Collusion among autonomous pricing algorithms utilizing function approximation methods

The increased prevalence of pricing algorithms incited an ongoing debate about new forms of collusion. The concern is that intelligent algorithms may be able to forge collusive schemes without being explicitly instructed to do so. I attempt to examine the ability of reinforcement learning algorithms to maintain collusive prices in a simulated oligopoly of price competition. To my knowledge, this study is the first to use a reinforcement learning system with linear function approximation and eligibility traces in an economic environment. I show that the deployed agents sustain supra-competitive prices, but tend to be exploitable by deviating agents in the short-term. The price level upon convergence crucially hinges on the utilized method to estimate the qualities of actions. These findings are robust to variations of parameters that control the learning process and the environment.