错误地勾结：算法复杂性驱动力是否具有竞争力？.pdf

1、IMF IMF|Statistics1Collusion by mistake:does algorithmic sophistication drive supra-competitive profits?NOVEMBER 21,2024Ibrahim AbadaGrenoble Ecole de ManagementIn collaboration with Xavier Lambin and Nikolay Tchakarov(ESSEC Business School)IMF IMF|Statistics2A DisclaimerA DisclaimerThe opinions exp

2、ressed in this presentation are those of the authors alone and might not represent the views of GEM or ESSEC.IMF IMF|Statistics3ContextContext The literature consistently reports that simple reinforcement learning algorithms systematically reach seemingly collusive outcomes.The drivers of cooperatio

3、n are being investigated:sophisticated punishment strategies to sustain the cartel(Calvano et al.2002b),numerical biases(cooperation bias Banchio and Mantegazza 2023),correlated learning(Lambin 2024),etc.Often simple Q-learning algorithms are tested with an implicit asusmption:“The enhanced sophisti

4、cation of learning algorithms makes it more likely that AI systems will discover profit-enhancing collusive pricing rules”in Calvano et al.2020a.IMF IMF|Statistics4The research questionsThe research questions Is algorithmic collusion always the aftermath of sophisticated punishment schemes deployed

5、by the algorithms?We develop a simple theoretical illustration of competing Q-learning algorithms in a basic social dilemma and show that(seeming)collusion can be an aftermath of imperfect exploration.We validate our results via simulations in a market environment.Does algorithmic sophistication mak

6、e seeming collusion easier?We simulate the competition between more sophisticated algos(Deep Learning Actor-Critic networks,Reinforce,and Exp3)and demonstrate that seeming collusion disappears.When agents are endowed with the possibility to choose the level of sophistication of the algorithms they u