Studying Cognition via Information Search in Game Experiments

许多关于战略行为的悬而未决的问题涉及战略的成熟度，即玩家将自己的环境作为游戏分析的程度，并考虑到其结构和其他人的激励。复杂性是传统博弈论的行为假设和适应性学习模型之间的主要区别，前者认为博弈论的行为假设是无限的，后者认为博弈论的行为假设是不存在的或严重受限的。这也是主要的学习模式的主要区别，基于信念的学习模式和强化学习模式。即使在确保收敛到均衡的情况下，复杂性也会通过玩家的初始信念和他们学习规则的结构来影响有限的结果。复杂性对收敛、均衡选择以及对环境变化的反应的性质和时机的影响使其在将博弈论应用于经济、政治和社会互动方面发挥主导作用。这个项目涉及与英国约克大学教研室副教授米格尔·科斯塔-戈麦斯设计和进行的实验。实验允许测试游戏中的行为理论所需的控制，这些理论对环境细节很敏感。实验还允许我们通过监控受试者对隐藏信息的搜索，更直接地研究认知的复杂性和其他方面。在关于认知和信息搜索的看似合理的假设下，最初决定和学习的主要理论对搜索的影响截然不同。根据另一种行为理论的认知含义来解释受试者的搜索有助于理解他们的战略思维并预测他们的行为。一些实验将用于研究受试者对两人矩阵游戏的初始反应的MouseLab设计应用于双人猜谜游戏，在双人猜谜游戏中，受试者具有隐藏的、独立可变的收益参数，监控他们对自己和伴侣的参数的搜索。这种新的设计更好地分离了决策规则；在已知结构中搜索收益参数与在矩阵游戏中直接搜索收益是一项非常不同的认知任务，这扩大了我们对受试者战略思维的视野。另一组MouseLab实验采用我们之前的设计，研究具有多个均衡的两人矩阵博弈与其他旨在分离有限理性决策规则的博弈的初始反应。其主要目标是将传统的均衡选择理论与这样一种观点分开，即对具有多个均衡的博弈的反应是相同类型的有限理性规则的副产品，这些规则似乎在其他博弈中占主导地位。传统的均衡选择理论将风险或收益主导等概念应用于均衡集合。第三组实验监控受试者对隐藏的收益参数的搜索，以及在重复玩各种结构的正常形式游戏时的决策和实现收益的历史。我们的目标是利用信息需求的明显分离来区分强化学习和基于信念的学习，以及经验加权吸引学习等混合学习。这些实验需要开发一种具有扩展搜索功能的新界面。

Many unresolved questions about strategic behavior concern strategic sophistication, the extent to which players analyze their environment as a game, taking its structure and others' incentives into account. Sophistication is the main difference between the behavioral assumptions of traditional game theory, which take it to be unlimited, and adaptive learning models, which take it to be nonexistent or severely limited. It is also the main difference between the leading kinds of learning models, beliefs-based and reinforcement. Even when convergence to equilibrium is assured, sophistication can affect limiting outcomes through players' initial beliefs and the structure of their learning rules. The influence of sophistication on convergence, equilibrium selection, and the nature and timing of responses to changes in the environment give it a leading role in applications of game theory to economic, political, and social interactions.This project concerns experiments designed and conducted with Faculty Associate Miguel Costa-Gomes of the University of York, UK. Experiments allow the control needed to test theories of behavior in games, which are sensitive to environmental details. Experiments also allow us to study sophistication and other aspects of cognition more directly, by monitoring subjects' searches for hidden information. Under plausible assumptions about cognition and information search, leading theories of initial decisions and of learning both have sharply separated implications for search. Interpreting subjects' searches in the light of the cognitive implications of alternative theories of behavior help to understand their strategic thinking and predict their behavior.Some experiments adapt the MouseLab design used to study subjects' initial responses to two-person matrix games to two-person guessing games in which subjects have hidden, independently variable payoff parameters, monitoring their searches for their own and their partners' parameters. This new design better separates decision rules; and searching for payoff parameters within a known structure is a very different cognitive task than searching directly for payoffs in matrix games, which expands our view of subjects' strategic thinking. Another set of MouseLab experiments adapts our previous design to study initial responses to two-person matrix games with multiple equilibria mixed with other games designed to separate boundedly rational decision rules. The main goal is to separate traditional theories of equilibrium selection that apply notions like risk- or payoff-dominance to the set of equilibria from the view that responses to games with multiple equilibria are a by-product of the same kinds of boundedly rational rules that appear to predominate in other games. A third set of experiments monitor subjects' searches for hidden payoff parameters and the history of decisions and realized payoffs in repeated play of normal-form games with a variety of structures. The goal is to exploit sharp separation of information requirements to distinguish reinforcement and beliefs-based learning and hybrids such as experience-weighted attraction learning. These experiments require the development of a new interface with expanded search capabilities.