AI Collusion in Procurement Auctions

Abstract

Auctions are an important format for defense procurement, historically accounting for at least $1 billion per year (DoD Inspector General, 2012; GAO, 2018). A common format is the reverse auction, in which potential sellers submit bids for a contract, and the buyer selects the lowest bid (Alper & Boning, 2003; Coughlan et al., 2008). This design is well-suited for small uniform contracts where price is the only dimension of choice; common applications include commodities and routine services. However, this setting is also conducive to the application of AI bidding algorithms by sellers. Such algorithms generally seek to maximize profit for their owners. An obvious way to do this is to collude. AI algorithms for bidding introduces the possibility that competing algorithms collude to increase prices without explicitly communicating with each other. I conduct simulations in a simplified environment to test this possibility. I find that Q-learning algorithms indeed converge to supra-competitive prices in reverse auctions, allowing firms to extract higher than competitive profit. The learned policies also suggest they sustain collusion by punishing competitors for deviations.