Multi-agent RL bidders learn tacit collusion in electricity markets without ever being told to

arXiv·medium signalXBlueskyLinkedInCopy link

Posted 27 August, this paper models strategic bidding as a repeated game with imperfect public monitoring and runs multi-agent reinforcement learning over it, on the grounds that electricity markets are oligopolistic with repeated interaction among few participants and are structurally susceptible to non-competitive behavior. The authors build a multi-dimensional criteria set for judging collusion that goes beyond comparing profit against Nash equilibria. Agents learned to sustain supra-competitive outcomes matching tacit-collusion indicators despite never being instructed to collude, which is the clearest domain-specific evidence yet that autonomous pricing agents produce a regulatory problem their operators did not choose.