Journal article
A trainable controller based on neural network
Neural networks, v 1(suppl), pp 336-336
01 Jan 1988
Abstract
We employ neuromorphic architecture in the design of Trainable Adaptive Controllers (TACs). TACs are controllers that can be designed by training rather than programming, where the trainer can be a human or automated expert. We show that a neural network (NN) can function as a real time, robust adaptive controller. The system can be trained using the responses of a human teacher, or any system that has achieved the desired control performance. Thus, a neurocontroller can be designed even when no explicit form of a control law is known. The TAC architecture consists of the controller the teacher and the controlled process. Our example is a cart-pole system, a simulated four dimensional nonlinear dynamic system. The cart-pole system has been the subject of previous research by (Widrow) and (Barto et. al.), whose results provide a comparison for the effectiveness of our TAC architecture. The system operates in the vertical plane and is controlled by application of a horizontal force. The goal of the controller is to return the cart-pole system to the origin of the state space from any initial state.
Metrics
3 Record Views
2 citations in Scopus
Details
- Title
- A trainable controller based on neural network
- Creators
- A GuezJ Selinsky
- Publication Details
- Neural networks, v 1(suppl), pp 336-336
- Publisher
- Elsevier
- Resource Type
- Journal article
- Language
- English
- Academic Unit
- Electrical and Computer Engineering
- Scopus ID
- 2-s2.0-0024173720
- Other Identifier
- 991020532095004721