Mastersthesis,
Games and Learning in Auctions
Maastricht University, (2008)
Abstract
Auctions are pervasive in today���s society and provide a variety of markets, ranging from government-to-business auctions for licenses to consumer-to- consumer online auctions. The success of trading strategies in auctions is highly dependent on the present competitors, hence traders are forced to adapt to the competition to maintain a high level of performance. This adaptation may be modeled by reinforcement learning algorithms, which have a proven relation to evolutionary game theory. This thesis facilitates a strategic choice between a set of predefined trad- ing strategies. It is based on previous work, which suggests to capture the payoff of trading strategies in a heuristic payoff table. A new methodology to approximate heuristic payoff tables by normal form games is introduced, and it is evaluated by a case study of a 6-agent clearing house auction. Learning models of exploration and exploitation, that link to selection and mutation in an evolutionary perspective, are subsequently applied to compare three common automated trading strategies. The information loss in the normal form approximation is shown to be reasonably small, such that the concise normal form representation can be used to derive strategic decisions in auctions. Furthermore, the learning model shows that learners with exploration may converge to different strate- gies than learners of pure exploitation. The devised methodology establishes a bridge between empirical data in heuristic payoff tables and the means from classical game theory. It might therefore become the basis for a more general framework to analyze strategic interactions in complex multi-agent systems.
