On Applications of New Soft and Evolutionary Computing Techniques to Direct and Inverse Modeling Problems

Abstract

Adaptive direct modeling or system identification and adaptive inverse modeling or channel equalization find extensive applications in telecommunication, control system, instrumentation, power system engineering and geophysics. If the plants or systems are nonlinear, dynamic, Hammerstein and multiple-input and multiple-output (MIMO) types, the identification task
becomes very difficult.

Further, the existing conventional methods like the least mean square (LMS) and recursive least square (RLS) algorithms do not provide satisfactory training to develop accurate direct and inverse models. Very often these (LMS and RLS) derivative based algorithms do not lead to optimal solutions in pole-zero and Hammerstein type system identification problem as they have tendency to be trapped by local minima.

In many practical situations the output data are contaminated with impulsive type outliers in addition to measurement noise. The density of the outliers may be up to 50%, which means that about 50% of the available data are affected by outliers. The strength of these outliers may be two to five times the maximum amplitude of the signal. Under such adverse conditions the available learning algorithms are not effective in imparting satisfactory training to update the weights of the adaptive models. As a result the resultant direct and inverse models become inaccurate and improper.

Hence there are three important issues which need attention to be resolved. These are :
(i) Development of accurate direct and inverse models of complex plants using some novel architecture and new learning techniques.
(ii) Development of new training rules which alleviates local minima problem during training and thus help in generating improved adaptive models.
(iii) Development of robust training strategy which is less sensitive to outliers in training and thus to create identification and equalization models which are robust
against outliers.

These issues are addressed in this thesis and corresponding contribution are outlined in seven Chapters. In addition, one Chapter on introduction, another on required architectures and algorithms and last Chapter on conclusion and scope for further research work are embodied
in the thesis.

A new cascaded low complexity functional link artificial neural network (FLANN) structure is proposed and the corresponding learning algorithm is derived and used to identify nonlinear dynamic plants. In terms of identification performance this model is shown to outperform the multilayer perceptron and FLANN model. A novel method of identification of IIR plants is proposed using comprehensive learning particle swarm optimization (CLPSO) algorithm. It is shown that the new approach is more accurate in identification and takes less CPU time
compared to those obtained by existing recursive LMS (RLMS), genetic algorithm (GA) and PSO based approaches. The bacterial foraging optimization (BFO) and PSO are used to develop efficient learning algorithms to train models to identify nonlinear dynamic and MIMO plants. The new scheme takes less computational effort, more accurate and consumes less input samples for training. Robust identification and equalization of complex plants have been carried out using outliers in training sets through minimization of robust norms using PSO and BFO based methods. This method yields robust performance both in equalization and identification tasks. Identification of Hammerstein plants has been achieved successfully using PSO, new clonal PSO (CPSO) and immunized PSO (IPSO) algorithms. Finally the thesis proposes a distributed approach to identification of plants by developing two distributed learning algorithms : incremental PSO and diffusion PSO. It is shown that the new approach is more efficient in terms of accuracy and training time compared to centralized PSO based approach. In addition a robust distributed approach for identification is proposed and its performance has been evaluated.

In essence the thesis proposed many new and efficient algorithms and structure for identification and equalization task such as distributed algorithms, robust algorithms, algorithms for ploe-zero identification and Hammerstein models. All these new methods are shown to be better in terms of performance, speed of computation or accuracy of results.