Bayesian Network Models for Pattern and Plan Recognition
Abstract
The aim of this AASERT supplement was to investigate the use of Bayesian networks to capture structural regularities in domains for pattern recognition. We focused on plans as a particular form of pattern, where the recognition elements are actions and their effects. Plans are distinguished by the fact that they are generated by agents to serve some objectives, and this causal relationship can be exploited in developing specific models to support the plan recognition task. This AASERT supplement augmented our AFOSR project that covered more generally the scope of dynamic decision making under uncertainty. It primarily supported the graduate studies of David Pynadath, who successfully completed his dissertation (Pynadath 1999) in February 1999. The specific results of this project comprised several advances in plan recognition under uncertainty, most notably: (1) a general Bayesian framework for plan recognition, (2) a generalization of techniques for probabilistic context-free grammars based on encoding the space of parse trees in Bayesian networks, (3) a new representation, probabilistic state-dependent grammars, exploiting the advantages of state-based and grammatical approaches and (4) demonstrations of the new techniques simplified versions of the application domains of highway traffic and air combat.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 16, 2000
- Accession Number
- ADA379538
Entities
People
- John Rushby
Organizations
- University of Michigan