Goal Seeking Components for Adaptive Intelligence: An Initial Assessment.
Abstract
This report assesses the promise of a network approach to adaptive problem solving in which the network components themselves possess considerable adaptive power. We show that components designed with attention to the temporal aspects of reinforcement learning can acquire knowledge about feedback pathways in which they are embedded and can use this knowledge to seek their preferred inputs, thus combining pattern recognition, search, and control functions. A review of adaptive network research shows that networks of components having these capabilities have not been studied previously. We demonstrate that simple networks of these elements can solve types of problems that are beyond the capabilities of networks studied in the past. An associative memory is presented that retains the generalization capabilities and noise resistance of associative memories previously studied but does not require a 'teacher' to provide the desired associations. It conducts active, closed-loop searches for the most rewarding associations. We provide an example in whcih these searches are conducted through the system's external environment and an example in which they are conducted through an internal predictive model of that environment. The latter system is capable of a simple form of latent learning.
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 01, 1981
- Accession Number
- ADA101476
Entities
People
- Andrew G. Barto
- Richard S. Sutton
Organizations
- University of Massachusetts Amherst