Neural Networks for Identification, Prediction and Control

Artificial neural networks are computational models of the brain. There are many types of neural networks representing the brain’s structure and operation with varying degrees of sophistication. This chapter provides an introduction to the main types of networks and presents examples of each type.

A dynamic system can be described by two types of models: input-output models and state-space models. This chapter describes the use of feedforward neural networks to learn to act as both types of models.

As mentioned in Chapter 1, neural networks can be classified as feedforward networks and recurrent networks. In feedforward networks, the processing elements are connected in such a way that all signals flow in one direction from input units to output units. In recurrent networks there are both feedforward and feedback connections along which signals can propagate in opposite directions.

Chapters 2 and 3 have shown that neural networks can be employed to identify dynamic systems. The main advantages of neural networks over conventional identification methods include simplicity of implementation and good approximation properties [Warwick et aI, 1992]. In feedforward network based identification schemes, neural networks are used to represent the implied static mapping between the available input and output data. The network structures (number of layers and number of units in each layer) are predefined and remain unchanged both during and after training. Successful identification is often dependent on proper pre-estimation of the network structure.

This chapter discusses the application of neural networks to stock market and exchange rate prediction. The first section illustrates the use of the multilayer perception network for stock market prediction. The remainder of the chapter describes experiments with the multilayer perception network, the GMDH network, and the Elman network in exchange rate prediction.

Neural networks are developed by morphologically and computationally simulating a human brain. Although, as seen in previous chapters, the precise operation details of artificial neural networks are quite different from human brains, they are similar in three aspects. First, a neural network consists of a very large number of simple processing elements (the neurons). Second, each neuron is connected to a large number of other neurons. Third, the functionality of the networks is determined by modifying the strengths of connections during a learning phase [Psaltis et al, 1988; Hunt et al, 1992; Warwick et ai, 1992]. Efforts have been made to find efficient approaches for control from physiological studies of the brain. Research over the last twenty years has revealed the architecture and performance characteristics of the brain as a controller [Albus, 1975; Ito, 1984; Kawato et al, 1987] and has shown that neural network controllers have important advantages over conventional controllers. The first advantage is that a neural network controller can efficiently utilise a much larger amount of sensory information in planning and executing a control action than an industrial controller can. The second advantage is that a neural network controller has the collective processing capability that enables it to respond quickly to complex sensory inputs while the execution speed of sophisticated control algorithms in a conventional controller is severely limited.

This chapter shows that a single-input single-output (SISO) Fuzzy Logic Controller (FLC) [Mamdani, 1974; Lee, 1990a, 1990b] can be modelled as a neural network which can be trained using a Genetic Algorithm (GA). The GA is employed to determine the membership functions for the input variable, the quantisation levels of the output variable and the elements of the relation matrix of the FLC. The reasons for such a neuromorphic FLC are provided. The structure of the NN model proposed for an FLC and its GA-based training procedure are explained. Results for the simulated control of a time-delayed linear second-order plant and a non-linear plant are also given. In this chapter, it is assumed that the reader is familiar with fuzzy logic control and genetic algorithms. F or a basic introduction to these topics, see Appendix B and Appendix C.

The control of a multi-input-multi-output (MIMO) plant is a difficult problem when the plant is nonlinear and time-varying and there are dynamic interactions between the plant variables. A good example of such a plant is an articulated robot with two or more joints handling a changeable load.