Abstract
This paper describes an architecture that combines the complementary strengths of declarative programming, probabilistic graphical models, and reinforcement
learning. Reasoning with different descriptions of incomplete domain knowledge and uncertainty is based on tightly-coupled representations at two different resolutions. For any given goal, non-monotonic logical inference with the
coarse-resolution domain representation provides a plan of abstract actions. Each abstract action is implemented as a sequence of concrete actions by
reasoning probabilistically over a relevant part of the fine-resolution representation, committing high probability beliefs to the coarse-resolution representation. Unexplained plan step failures trigger relational reinforcement learning for incremental and interactive discovery of domain axioms. These capabilities are illustrated in simulated domains and on a physical robot in an indoor domain.
learning. Reasoning with different descriptions of incomplete domain knowledge and uncertainty is based on tightly-coupled representations at two different resolutions. For any given goal, non-monotonic logical inference with the
coarse-resolution domain representation provides a plan of abstract actions. Each abstract action is implemented as a sequence of concrete actions by
reasoning probabilistically over a relevant part of the fine-resolution representation, committing high probability beliefs to the coarse-resolution representation. Unexplained plan step failures trigger relational reinforcement learning for incremental and interactive discovery of domain axioms. These capabilities are illustrated in simulated domains and on a physical robot in an indoor domain.
| Original language | English |
|---|---|
| Title of host publication | Proceedings of the IJCAI 2016 Workshop - Closing the Cognitive Loop: 3rd Workshop on Knowledge, Data, and Systems for Cognitive Computing |
| Publication status | Published - 11 Jul 2016 |