RI: Small: Speedup Learning for Online Planning Under Uncertainty

许多复杂的随机规划领域，如物流、应急响应、弹性电网和机器人，都需要在严格的时间限制下做出高质量决策的能力。该项目通过加速学习的新理论和算法解决了对高质量、但计算效率高的决策制定的需求，这将使规划者能够根据先前的规划经验学习加速他们的绩效。这种加速学习方法的灵感来自于这样一个事实，即人类通常会通过经验来学习加速他们的推理过程，而不会牺牲决策质量。同样，通过加速学习，一个低效的计划者产生高质量的决策，将转变为一个更快的计划者，决策质量几乎没有损失。该项目涉及在四个方面的不确定性下推进在线规划的加速学习。首先，通过引入原始加速学习（PSL）的规范问题来形式化加速学习问题，并研究如何使用PSL来解决各种加速目标。第二，正在设计和开发一种新的在线规划框架，它包含了许多现有的框架，并实现了许多潜在的加速机会。第三，该项目正在为新框架生成新的加速学习算法，这些算法可以学习各种类型的知识，并可以利用深度神经网络（DNN）技术。最后，该研究正在产生广泛的经验评估，包括在电网控制、市政应急响应和基准规划领域的重要问题上的应用。该项目有可能对必须在随机环境中做出时间敏感决策的应用产生重大而广泛的影响。它将直接促进两个应用领域的进步：电网的补救行动控制，以最大限度地减少级联停电，以及城市消防和救援行动等市政紧急情况的规划。该项目还将通过研究助学金促进研究生教育，并通过夏季和学期的本科生研究经验促进本科教育。专题研究生课程将在俄勒冈州立大学的规划和学习领域教授，所有课程材料将开放获取。

Many complex stochastic planning domains such as logistics,emergency response, resilient power grids, and robotics require theability to make high-quality decisions under tight timeconstraints. This project addresses the need for high-quality, butcomputationally efficient, decision making via new theory andalgorithms for speedup learning, which will enable planners tolearn to speedup their performance based on prior planningexperience. This speedup-learning approach is loosely inspired bythe fact that humans routinely learn to speedup their reasoningprocesses with experience, without sacrificing decision quality.Similarly, through speedup learning, an inefficient planner thatproduces high-quality decisions will be transformed into a muchfaster planner with little loss in decision quality.The project involves advancing speedup learning for online planningunder uncertainty on four fronts. First, the speedup-learningproblem is formalized by introducing the canonical problem ofPrimitive Speedup Learning (PSL) and studying how PSL can be usedto solve various speedup objectives. Second, a novel onlineplanning framework, which subsumes many existing frameworks andenables many potential speedup opportunities, is being designed anddeveloped. Third, the project is producing new speedup learningalgorithms for the new framework, which learn various types ofknowledge and that can exploit deep neural network (DNN)techniques. Finally, the research is producing extensive empiricalevaluations including applications to the important problems ofpower grid control, municipal emergency response, and benchmarkplanning domains. The project has the potential for significant broader impact onapplications where time-sensitive decisions must be made withinstochastic environments. It will directly contribute to advances intwo applications in particular: remedial action control inelectrical grids to minimize cascading power outages, and planningfor municipal emergencies such as fire and rescue operations incities. The project will also serve to advance graduate educationthrough research assistantships and undergraduate education throughsummer and academic term research experiences for undergraduates. Aspecial topics graduate course will be taught on the area ofplanning and learning at Oregon State University and all coursematerials will be open access.