Research on Bounded Rationality

这项研究开发了模仿人类决策者认知过程的智能算法。这种算法像分配其他稀缺资源一样分配认知资源。注意力只被分配到它能带来最大好处的任务上。在解决高度复杂的问题时，这些算法满足于近似解决方案和“有经验的猜测”。这项研究开发了预测这种有根据的猜测的形式的模型，提供了一种可实现的人工智能模型。为了取得成功，这些模式必须在无所不知的极端理性和机械策略的极端短视之间找到一个合理的中间地带。一台试图考虑一切(无所不能)的智能机器在做出决定时会耗尽时间。相比之下，一台短视的机器会很快犯下人类永远不会犯的明显错误。要在全知和近视的两极之间找到一个中间地带，就需要回答许多关于认知的问题。人们使用哪些信息？这些信息是如何被操纵的？个体如何决定何时停止处理一个复杂的问题，转而根据自己最好的猜测采取行动？解决这些问题的“定向认知模型”被表示为一个三步算法。首先，该算法评估各种认知操作的预期收益。预期的收益与认知操作将揭示有关即将做出的决定的有用信息的预测可能性有关。其次，该算法以最大的预期收益执行认知操作。第三，算法反复循环前两个步骤，当分析的认知成本超过预期收益时停止。定向认知模式实现了三个目标。首先，该模型在心理学上是可信的，它预测了许多观察到的心理现象(例如，突显、近视和锚定)，并与实验对象声称使用的认知策略相匹配。其次，该模型生成了可以进行经验检验的精确的定量预测。最初的实验数据压倒性地拒绝了完全理性模型，转而支持定向认知模型。第三，由于该模型是通用的，所以它可以应用于广泛的一类问题。通过扩展定向认知模型并将其与其他成本效益认知模型相结合，本研究发展了一种包括有界理性动态规划在内的一般有限理性优化方法。这种方法的核心是内生近似理论。目前的应用包括契约理论和消费理论。合同理论的应用解释了合同不完备性的存在和形式，包括样板合同。有限理性消费模型解释了为什么家庭对经济环境的变化适应得太慢，以及为什么家庭同时对当前收入等显著变量表现出过度敏感。定向认知模型准确预测了实验对象在一个由多个部分组成的测验中选择花在每个问题上的时间。该模型还预测了当给定问题的允许时间由实验者固定并随受试者的不同而变化时，受访者的回答质量将如何变化。如果定向认知模型继续得到经验验证，它将代表第一批能够正式预测决策问题难度的经济模型之一-即，该模型预测分析问题所花费的时间与结果决策的最佳性/准确性之间的定量关系。最终，可以开发出一个相对通用的人工智能决策模型，该模型可以在广泛的选择问题中应用和测试。

This research develops intelligent algorithms that mimic the cognitive processes of human decision-makers. Such algorithms allocate cognitive resources like other scarce resources. Attention is only allocated to tasks in which it will do the most good. The algorithms settle for approximate solutions and "educated guesses" when solving highly complex problems. The research develops models that predict the form of such educated guesses, providing an implementable model of artificial intelligence. To succeed such models must find a sensible middle ground between the extreme rationality of omniscience and the extreme myopia of mechanistic strategies. An intelligent machine that tried to think of everything (omniscience) would run out of time when making a decision. By contrast, a machine that acted myopically would quickly blunder into obvious mistakes that humans would never make. Finding a middle ground between the extremes of omniscience and myopia will require answers to numerous questions about cognition. What information do people use? How is that information manipulated? How do individuals decide when to stop working on a complex problem and act on their best guess?The "directed cognition model" that addresses these questions is expressed as a three-step algorithm. First, the algorithm evaluates the expected benefit of various cognitive operations. The expected benefit is related to the predicted likelihood that a cognitive operation will reveal useful information about an upcoming decision. Second, the algorithm executes the cognitive operation with the greatest expected benefit. Third, the algorithm repeatedly cycles through these first two steps, stopping when the cognitive costs of analysis outweigh the expected benefits. The directed cognition model realizes three goals. First, the model is psychologically plausible, predicting numerous observed psychological phenomena (e.g., salience, myopia, and anchoring) and matching the cognitive strategies that experimental subjects claim to use. Second the model generates precise quantitative predictions that can be empirically tested. Initial experimental data overwhelmingly rejects the perfectly rational model in favor of the directed cognition model. Third, because the model is general it can be applied to a wide class of problems. By extending the directed cognition model and integrating it with other models of cost-effective cognition, the research develops a general bounded rationality approach to optimization, including boundedly rational dynamic programming. At the core of this approach is a theory of endogenous approximation. Current applications include contract theory and consumption. Contract theory applications explain both the presence and form of contract incompleteness, including boilerplate contracts. Boundedly rational consumption models explain why households adjust too slowly to changes in their economic environment and why households simultaneously exhibit excessive sensitivity to salient variables like current income.The directed cognition model makes sharp predictions about how much time experimental subjects will choose to spend on each problem in a multi-part quiz. The model also predicts how the quality of respondents' answers will vary when the amount of time allowed for a given problem is fixed by the experimenter and varied across subjects. If the directed cognition model continues to be empirically validated, it will represent one of the first economic models that can formally predict the difficulty of a decision problem --- i.e., the model predicts the quantitative relationship between time spent analyzing a problem and optimality/accuracy of the resulting decision. Ultimately, a relatively general model of artificially intelligent decision-making may be developed, which can be applied and tested in a wide range of choice problems.