Milorad M. Bozic

An adequate training set for the AIMNC strategy for typical industrial processes

Here we have discussed how the training data set should be selected for the Approximate Internal Model-based Neural Control (AIMNC) applied to the typical industrial processes. In the considered control strategy only one neural network (NN), Multi Layer NN (MLNN), which is the neural model of the plant, should be trained off-line. An inverse neural controller can be directly obtained from the neural model without necessity of a further training. Simulations demonstrate performance of the AIMNC strategy for NN model obtained with adequate training set.

On sensitivity of neural adaptive filters with respect to the slope parameter of a neuron activation function

Sensitivity analysis of neural adaptive filters with respect to the slope parameter of a neuron activation function is performed. The analysis is provided both for a feedforward neural adaptive filter and a recurrent perceptron. The slope affects stability and convergence characteristics of a filter via inherent relationship between the slope and the learning rate parameter. In addition, it determines character of an activation function, i.e. whether it is contractive or expansive mapping. Presented analysis shows that gradient-descent based learning algorithms with an adaptive learning rate significantly reduce sensitivity of a neural adaptive filter with respect to the slope parameter, when compared with learning algorithms with a constant learning rate. Experimental results on the test speech and HRV signals support the analysis.

A neural nonlinear adaptive filter with a trainable activation function

The normalized nonlinear gradient descent learning algorithm (NNGD) for a class of nonlinear finite impulse response (FIR) adaptive filters (dynamical perceptron) is extended to the case where the amplitude of the nonlinear activation function is made gradient adaptive. This makes the adaptive amplitude normalized nonlinear gradient descent (AANNGD) algorithm. The AANNGD is suitable for processing of nonlinear and nonstationary signals with a large dynamical range. Experimental results show that AANNGD outperforms the standard LMS, NGD, NNGD, the fully adaptive (FANNGD) and the sign algorithm on nonlinear input with large dynamics.

Global asymptotic stability for RNNs with a bipolar activation function

Conditions for global asymptotic stability of a nonlinear relaxation process realized by a recurrent neural network with a hyperbolic tangent activation function are provided. This analysis is based upon the contraction mapping theorem and corresponding fixed point iteration. The derived results find their application in the wide area of neural networks for optimization and signal processing.

A neuro-adaptive control of nonlinear slow processes

A Neuro-Adaptive Internal Model-based Control (NAIMC), using the Fast Clustered Radial Basis Function Network (FCRBFN) equipped by the Stochastic Gradient Descent (SGD) learning algorithm is proposed to control the nonlinear plant with slow dynamics. As a first step in this design approach, the classical feedback controller is applied to improve the overall dynamic characteristics of the obtained local closed loop. Such local loop is further on considered as an equivalent plant to which the NAIMC can be applied. The improved characteristics of the equivalent plant can be usually obtained by some kind of the PD control law and we used this approach at the NAIMC design of the nonlinear slow process. To achieve a zero-steady state error in cases of the piecewise constant changes of the reference and disturbance at output of the plant, we applied the method of Gain Scheduling (GS) for adjusting the gain of the Q controller in the NAIMC based structure. We illustrate the performance of the proposed NAIMC design using simulation results for the control of a double tank system.

A class of neural adaptive FIR filters for complex-valued load prediction

Load prediction is a necessity in a deregulated electrical energy sector. It is important financially and technically. In order to cope with nonlinear and non stationary character of a load signal, an efficient adaptive predictor should be employed. Also, power utilities manage load information as a complex-valued signal. To this cause, performance of a class of complex-valued gradient descent (GD) neural adaptive finite impulse response (FIR) filters is analyzed. It is shown that fully complex nonlinear GD algorithms have the best performance in a load prediction task. To support the analysis, experiments are carried out on the test load signal, metered on a medium voltage feeder.