RI: Small: Towards a Formal Theory of Blameworthiness, Intention, and Moral Responsibility

随着我们进入一个自动驾驶汽车、自动无人机和机器人帮助人们完成工作和日常任务的时代，越来越需要发展一种道德或伦理行为理论。 政策制定者已经在努力解决其中的一些问题。 世界各地的司法管辖区都开始制定考虑到无人驾驶汽车存在的公路法规。 这个项目的目标是提供一个理论的责备，意图和道德责任，这样的代理人和应用程序的基础。 这项研究将根据人们做出的道德判断来评估这一理论，并采取第一步将由此产生的政策应用于实践。 一个好的理论必须考虑到许多相互竞争、有时是相互冲突的目标之间的权衡。 这表明，概率和效用将涉及。 该理论还必须处理反事实问题，也就是说，推理如果情况或行动不同会发生什么。 因果关系的问题将揭示代理人的行为是否对结果负责，或者该行为是否与结果巧合。 另一个相关问题是确定意图：即确定行为人在实施行为时是否有意取得某种结果。珀尔的结构方程框架可以模拟反事实（干预），并已被用作因果关系定义的基础。 在这个项目中，该框架将被扩展到捕获的程度的责任和意图。 这些定义是相对于一个代理的认知状态，它描述了代理的信念和偏好。 最初，信念将由概率度量指定，偏好由效用函数指定。 然后，该框架将扩展到考虑信念和偏好的更多定性表示，以考虑正常性和典型性问题（例如，以捕获社会惯例），以及具有多个代理的设置（例如，多个通信的无人驾驶汽车或机器人），其中出现了博弈论的关注和群体责任与个人责任的问题。 将对在这一框架下制定的定义进行测试，以了解它们是否抓住了人们对这些问题的直觉，并将其应用于实际感兴趣的情况。 然后将对理论进行改进，以考虑这些实验的结果。

As we move to an era of self-driving cars, autonomous drones, and robots helping people in work and everyday tasks, there is an increasing need to develop a theory of moral or ethical behavior. Policy-makers are already grappling with some of the issues. Jurisdictions around the world are beginning to formulate highway codes that take into consideration the presence of driverless cars. The goal of this project is to provide the foundations for a theory of blameworthiness, intention, and moral responsibility for such agents and applications. The study will evaluate this theory against the moral judgments that people make, and take the first steps to applying resulting policies in practice. A good theory will have to take into consideration tradeoffs among many competing and sometimes conflicting goals. This suggests that probability and utility will be involved. The theory will also have to deal with counterfactuals--that is, reasoning about what would have happened had circumstances or actions differed. Issues of causality will reveal whether an agent's action was responsible for an outcome, or if the action was coincidental to the outcome. Yet another relevant issue is determining intent: that is, determining whether an agent intended an outcome when performing an act. Pearl's structural-equations framework can model counterfactuals (interventions) and has been used as the basis of a definition of causality. In this project, the framework will be extended to capture degree of blameworthiness and intention. These definitions are given relative to an agent's epistemic state, which describes the agent's beliefs and preferences. Initially, the beliefs will be specified by a probability measure and the preferences by a utility function. The framework will then be extended to consider more qualitative representations of beliefs and preferences, to take into account issues of normality and typicality (e.g., to capture societal conventions), and to settings with multiple agents (e.g., multiple communicating driverless cars or robots), where game-theoretic concerns and issues of group responsibility vs. individual responsibility arise. The definitions developed under this framework will be tested to see if they capture people's intuitions about these issues, and will be applied to situations of practical interest. The theory will then be refined to take into account the outcomes of these experiments.