A learning architecture for control based on back-propagation neural networks

Abstract

A neural-network-based control architecture has been developed which can autonomously learn to perform kinematic control of an unknown system and/or adapt to a system which changes over time. It can control continuous-valued system variables to arbitrary accuracy using a small number of neurons. It learns to control the system more accurately than an analytically calculated controller. It is fault-tolerant in the presence of a large number (e.g., 30%) of component failures. The architecture has been used to learn to control a simulated robot arm of initially unknown characteristics. The simulations run in near real time.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>