Intelligent coordination of multiple systems with neural networks.

Abstract

Many control applications require cooperation of two or more independently designed, separately located, but mutually affecting, subsystems. In addition to the proper functioning of each subsystem, their effective coordination is very important to achieve the desired performance. This has led to the development of new multiple-system coordinators and the evaluation of their performance in real-time implementation. Two coordination schemes have been proposed: a knowledge-based coordinator (KBC) and a neural network-based coordinator (NNBC). Either of them functions as a high-level coordinator in a hierarchical system. The basic idea of the KBC is to estimate the effects of commands to low-level subsystems using a predictor and modify them by searching a knowledge base. By introducing the predictor, the knowledge base for multiple-system coordination is greatly simplified, and each command is evaluated before its execution. The basic structure of the NNBC is a multilayer perceptron. Neural networks (NNs) are usually trained by using the errors at its output layer. However, when an NN is used to control a plant directly, these errors are unknown, since the desired control actions are unknown. This implies that the conventional back propagation training algorithm cannot be applied to control problems directly. A simple training algorithm has been developed which enables the NNBC to be trained by using the output errors of the controlled plant. The effects of computing time delay on the performance of controlled systems are studied. For the qualitative analysis, a generic criterion is derived in terms of system stability. For quantitative analysis, upper bounds of computing time delay on system stability and performance are derived. These upper bounds can be used as extra constraints on controller design and CPU selection in implementation. The coordination of two robots holding an object, the coordination of multiple robots to avoid collision, and the control of a nonlinear thermo-process are investigated to test the capability of the proposed schemes. Because the internal structure and parameters of the low-level subsystems are not affected by using either the KBC or the NNBC, some commercially-designed servo controllers can be coordinated to accomplish more sophisticated tasks than originally intended.