Working Papers
Recommender systems and efficient social learning (Draft)
Abstract: In this paper, I study a model of a streaming platform (for instance YouTube or Netflix) with a catalog of items whose quality are unknown. Short-lived users arrive in sequence and browse the catalog, paying a small cost each time they want to examine a new alternative. The platform observes their behavior and uses it to deduce information about the items' quality so as to enhance future user's experience. I show that the platform is able to distinguish high-quality items if and only if a condition linking the cost of browsing and the thickness of the tail of the distribution from which the quality of the items is drawn holds true. This condition relates to the ability of the platform to incentivize users to explore new items using the information gathered with the previous users. I pin down an easily implementable policy that guarantees efficient learning.
Algorithmic collusion with endogenous exploration (Draft)
Abstract: I study a two-stage model in which two players simultaneously choose an exploration parameter for their Q-learning algorithms, which then repeatedly play a one-shot game chosen from a class of social dilemmas including the prisoner's dilemma, first and second price auctions as well as Bertrand competition with horizontally differentiated products. The players collect the limit average payoffs obtained by their algorithms. I show that all equilibria are collusive : both players receive payoffs that are strictly higher than the payoffs received in the unique strict Nash equilibrium of the one-shot game. I then use extensive numerical simulations in a Bertrand duopoly and a parameterized prisoner's dilemma. Their results allow to gain insight on (i) the mechanism causing algorithmic collusion and (ii) the strategic role of exploration levels in the game. They reveal that in equilibrium, the players tend to choose algorithms that over-explore, which comes at the detriment of joint payoff. These findings have important implications for algorithmic collusion.
Algorithmic collusion with unsynchronized updating (with G. Abel and A. Kalogeratos) (Draft coming soon)
Abstract: We study a model of algorithmic collusion in continuous time in which two firms use Q-learning algorithms to set prices in a Bertrand duopoly. The firms update their prices at times dictated by a Poisson clock. We introduce a simple parameter controlling for the synchronicity of the firms' algorithms' updates, ranging from perfect synchronicity like in previous models of algorithmic collusion, to independence. Using extensive numerical experiments, we show that algorithmic collusion gradually disappears when the synchronicity between the algorithms' updates decreases, and completely vanishes when they update independently. Specifically, we show that (i) the payoffs collected in the long run by the algorithms are no different than the ones collected by algorithms playing randomly (ii) the punishment-reward strategies gradually appear when the level of synchronicity increases, and are absent when the algorithms update independently. We show the last point by recording a large number of the algorithms' reactions to unilateral price cuts and compare them with the reactions of strategy-agnostic algorithms. Performing a clustering task on these reactions reveals that for synchronous algorithms, the reactions to price cuts are very well distinguishable from that of random algorithms, while they are not distinguishable at all for algorithms that update independently. These findings indicate that algorithmic collusion critically relies on synchronous updating and is thus unlikely to happen in many economic situations.
Published Papers
Work in progress
Taking decisions together promotes cooperation in spatial social dilemmas (with G. Abel and A. Kalogeratos)
Pandora's box with endogenous entry