Algorithmic Prediction of Human Strategic Behaviour

点击翻译按钮获取中文摘要

Settings in which participants reason strategically are widespread in artificial intelligence; they include any system that interacts with a human. These range from systems with large numbers of participants accounting for billions of dollars in transactions, such as ad auctions and radio spectrum allocation, to interactions with just a few participants, such as computer games or intelligent user interfaces. Regardless of the scale of such a system for interacting with humans, alternative designs must be evaluated either by trying each one out--frequently a very expensive and time-consuming proposition--or via their predicted performance based on a model of how participants will react to them. Hence, accurate behavioural models are valuable tools for improving both the design of agents and the design of the systems themselves. One possible candidate for such a model is game theory, the standard mathematical framework for understanding strategic interactions. However, it has been demonstrated from experimental data that standard game theoretic models describe human behaviour poorly. This inadequacy stems from their assumption that participants are perfectly rational agents. However, thanks to algorithmic advances and increased computing power and data availability, we now have the ability to go beyond the classical economic rationality assumptions to accurately predict actual human behaviour. The proposed research will develop substantially stronger models by taking realistic domain-specific problem structure into account. The realism to be considered will include: participants with large action sets; reasoning about large sets of other participants; receiving and learning from feedback as they interact with a system; reasoning about the learning of other participants. I will build on my previous work to construct models that predict behaviour in these richer environments. I will focus on two kinds of strategic interactions. The first is interactions with humans in relatively fixed environments. The specific domains will be allocation of security resources for malware and fraud detection, more realistic play in games, and optimal traffic advice. The other kind of interaction is the design of policies and incentives for coordinating the behaviour of multiple people, with a focus on the design of peer grading incentives that cannot be "gamed". The second problem of designing the strategic environment itself is referred to as mechanism design, and it is an increasingly important area of algorithmic game theory. There is a great opportunity for behaviourally informed mechanism design, as existing work is still very tied to the full rationality model.