RI: Small: Plan Execution for Continuous Dynamical Systems Within Risk Bounds

在许多应用中，从海底到太空，自治代理被赋予一组最佳实现的操作目标，同时考虑到不可控事件引起的不确定性。例如，在对自主水下航行器（AUV）的监测使命中，目标可能是在避免危险的同时最大限度地提高科学回报。由于不确定性，对于许多现实世界的使命来说，保证100%的成功是不现实的。在决策理论规划界广泛探讨的一种方法是最大化一个以风险换取效用的目标。但是，这并不能提供任何硬性保证。工程实践中常用的另一种方法是将风险指定为硬约束，即使命失败的上限（称为机会约束）。例如，NASA的火星任务旨在满足或超过成功着陆概率的要求;人类额定车辆的设计要求类似。给定一个机会约束的使命，执行该使命的智能体可以努力最大化预期奖励，同时确保满足机会约束和其他操作约束。 这项研究正在开发一个基于模型的执行，实现特定的风险范围内的目标水平的计划，同时试图最大化预期的回报。该执行程序的关键属性包括：1）用户指定的时间演变的目标行为; 2）通过生成一系列离散和连续的行动来控制工厂的计划执行;以及3）在风险范围和动态约束内的控制行动的最佳随机规划。这项赠款下的研究正在为长期目标奠定基础，即随着计划执行期间不确定因素的解决，促进对最初计划的不断调整和风险分配。在其他应用中，我们计划与参与科学任务的AUV进行测试，测量与障碍物相关的时间和能量范围内的性能。

In many applications, from undersea to space, an autonomous agent is given a set of operational goals to achieve optimally, while taking into account the uncertainties that arise from uncontrollable events. For example, in a monitoring mission for an autonomous under-water vehicle (AUV), the goal might be to maximize scientific return, while avoiding hazards. Due to uncertainty, it is unrealistic for many real-world mission to guarantee 100% success. One approach explored extensively within the decision-theoretic planning community is to maximize an objective that trades risk for utility. However, this does not provide any hard guarantees. An alternative approach, commonly employed in engineering practice is to specify risk as a hard constraint, in terms of an upper bound on mission failure (called a chance constraint). For example, NASA Mars missions are designed to meet or exceed a requirement on the probability of successful landing; human-rated vehicles are designed to similar requirements. Given a chance-constrained mission, an agent performing the mission may strive to maximize expected reward, while ensuring that the chance constraint and other operating constraints are met. This research is developing a model-based executive that achieves goal-level plans within specified risk bounds, while attempting to maximize expected reward. Key attributes of this executive include: 1) user specification of time-evolved goal behaviors; 2) plan execution by generating a sequence of discrete and continuous actions for controlling the plant; and 3) optimal, stochastic planning of control actions within risk bounds and dynamic constraints. The research under this grant is laying the groundwork for longer-term objectives of facilitating continuous adaptation of the initial plan and allocation of risk as uncertainties are resolved during plan execution. Among other applications, we plan tests with AUVs engaged in scientific missions, measuring performance within obstacle-related, time and energy bounds.