We study planning problems for continuous control systems with uncertainty caused by measurement and process noise. The goal is to find an optimal plan that guarantees that the system reaches a desired goal state within finite time. Measurement noise causes limited observability of system states, and process noise causes uncertainty in the outcome of a given plan. These factors render the problem undecidable in general. Our key contribution is a novel abstraction scheme that employs Kalman filtering as a state estimator to obtain a finite-state model, which we formalize as a Markov decision process (MDP). For this MDP, we employ state-of-the-art model checking techniques to efficiently compute plans that maximize the probability of reaching goal states. Moreover, we account for numerical imprecision in computing the abstraction by extending the MDP with intervals of probabilities as a more robust model. We show the correctness of the abstraction and provide several optimizations that aim to balance the quality of the plan and the scalability of the approach. We demonstrate that our method can handle systems that result in MDPs with thousands of states and millions of transitions.
翻译:我们研究的是持续控制系统的规划问题,由于测量和过程噪音造成不确定性。目标是找到一个最佳计划,保证系统在一定时间内达到理想的目标状态。测量噪音造成系统状态的可观察性有限,而过程噪音则造成特定计划结果的不确定性。这些因素使得问题在总体上无法消除。我们的主要贡献是一个新的抽象计划,利用卡尔曼过滤器作为州测算器获得一个有限状态模型,我们正式确定该模型为马尔科夫决定程序(MDP ) 。对于这个MDP,我们使用最先进的模型检查技术来有效地计算能够最大限度地达到目标状态的计划。此外,我们考虑到在计算抽象时数字不精确,我们延长了MDP的概率间隔,将其作为一个更强有力的模型。我们展示了抽象的正确性,并提供了若干优化,目的是平衡计划质量和方针的可扩展性。我们证明我们的方法可以处理导致MDP的系统以及数千个州和数百万个转型国家。