ICES: Small: Artificial Human Agents for Virtual Economies

该项目的目标是能够在研究两人游戏时用人工代理取代人类代理。这个项目如果成功，将通过用模拟部分取代实验室实验，大大提高经济学家检验经济理论的能力。在过去的几十年里，实验室研究证明了经济理论的有效性（和无效性），以及在实际实施之前在实验室中测试机制（如拍卖）的实际目的。用人工代理人来代替实验室里的真人实验的能力，既可以减少验证和测试的成本，也可以迅速探索更广泛的理论和政策选择。人工智能还增强了我们对人类行为的理解，丰富了我们对人类和人工智能之间联系的认识。基于主体的建模是一种新兴的、有吸引力的验证经济理论的方法。现有的研究主要集中在简单和幼稚的代理。该项目提出的下一步是开发能够模仿人类实验室受试者在两个玩家的随机移动游戏中的行为的部分人类（经济）代理。在这些条件下，人类玩耍的大量和详细的数据是可用的。现有的算法只是稍微缺乏模仿人类游戏的能力，但还没有实现完全自主的代理。基于隐马尔可夫模型，PI提出发展和研究一个信念学习的框架，这个框架足够广泛，可以包含许多现有的学习算法，包括强化学习、虚拟游戏和平滑虚拟游戏。此外，PI还提出在这个框架的基础上衍生出更复杂的学习方法，以充分开发人工智能体。本研究将追求两个重要方向。首先，PI将介绍基于可用信息和认知层次模型的先验的初始校准。第二，当“意外”发生时，PI将允许重新考虑现有模型。该项目将导致经济学和计算机科学研究的下一阶段，扩大人工代理的类别，以攻击更广泛和更经济重要的任务。这也将丰富人工智能中图模型学习的研究。

The goal of the project is to be able to replace human agents withartificial agents in studying two-player games. This project, ifsuccessful, will greatly enhance the ability of economists to testeconomic theories by partially replacing laboratory experiments withsimulations. Over the last decades laboratory studies have proveninvaluable both for the validation (and invalidation) of economictheories, and for the practical purpose of testing mechanisms (such asauctions) in the laboratory prior to practical implementation. Theability to use simulations with artificial agents in place oflaboratory experiments with live human beings will both reduce thecost of validation and testing, and make it possible to explorequickly a much wider range of theories and policy alternatives. Itwill also enhance our understanding of human behavior and enrich ourknowledge of the connection between human and artificial intelligence.Agent-based modeling is an emerging and attractive approach tovalidating economic theories. Existing research has focused on simpleand naive agents. This project proposes as the next step to developartificial human (economic) agents capable of mimicking the behaviorof human laboratory subjects in the context of two player simultaneousmove games. Substantial and detailed data is available on human playunder these conditions. Existing algorithms fall only slightly shortof the ability to mimic human play but do not yet implement fullyautonomous agents. Based on hidden Markov models, the PIs propose todevelop and investigate a framework of belief learning that is broadenough to encompass many existing learning algorithms includingreinforcement learning, fictitious play, and smooth fictitious play.Moreover, based on this framework, the PIs propose to derive moresophisticated learning methods to fully develop artificial agents. Thisresearch will pursue two important directions. First, the PIs willintroduce the initial calibration of priors based on availableinformation and a cognitive hierarchy model. Second, the PIs will allow forthe reconsideration of the existing model when "surprises" occur. Theproject will lead to the next stage of research in both economics andcomputer science in broadening the class of artificial agents toattack broader and more economically important tasks. It will alsoenrich the research on graphical model learning for artificialintelligence.