Prior knowledge elicitation and policy explanation for decision-theoretic planning and learning

考虑一下口语对话管理器、移动机器人控制器、老年痴呆症患者的自动监控/提示系统，或者任何其他必须完成相当复杂任务的复杂系统。由于传感器的噪声性质(例如，噪声语音识别、噪声声纳)以及系统动作的不确定和相互依赖的影响(例如，提示对老年人的不确定影响、机器人中的相互依赖和噪声的马达控制)，这种系统的概念特别具有挑战性。因此，通过手工编码控制策略来设计复杂的鲁棒系统通常是不可能的。决策理论规划和学习领域在自动化技术的发展方面取得了重大进展，以产生可能彻底改变下一代计算机系统的稳健的控制策略。不是直接对策略进行编程，而是使用算法基于系统及其环境的模型或模拟器来优化策略。然而，获取指定模型或模拟器所需的领域知识，以及验证/解释产生的策略是研究界忽视的两个主要瓶颈，它们阻碍了这种颠覆性技术的采用。知识获取和政策解释尤其具有挑战性，因为非技术领域的专家往往具有不全面和不精确的知识，往往需要对政策进行高层次的解释，其中技术细节被抽象出来，以更好地传达直觉。因此，这项研究的目标是：i)设计通用的原则性技术，以获取和编码有关系统、环境和所需策略的部分/不精确的领域知识；ii)开发能够利用尽可能多的领域知识来提高可伸缩性的算法；以及iii)创建通用工具，以验证和解释开发人员和非技术专家在适当级别做出的策略决策。

Consider spoken-dialogue managers, mobile robot controllers, automated monitoring/prompting systems for seniors with dementia or any other complex system that must accomplish a fairly complicated task. The conception of such systems is particularly challenging due to the noisy nature of the sensors (e.g., noisy speech recognition, noisy sonars) as well as the uncertain and interdependent effects of system actions (e.g., uncertain effect of prompts on seniors, interdependent and noisy motor controls in robotics). As a result, it is generally impossible to design complex robust systems by hand coding control policies. The fields of decision-theoretic planning and learning have made significant advances in the development of automated techniques to generate robust control policies that could revolutionize the next generation of computer systems. Instead of programming a policy directly, an algorithm is used to optimize a policy based on a model or simulator of the system and its environment. However, eliciting the domain knowledge necessary to specify a model or simulator, and validating/explaining the resulting policy are two major bottlenecks ignored by the research community that are holding back the adoption of this disruptive technology. Knowledge elicitation and policy explanation are particularly challenging since non-technical domain experts tend to have partial and imprecise knowledge, and often need high-level explanations of the policy where technical details are abstracted away to better convey the intuition. Hence, the objectives of this research are i) to design general and principled techniques to elicit and encode partial/imprecise domain knowledge about the system, the environment and the desired policy, ii) to develop algorithms that can exploit as much domain knowledge as possible to improve scalability, and iii) to create generic tools to validate and explain the decisions made by a policy at an appropriate level for developers and non-technical experts.