Simulation Experiments with Goal-Seeking Adaptive Elements.
Abstract
This report describes results obtained from computer simulation experiments designed to systematically develop and evaluate an approach to learning by networks of neuronlike adaptive elements. The characterizing feature of this approach is that the network components, or adaptive elements, are self-interested, goal-seeking agents that implement robust algorithms for furthering their individual interests. These algorithms combine stochastic search and associative learning methods. Results show how these adaptive elements can cooperate as components of layered networks that adaptively create new features by combining existing features. Other simulation experiments are described that systematically examine the performance of individual adaptive elements in tasks that resemble the tasks faced by elements embedded within networks. The performances of a variety of algorithms are compared and contrasted. These results demonstrate the shortcomings of certain algorithms and the effectiveness of others. Experiments examine the effects on learning of delayed reinforcement. We justify the adaptive heuristic critic algorithm for reducing the severity of the temporal credit-assignment problem for tasks with delayed reinforcement. We illustrate the effectiveness of this algorithm in a task requiring the system to learn to control an unstable dynamical system under the influence of low-quality evaluative feedback.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 01, 1984
- Accession Number
- ADA140295
Entities
People
- A. G. Barto
Organizations
- University of Massachusetts Amherst