies (overshoot, convergence to supracompetitive prices, cycles and endogenous market-sharing).
It is not clear what regulators could or should ban; can they detect AC or limit its consequences? We consider: restricting information available to firms; constraining price dynamics; coding standards that incorporate regulatory compliance in ML objectives; and algorithmic detection of specified anticompetitive behaviours. For instance, likelihood-ratio policy gradient reinforcement learning algorithms are more likely to converge to collusive behaviours when they take other firms’ learning into account and able to shape others’ learning with suitable prevalence of AI and network topology.
KeywordsAlgorithmic collusion Machine learning Antitrust
- 1.Cave, J., Gelsomini, L.: Single-bank proprietary platforms (2010). https://efmaefm.org/0EFMAMEETINGS/EFMA%20ANNUAL%20MEETINGS/2012-Barcelona/papers/EFMA2012_0482_fullpaper.pdf. Accessed 02 Sept 2019
- 3.Katz, M.L., Shapiro, C.: Network externalities, competition, and compatibility. Am. Econ. Rev. 75(3), 424–440 (1985)Google Scholar
- 12.Hu, J., Wellman, M.P.: Multiagent reinforcement learning: theoretical framework and an algorithm. In: ICML, vol. 98, pp. 242–250, July 1998Google Scholar
- 13.Greenwald, A., Hall, K., Serrano, R.: Correlated Q-learning. In: ICML, vol. 3, pp. 242–249, August 2003Google Scholar