基于因果构造和推理的专家判断关键技术研究

项目名称： 基于因果构造和推理的专家判断关键技术研究

项目编号： No.71271014

项目类型： 面上项目

立项/批准年度： 2013

项目学科： 管理科学

项目作者： 杨敏

作者单位： 北京航空航天大学

项目金额： 53万元

中文摘要： 不确定性事件概率分布是决策理论和风险分析的核心要素，在可用数据匮乏的情况下，主要通过专家判断来获取。由于受到认知、动机、情绪与社会等多种因素影响，同时也缺乏规范获取技术的支持，目前专家判断获取的质量难于保证。 最近的研究表明，因果构造和推理是专家组合各种线索做出概率判断的重要途径。课题采用基于因果分析的思路，研究1）利用因果贝叶斯网络开发概率判断生成模型，用于解释和消除专家判断中的基率谬误、过度自信和联合谬误等认知偏差；2）基于因果的多信息源集成技术，用于多专家判断，以及与其他数据的综合；3）基于Web文本挖掘和价值决策的专家选择模型，用于选择适合判断任务的专家；4）寻找群体一致判断，用于度量动机偏差；5）利用基于主体的股票价格预测模型进行计算验证，并应用于复杂产品研制风险和试验安全性评估中。 课题将提供理论上严密、应用中可行的专家判断获取关键技术，可应用于所有依赖专家判断的场合。

中文关键词： 专家判断抽取与集成；因果贝叶斯网络；工程风险分析；因果分析；基于主体的建模

英文摘要： The probability distribution of event under uncertainty is essential in decision theory and risk analysis. It is obtained mainly by expert judgment elicitation (EJE) while lacking available data, but the elicitation quality is hard to be guaranteed because of the cognitive, motivational, emotional and social influences, and the lack of formal elicitation techniques. Recent researches show that causality constructing and reasoning is the crucial way for expert to combine the cues to make probability judgment. So，this project employs causality analysis methods to 1) develop generative probability judgment model using causal Bayesian network for explaining and debiasing cognitive biases, such as base rate fallacy, overconfidence and conjunction fallacy; 2) construct causality-based aggregation technique for multiple information sources, e.g. expert judgment and statistical data; 3) build an expert selection model using Web-based text mining methods and value-focused decision-making, in order to choose adequate experts for judgment task; 4) identify consensus judgment for measure motivational biases; 5) conduct computational validation by forecasting stock price using agent-based modeling, and apply EJE techniques to evaluating the risk and the safety during complex product developing and testing. The project will

英文关键词： expert judgement elicitation and aggregation；causal Bayesian network；engineering risk analysis；causality analysis；agent-based modeling