RI: Small: Speedup Learning for Online Planning Under Uncertainty

Oregon State University, Corvallis OR

Investigators

Abstract

Many complex stochastic planning domains such as logistics, emergency response, resilient power grids, and robotics require the ability to make high-quality decisions under tight time constraints. This project addresses the need for high-quality, but computationally efficient, decision making via new theory and algorithms for speedup learning, which will enable planners to learn to speedup their performance based on prior planning experience. This speedup-learning approach is loosely inspired by the fact that humans routinely learn to speedup their reasoning processes with experience, without sacrificing decision quality. Similarly, through speedup learning, an inefficient planner that produces high-quality decisions will be transformed into a much faster planner with little loss in decision quality. The project involves advancing speedup learning for online planning under uncertainty on four fronts. First, the speedup-learning problem is formalized by introducing the canonical problem of Primitive Speedup Learning (PSL) and studying how PSL can be used to solve various speedup objectives. Second, a novel online planning framework, which subsumes many existing frameworks and enables many potential speedup opportunities, is being designed and developed. Third, the project is producing new speedup learning algorithms for the new framework, which learn various types of knowledge and that can exploit deep neural network (DNN) techniques. Finally, the research is producing extensive empirical evaluations including applications to the important problems of power grid control, municipal emergency response, and benchmark planning domains. The project has the potential for significant broader impact on applications where time-sensitive decisions must be made within stochastic environments. It will directly contribute to advances in two applications in particular: remedial action control in electrical grids to minimize cascading power outages, and planning for municipal emergencies such as fire and rescue operations in cities. The project will also serve to advance graduate education through research assistantships and undergraduate education through summer and academic term research experiences for undergraduates. A special topics graduate course will be taught on the area of planning and learning at Oregon State University and all course materials will be open access.