Models of information search: A theoretical and empirical synthesis

寻找相关信息是认知和决策的核心。医生在诊断病人时不能进行所有可用的医学测试。一个科学家不可能进行所有可能相关的实验。我们如何量化不同问题（测试，实验）的有用性，并确定哪些问题选择策略是最好的，给定一个特定的任务？根据人们的目标和对环境结构的信念，人们使用什么策略来选择问题？哲学家、心理学家和统计学家都在这些领域工作，往往没有意识到邻近学科的相关研究。我们的项目通过连接理论上密切相关但很少被一起考虑的三类模型来解决这些问题：i）概率确认措施，如认识论和科学哲学中所讨论的; ii）来自统计学的信息价值模型; iii）来自心理学的有限理性启发式模型。概率确认测量量化了所获得的数据（例如医学测试结果）为特定假设提供的支持量（即，患者患有某种疾病）。信息价值模型，如信息增益或概率增益，量化了测试结果的整体有用性，或实验的预期有用性，在所有考虑的假设。问题选择的启发式模型指定如何选择特定的测试，但可能不会尝试精确量化每个可用测试的值。我们将使用不同的方法，包括数学分析，计算机模拟和行为实验，以探索这些领域是如何相关的。我们将展示确认措施和信息价值模型如何从一个统一的数学框架中产生，以及一些启发式模型如何作为信息价值模型的特殊情况出现。我们将重点讨论使用不同方法将确认汇总为信息价值的含义。计算机模拟将调查许多启发式和信息价值策略的行为，为不同策略在不同环境中如何相互关联提供综合视角。行为实验将确定哪些策略最能描述儿童和成人对问题的选择。教育研究将测试哪些顺序搜索游戏可以最有效地培养儿童对概率和信息的直觉。我们的总体目标是把确认和信息在一个统一的框架，可以指导哲学和心理学研究，同时有助于信息获取的心理和规范接地理论。

Searching for relevant information is central to cognition and decision making. A doctor cannot conduct every available medical test when diagnosing a patient. A scientist cannot conduct all potentially relevant experiments. How can we quantify the usefulness of different questions (tests, experiments) and determine which question-selection strategies are best, given a particular task? Which strategies do people use to select questions, given their goals and beliefs about the structure of the environment? Philosophers, psychologists, and statisticians have all been working in these areas, often without awareness of related research in neighboring disciplines. Our project addresses these issues by connecting three classes of models that are closely related theoretically, but that have seldom been considered together: i) probabilistic confirmation measures, as discussed in epistemology and philosophy of science; ii) value-of-information models, from statistics; and iii) boundedly rational heuristic models, from psychology. Probabilistic confirmation measures quantify the amount of support that an obtained datum (such as a medical test result) provides for a particular hypothesis (i.e., that the patient has a particular disease). Value-of-information models, like information gain or probability gain, quantify the overall usefulness of a test result, or the expected usefulness of an experiment, over all considered hypotheses. Heuristic models for question selection specify how a particular test is chosen, but may not attempt to precisely quantify each available test's value. We will use diverse methodologies including mathematical analyses, computer simulations, and behavioral experiments to explore how these areas are related. We will show how confirmation measures and value-of-information models arise from a unified mathematical framework, and how a number of heuristic models arise as special cases of value-of-information models. We will focus on the implications of using different methods to aggregate confirmation into the value of information. Computer simulations will investigate the behavior of many heuristic and value-of-information strategies, providing an integrative perspective on how different strategies relate in different environments. Behavioral experiments will identify which strategies best describe children and adults' choices of questions to ask. Educational studies will test which sequential-search games can most effectively develop children's intuitions about probability and information. Our overarching goal is to bring confirmation and information together in a unified framework that can guide philosophical and psychological research, while contributing to a psychologically and normatively grounded theory of information acquisition.