In this chapter, we will introduce the patented 1-For-3 ANN Controller which was developed based on the Model-Free Adaptive Control Theory.
What is an Artificial Neural Network (ANN)
From the area of Bionics, the Artificial Neural Network (ANN) was originally proposed to emulate the structure and information processing of neurons in human brain. It is actually a set of algorithms that can perform amazing functions like a brain. It has been used successfully in the area of pattern recognition, image processing, and signal filtering due to its clear mathematical representations as well as its ability to learn, to approximate dynamics, and to classify patterns. [2]-[9]
Problems in Model Based ANN Control
An Artificial Neural Network, from the systems control point of view, can be considered as a large-scale, parallel processing, and nonlinear dynamic "block". In principle, the structure and parameters of an ANN can be adjusted based on some algorithms so that an ANN can perform as an adaptive controller.
During the last several years, steady progress has been made in the area of ANN based control. Quite a few ANN based control systems have been proposed. The most commonly used ANN in control is the multi-layered perceptron. The typical approach is to train the ANN controller by using either the approximations based on the mathematical model of the process if available, or an ANN model of the direct or even inverse dynamics of the process. [10]-[18]
The major difficulties with this approach are due to the ANN training which is usually time consuming and requires rich variety of signals. Also the local minimum problem in the deepest descent type of learning algorithm such as the back-propagation algorithm makes the training unreliable for dynamic systems; that is, we are not sure if the model obtained is useful or not. Other concerns may include the underlying assumption that the nonlinear static map generated by the ANN can adequately represent the system's behavior in the ranges of interest for the particular application.
It can be seen that these ANN control approaches have fallen into the same trap that troubled the traditional adaptive control for so long; that is, as long as an plant identification mechanism is used in the control scheme, the same old problems caused by the identification mechanism will still have to be resolved. To make improvement possible, we believe fundamental changes in the area of ANN control have to be made.
The followings points are covered in this chapter: