INSPIRE Track 1: Human reasoning and learning in a complex but tractable decision-making paradigm

INSPIRE奖的部分资金来自计算机和信息科学与工程局信息和智能系统部的稳健智能计划以及社会、行为和经济科学局行为和认知科学部的感知、行动和认知计划。该项目研究人类智能的一个标志，即超前思考的能力。在商业谈判、军事战略和教学等不同领域，预测自己和他人行为的后果至关重要。在这些领域中的每一个领域，一个人的决策质量取决于一个人对事件序列的心理模拟的质量，这可能受到认知能力的限制，或者一个人对决策空间复杂性的把握，或者两者兼而有之。该项目的目标是找出影响人们提前思考的表现的因素，并调查通过培训可以在多大程度上提高这种表现。该项目结合了心理学和行为经济学中的启发式(决策者使用的一般规则)的研究。超前思考很难在现实世界的问题中衡量和建模。因此，研究人员开发了一个两人战略决策任务，作为一个可控的实验环境。参与者轮流将代币放在一块4x11的黑板上，并尝试连续放置四个自己的代币。这些规则对受试者来说并不熟悉，但很容易学习。此任务的状态空间大小约为10^20，比国际象棋的状态空间(~10^47)小得多，但具有可观的复杂性，而且太大，人类无法轻松掌握。研究人员使用一种改进的阿尔法-贝塔修剪法“微弱地解决了”这项任务。它很可能也可以被很强地解决，这意味着一个人可以在任何给定的位置确定任何给定的决定是否是错误的。人类数据将通过三种任务模式收集：一种是给受试者一个位置，并必须在一定数量的动作中获胜；人与计算机；以及人与人。研究人员将跟踪受试者的眼球运动，这可能会揭示计划的各个方面，甚至可能有助于可视化心理模拟的过程。该项目的一个重要组成部分将是数据的计算建模。人类不能在任务结束前提前思考，所以我们假设他们使用简单的位置特征(启发式)来评估某些动作的价值。特征的例子可以是连续三个或相邻的开放式两个连续的存在。初步的人类数据表明，不正确的试探法的应用造成了“战略盲区”。研究人员的目标是根据某一特定位置的位置特征以及受试者有限的推理深度来预测该对象做出某一特定动作的概率。由此产生的模型将允许定量地解决这样一个问题，即学习主要是用于增加一个人的推理深度，还是改进一个人的启发式调色板。行为和眼动数据将为研究复杂决策环境中推理的神经基础奠定基础。该项目位于计算机科学、认知心理学、管理和决策科学以及教育的交叉点，并有可能对这些领域做出贡献。从长远来看，该项目可能有助于理解并避免在现实生活中解决问题时“跳出框框”思考的失败。此外，像这个项目中使用的这样的战略任务可以作为测试关于教学方法的假设的微型环境。

This INSPIRE award is partially funded by the Robust Intelligence Program in the Division of Information and Intelligent Systems in the Directorate for Computer and Information Science and Engineering and the Perception, Action, and Cognition Program in the Division of Behavioral and Cognitive Sciences in the Directorate for Social, Behavioral, and Economic Sciences.This project studies a hallmark of human intelligence, namely the ability to think ahead. Anticipating the consequences of one's own actions and those of others is of crucial importance in areas as diverse as business negotiations, military strategy, and teaching. In each of these domains, the quality of one's decisions depends on the quality of one's mental simulations of event sequences, which might be limited by cognitive capacity limitations, one's grasp of the complexities of the decision space, or both. The project's goal is to identify the factors that affect people's performance in thinking ahead, and investigate to what extent this performance can be improved through training. The project ties into the study of heuristics (general rules used by decision-makers) in psychology and behavioral economics.Thinking ahead is difficult to measure and model in real-world problems. Therefore, the investigator has developed a two-person strategic decision-making task as a controllable experimental environment. Participants take turns to put tokens on a 4x11 board and try to get four of their own tokens in a row. The rules are unfamiliar to subjects, yet easy to learn. The size of the state space for this task is of the order of 10^20, much smaller than that of chess (~10^47), yet of appreciable complexity and much too large for humans to easily grasp. The investigators have "weakly solved" this task using an improved version of alpha-beta pruning. It can most likely also be solved strongly, which means that one can determine in any given position whether any given decision is an error. Human data will be collected in three task modes: one in which the subject is given a position and has to win in a set number of moves; human versus computer; and human versus human. The investigators will track subjects' eye movements, which could reveal aspects of planning and perhaps even serve to visualize the process of mental simulation. An important component of the project will be computational modeling of the data. Humans cannot think ahead to the end of the task, so we hypothesize that they use simple features of positions (heuristics) to value certain moves over others. Examples of features could be the presence of a three-in-a-row, or of an adjacent, open-ended two-in-a-row. Preliminary human data suggest "strategic blind spots" created by the application of the incorrect heuristics. The investigators aim to predict the probability that a subject will in a given position make a particular move, based on features of the position that would be created by that move, as well as the subject's limited depth of reasoning. The resulting model will allow to quantitatively address the question of whether learning mostly serves to increase one's depth of reasoning or to refine one's palette of heuristics. The behavioral and eye movement data will lay the foundation for studies of the neural substrates of reasoning in complex decision-making contexts.The project is positioned at the intersection of computer science, cognitive psychology, management and decision science, and education, and has the potential to contribute to each of these fields. In the long run, the project might be able to contribute to understanding and perhaps avoiding failures to think "out of the box" in real-life problem-solving. Moreover, strategic tasks like the one used in this project could serve as a mini-environment for testing hypotheses about teaching methods.