Tractable Reasoning

Despite the tremendous progress in developing solution techniques for dealing with large constraint satisfaction problems, there will always be problems that are too large even for these very powerful techniques. In such cases, it is necessary to apply methods that are guaranteed to be tractable, although perhaps not necessarily logically sound or complete.

CIRL

We have been exploring ways of effectively dealing with such problems using knowledge compilation, approximation, and context-limited reasoning.

We have also been working for some time to semantically characterize a large class of inferences that can be done quickly. Our goal is to use this information to develop ``anytime'' approximate reasoning systems that are guaranteed to perform a broad class of inferences quickly, and which can then do more complicated reasoning as time allows. Our framework supports a kind of iterative refinement of the models: a system can start with a partial model and gradually make it more concrete as the reasoning process continues, and it will never preclude possibilities that actually remain open by doing so. The reasoning systems characterized by our semantics are tractable, and can be incrementally extended to more complete reasoners without sacrificing polytime behaviour.

Current work involves extending the framework to provide a tractable limited inference mechanism for full first-order logic, combining the ideas from the propositional case with our ideas on context-limited reasoning.

Pointers

A Non-Deterministic Semantics for Tractable Inference
Characterizes a family of tractable fragments of propositional logic using a non-deterministic 3-valued semantics. Written by James Crawford and David Etherington; appeared in Proc. AAAI in 1998. Compressed postscript document.

Parent areas:
Reasoning

Subareas:
Knowledge Compilation
Contexts

Back to the CIRL Overview Page.