Rule Learning Across Dissimilar Normal-Form Games

该奖项资助了关于人们如何在战略情境中学习的实验室实验。大多数经济和社会环境的特点是代理人之间的重复互动。长期以来，学习模型一直在寻找代理如何在这种重复交互中学习的模型。最常见的学习过程是将行动映射到收益并使用每一组新证据更新这种映射的过程。虽然这种行为学习在广泛的情况下都很有用，但人类也倾向于表现出认知学习--学习如何对游戏进行推理，并预测其他玩家的行动和其他发展。行为方法和认知方法之间的一个主要区别是，在前者中，除非动作被标记为相同，否则游戏之间的学习是不可能的，而在后者中，游戏之间的学习是可能的，并且可以用适当的框架来表征和理解。在这方面，Stahl(1996)的规则学习框架是理想的。规则学习框架假定个人应用规则，并在规则和收益之间形成映射，而不是行动和收益。然后，强化发生在规则之上，个人会加班加点地学习，放弃历史上落后的规则，转而支持历史上成功的规则。PI团队已经证明了规则学习可以适应这种行为动态，他们现在将进行实验来测试这些预测。理解消费、投资或管理规则，并预测哪些规则将继续存在，哪些规则将消亡，这是经济学家的一项重要任务，也需要理解学习如何在不同情况下转移。本框架是朝着这一方向迈出的重要一步。

This award funds laboratory experiments about how people learn in strategic situations. Most economic and social settings are characterized by repeated interaction between agents. Learning models have long sought to model how agents learn in such repeated interactions. The learning process most commonly assumed is one that maps actions to payoffs and updates this mapping with each new set of evidence. While such behavioral learning is useful in a broad set of situations, humans also tend to display cognitive learning-- learning how to reason about a game and to anticipate other players' actions and other developments. One major difference between the behavioral and cognitive approaches is that in the former, learning between games is not possible unless the actions are labeled the same, whereas in the latter, learning between games is possibleand can be characterized and understood with the appropriate framework. The rule learning framework of Stahl (1996) is ideal in this regard. The rule-learning frameworkposits individuals who apply rules and form mappings between rules and payoffs ratherthan actions and payoffs. Then reinforcement occurs over rules and individuals learn overtime to abandon historically inferior rules in favor of historically successful rules. The PI team has demonstrated that Rule Learning can accommodate such behavioral dynamics, and they will now conduct experiments to test those predictions. Understanding consumption, investment or managerial rules and predicting which will rules will survive and which will perish is an important task for economists and one which is requires an understanding of how learning transfers between situations. The present framework is an important step in this direction.