A Probabilistic Approach to Anytime Algorithm for Intelligent Real-Time Problem Solving
Abstract
Our work on real time intelligent problem solving has focused on the tradeoff between deliberation and activity. Such a tradeoff is required, since an excess of deliberation will be defeated by the dynamical nature of the world and by errors in the predictive model, and a lack of deliberation will not provide the agent with sufficient flexibility to perform well in novel situations. Our framework for evaluating this tradeoff includes both an explicit and an implicit component. In the explicit work, we represent the uncertainties associated with inaccuracies in the model and the inability to completely monitor changes in the world by expanding our language to include probabilities, and making choices about when to act and when to deliberate further based upon these explicit uncertainty measures. In the implicit approach, we use reinforcement learning of a Markov Decision Process to place a strict bound on deliberation. The agent's knowledge is obtained through an active sensory system having limited bandwidth, overcoming the standard limitations of assuming complete knowledge, but requiring modifications to the standard learning algorithm. Learning time is decreased by the use of social learning mechanisms as well as task decomposition and dynamic policy merging.
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 04, 1992
- Accession Number
- ADA255709
Entities
People
- James F. Allen
- Josh Tenenberg
Organizations
- University of Rochester