Learning to Bid against Humans with Unclear Motives
Abstract: Learning to Bid against Humans with Unclear Motives
Auctions play a central role in corporate procurement and sales.
Game theoretical predictions provide a way for bidders to prepare in such auctions, but laboratory experiments have shown that human subjects often fail to follow equilibrium predictions in auctions.
One reason is that the intricate complexity and heterogeneity of the motivations that drive human behavior do not conform to conventional game-theoretic assumptions based on risk neutrality and symmetry.
For example, in addition to seeking financial gain, bidders avoid regret or they are risk-averse.
The influence of such behavioral biases depends on the auction format and the number of competitors.
Understanding equilibrium behavior with such behavioral motives and asymmetries is challenging analytically.
However, ignoring these asymmetries will lead to suboptimal strategies.
We introduce methods to estimate the behavioral motives of bidders in specific auction markets and derive bidding strategies in such asymmetric environments.
First, we propose a novel method for estimating human utility based on Bayesian optimization and equilibrium learning.
Second, we apply this new estimation technique to study overbidding in first-price winner-pay and all-pay auctions with different levels of competition.
Risk and regret aversion are the most prominent conjectures and we can characterize the level of risk and regret aversion that best characterizes certain market environments.
Based on these estimates, we ask the question of how a payoff-maximizing machine would best respond if it has to play against humans with such behavioral motives.
Such a machine might be an automated bidding agent or a decision support tool for bidders in the auction.
Interestingly, we find that a machine could improve its payoff by picking an appropriate utility function that might differ from pure payoff maximization.
Overall, our new methods allow analyses of strategic interaction in games populated by motivationally complex humans and/or machines with asymmetries that go far beyond the standard symmetric Bayes-Nash equilibrium analysis, which ignores the asymmetries that often matter in the field.