Margarita Terziyska

Modeling of Chaotic Time Series by Interval Type-2 NEO-Fuzzy Neural Network

This paper describes the development of Interval Type-2 NEO-Fuzzy Neural Network for modeling of complex dynamics. The proposed network represents a parallel set of multiple zero order Sugeno type approximations, related only to their own input argument. The induced gradient based learning procedure, adjusts solely the consequent network parameters. To improve the robustness of the network and the possibilities for handling uncertainties, Type-2 Gaussian fuzzy sets are introduced into the network topology. The potentials of the proposed approach in modeling of Mackey-Glass and Rossler Chaotic time series are studied.

Fuzzy-neural model predictive control of a building heating system

Abstract This paper describes the development of a Model Predictive Controller with supervision control of a building heating system. A fuzzy–neural model and optimizing procedure as a part of a nonlinear predictive controller are utilized on-line to determine the future values of control actions based on dependence between outdoor and indoor temperatures. A learning algorithm for parameters in fuzzy-neural implementation of the predictive model is additionally applied. Simulation results with a model of a single room heating system demonstrate that a better system performance can be achieved in comparison to classical PID control.

Adaptive supervisory tuning of nonlinear model predictive controller for a heat exchanger

Abstract It is presented in this paper an adaptive predictive supervisory algorithm to the temperature control of a heating system with a heat exchanger. The nonlinear predictive control strategy is designed on the basis of a Takagi-Sugeno fuzzy-neural model and a simple optimization procedure. An additional supervisory level in the control system is introduced for adaptive tuning of a weighting factor in the predefined optimization criterion. Using the proposed algorithm a higher system performance can be achieved which leads to reduction of the energy consumption into the heating system. The proposed approach is studied by experimental simulations to control a temperature in the heating system.

State-space fuzzy-neural network for modeling of nonlinear dynamics

This paper describes a novel idea for designing a fuzzy-neural network for modeling of nonlinear system dynamics. The presented approach assumes a state-space representation in order to obtain a more compact form of the model, without statement of a great number of parameters needed to represent a nonlinear behavior. To increase the flexibility of the network, simple Takagi-Sugeno inferences are used to estimate the current system states, by a set of a multiple local linear state estimators. Afterwards, the output of the network is defined, as function of the current and estimated system parameters. A simple learning algorithm based on two step Gradient descent procedure to adjust the network parameters, is applied. The potentials of the proposed modeling network are demonstrated by simulation experiments to model an oscillating pendulum and a nonlinear drying plant.

Input space selective fuzzification in intuitionistic semi fuzzy-neural network

In this paper, the influence of the selective fuzzification of the input space in Intuitionistic Semi-Fuzzy Neural Network (ISFNN) is investigated. The ISFNN represents a structure modification of the classical fuzzy-neural approach where selective fuzzification as a means to reduce the number of the generated fuzzy rules is proposed, thus expected to reduce the number of the associated learning parameters and to achieve a degree of computational simplicity. On the other hand, the potentials of the network are supplemented by intuitionistic fuzzy logic, in order to handle uncertain data variations. As a learning procedure for the proposed structure, a two-step gradient descent algorithm is employed. To investigate the influence of input space fuzzificaton, several test experiments in modeling of a two benchmark chaotic systems — Mackey-Glass and Rossler chaotic time series are made.

Intuitionistic Neo-Fuzzy predictive control

This paper presents an implicit predictive control strategy based on Intuitionistic Neo-Fuzzy predictor, as a first attempt to investigate the potentials of the intuitionistic fuzzy logic for the purpose of control applications. The proposed predictor represents a simple fuzzy-neural network as fusion from the concepts of the intuitionistic fuzzy logic, the neo-fuzzy neuron theory and the classical Takagi-Sugeno inference mechanism. The predictions are then coupled into generalized predictive control scheme where a standard quadratic control cost function is minimized over a set of predefined horizons. For simplicity, the considered process variables and the calculated output control sequence are iteratively bounded instead of explicitly constrained, in order to investigate the computational procedures related to implementation of an intuitionistic fuzzy logic. To investigate the potentials of the proposed predictive control approach, numerical experiments to control a Continuous Stirred Tank Reactor (CSTR) under uncertain conditions are studied.

Intuitionistic Neo-Fuzzy Network for modeling of nonlinear systems dynamics

In this paper, an approach to design an Intuitionistic Neo-Fuzzy Network (INFN) is presented. The proposed architecture combines the advantages of the Intuitionistic Fuzzy Logic (IFL) to deal with uncertainties and the Neo-Fuzzy Neural Network approach to represent nonlinear systems with topologies including small number of parameters. As a learning approach for the consequent fuzzy rules parameters, the gradient optimization procedure is proposed. The investigate the potentials of the generated INF structure, the modeling of a three benchmark chaotic time series - Mackey-Glass, Lorenz and Rossler under uncertain conditions is investigated. The obtained results prove the flexibility of the approach and its further extension to Model Predictive Control is investigated too.

NEO-Fuzzy Neural Networks for Knowledge Based Modeling and Control of Complex Dynamical Systems

Capturing the dynamics and control of fast complex nonlinear systems often requires the application of computationally efficient modeling structures in order to track the system behavior without loss of accuracy and to provide reliable predictions on purpose to process control. An available approach is to employ fuzzy-neural networks, whose abilities to handle dynamical data streams and to build rule-based relationships makes them a flexible solution. A major drawback of the classical fuzzy-neural networks is the large number of parameters associated with the rules premises and consequents parts, which need to be adapted at each discrete time instant. Therefore, in this chapter several structures with reduced number of parameters lying in the framework of a NEO-Fuzzy neuron are proposed. To increase the robustness of the models when addressing to uncommon/uncertain data variations, Type-2 and Intuitionistic fuzzy logic are introduced. An approach to design a simple NEO-Fuzzy state-space predictive controller shows the potential applicability of the proposed models for process control.

Efficient Error Based Metrics for Fuzzy-Neural Network Performance Evaluation

In this paper the effectiveness of different error metrics for assessment of the capabilities of an advanced fuzzy-neural architecture are studied. The proposed structure combines the potentials of the Intuitionistic Fuzzy Logic with the simplicity of the Neo-Fuzzy Neuron theory for implementation of robust modeling mechanisms, able to capture uncertain variations in the data space. A major concern when evaluating the performance of such kind of models is the selection of appropriate error metrics in order to assess their potential to capture a wide range of system behaviours. Therefore, different error metrics to evaluate the functional properties of a proposed Intuitionistic Neo-fuzzy network are studied and a comparative analysis in modeling of chaotic time series is made.

Intuitionistic fuzzy radial basis functions network for modeling of nonlinear dynamics

This paper deals with a design methodology for a neural network with improved robust qualities in notion to handling uncertain input data space variations. The proposed network topology combines the simplicity of the radial basis functions networks to interpret or classify data pairs and the abilities of the intuitionistic fuzzy logic to deal with the vagueness of the data space. A simplified gradient optimization procedure as a learning approach for the designed hybrid neural network is proposed. To investigate the effects of the generated structure throughout varying network parameters, the modeling of a two benchmark chaotic time series — Mackey-Glass and Rossler under uncertain conditions is investigated. The obtained results prove the flexibility of the approach and its potentials to cope with data variations.