Bidding Heuristics for Simultaneous Auctions: Lessons from TAC Travel

Abstract

We undertake an experimental study of heuristics designed for the Travel division of the Trading Agent Competition. Our primary goal is to analyze the performance of the sample average approximation (SAA) heuristic, which is approximately optimal in the decision-theoretic (DT) setting, in this game-theoretic (GT) setting. To this end, we conduct experiments in four settings, three DT and one GT. The relevant distinction between the DT and the GT settings is: in the DT settings, agents’ strategies do not affect the distribution of prices. Because of this distinction, the DT experiments are easier to analyze than the GT experiments. Moreover, settings with normally distributed prices, and controlled noise, are easier to analyze than those with competitive equilibrium prices. In the studied domain, analysis of the DT settings with possibly noisy normally distributed prices informs our analysis of the richer DT and GT settings with competitive equilibrium prices. In future work, we plan to investigate whether this experimental methodology—namely, transferring knowledge gained in a DT setting with noisy signals to a GT setting—can be applied to analyze heuristics for playing other complex games.