RI: Medium: Hierarchical Decision Making for Physical Agents

点击翻译按钮获取中文摘要

Research under this award addresses a core problem in the development of intelligent systems: the generation of effective, deliberate activity over substantial time scales in complex state spaces. Whereas current methods for plan generation are limited either to very short plans or to very repetitive behavior in highly structured and simplified environments, the investigators are developing a new mathematical framework for hierarchical decision making. The framework, based on so-called "angelic" nondeterministic semantics for high-level actions, underpins new algorithms capable of generating provably optimal high-level plans without expanding those plans into primitive actions. The algorithms therefore satisfy the "downward refinement property," resolving a problem that has been open for over 30 years at the heart of AI planning research. By combining hierarchical deliberation, both offline and online, with hierarchical reinforcement learning and apprenticeship learning for low-level physical skill acquisition, the research seeks to enable dexterous human-scale robots to operate in unstructured environments such as kitchens, offices, and environments of still greater complexity. The research promises to yield a deeper understanding of, and better tools for, large-scale decision making in general, with implications in the social, governmental, corporate, and military spheres. This award is funded under the American Recovery and Reinvestment Act of 2009 (Public Law 111-5).