DEVELOPMENT OF DYNAMIC NEURAL STRUCTURES WITH CONTROL APPLICATIONS

Dynamic neural networks, because they offer computational advantages over purely static neural networks, have many potential applications in a number of fields. The objective of the research described in this thesis was to develop dynamic neural structures for control applications. A dynamic model of the biological neuron called the dynamic neural unit (DNU) was developed for this purpose. The structure of the DNU is inspired by the topology of a reverberating circuit in a neuronal pool of the central nervous system. The DNU consists of internal feedforward and feedback synaptic weights followed by a nonlinear activation operator. It is thus different from the conventionally assumed structure of an artificial neuron. It is demonstrated in this thesis that a DNU can control unknown linear and simple nonlinear systems to track adaptively desired trajectories. The efficacy of artificial neural networks comes more from the number of neurons connected in the network and from the topology rather than from the computational ability of an isolated neuron. Considering the DNU as the basic functional element, a multi-stage dynamic neural network has been developed. One of the most important characteristics of neural networks is their ability to approximate arbitrary nonlinear functions. While most of the research work in this area has concentrated on static neural networks, a theoretical foundation of functional approximation using a dynamic neural network has been developed. Computer simulation studies are provided to substantiate the theoretical developments. Following this development, the dynamic neural network has been used in a direct adaptive control mode to cause unknown nonlinear systems to follow desired reference signals. In conventional static neural structures, the optimum slope of a nonlinear activation function is usually determined by trial and error. An improper selection of the slope may lead to instability. The importance of using an adaptive activation operator in neural networks has been demonstrated through computer simulations. In this context, the concept of somatic adaptation for dynamic neural structures has been introduced. The significance of this concept as applied to the control of unknown nonlinear dynamic systems has been extensively studied through computer simulations. A new dynamic neural structure called the dynamic neural processor (DNP) that emphasizes the aggregate dynamic properties of a neural population has also been proposed and reported in this thesis. This structure is based upon the hypothesis that the neurophysiological activities of any complexity are dependent upon the interaction of antagonistic (excitatory and inhibitory) neural subpopulations. The DNP consists of two DNUs which are configured to function as excitatory and inhibitory neurons. A mathematical model and an algorithm to modify the parameters of the DNP have been developed. Four applications of the DNP, the functional approximation of nonlinear functions, computation of inverse kinematic transformations of a two-link robot, control of unknown single-input-single-output nonlinear systems, and coordination and control of multiple-input-multiple-output systems, are presented. A brief comparative study of the performance of this neural model with that of conventionally used recurrent neural networks has also been presented. A generalized dynamic neural model based on the concept of neural subpopulations has been proposed in this thesis. It is shown that many existing neural structures can be derived from this generalized neural model