Alex Alexandridis

A new algorithm for developing dynamic radial basis function neural network models based on genetic algorithms

Abstract A new method for extracting valuable process information from input–output data is presented in this paper. The proposed methodology produces dynamical radial basis function (RBF) neural network models based on a specially designed genetic algorithm (GA), which is used to auto-configure the structure of the networks and obtain the model parameters. The new RBF network training technique formulates a complete optimization problem, which includes the network structure into the set of free variables that are used to minimize the prediction error. This is a different approach compared with the local search methods employed by other structure selection mechanisms, which are often trapped to local minima. Another advantage of the proposed method is that only one run of the algorithm is required to obtain the optimal network structure, in contrast to the standard RBF training techniques, where the produced model is selected by trial and error. The effectiveness of the method is illustrated through the development of dynamical models for two sets of data: simulated data from a Continuous Stirred Tank Reactor (CSTR) and true data collected from a Kamyr digester, which is a rather complicated reactor used in the pulp and paper industry.

A new algorithm for online structure and parameter adaptation of RBF networks

This paper deals with the problem of online adaptation of radial basis function (RBF) neural networks. A new adaptive training method is presented, which is able to modify both the structure of the network (the number of nodes in the hidden layer) and the output weights, as the algorithm proceeds. These adaptation capabilities make the algorithm suitable for modeling dynamical time varying systems, where not only the dynamics but also the operating region changes with time. Therefore, the important issue of extrapolation is faced successfully, but at the same time the algorithm takes care of the size of the network, by deleting the hidden node centers that remain inactive for a long time. The selection of the network centers is based on a fuzzy partition of the input space, which defines a number of fuzzy subspaces. The algorithm considers the centers of the fuzzy subspaces as candidates for becoming hidden node centers and makes the selections, so that at least one center is close enough to each input example. The proposed technique is illustrated through the application to time varying dynamical systems and is compared to other adaptive training methods.

A cellular automata model for forest fire spread prediction: The case of the wildfire that swept through Spetses Island in 1990

We present and illustrate the simulation results of a cellular automata model describing the dynamics of a forest fire spread on a mountainous landscape taking into account factors such as the type and density of vegetation, the wind speed and direction and the spotting phenomenon. The model is used to simulate the wildfire that broke up on Spetses in August of 1990 and destroyed a major part of the Island’s forest. We used a black-box non-linear optimization approach to fine-tune some of the model’s parameters based on a geographical information system incorporating available data from the real forest fire. The comparison between the simulation and the actual-observed results showed that the proposed model predicts in a quite adequate manner the evolution characteristics in space and time of the real incident and as such could be potentially used to develop a fire risk-management tool for heterogeneous landscapes.

A fast training algorithm for RBF networks based on subtractive clustering

Abstract A new algorithm for training radial basis function neural networks is presented in this paper. The algorithm, which is based on the subtractive clustering technique, has a number of advantages compared to the traditional learning algorithms, including faster training times and more accurate predictions. Due to these advantages the method proves suitable for developing models for complex nonlinear systems.

A classification technique based on radial basis function neural networks

In this paper, a new classification method is proposed based on the radial basis function (RBF) neural network architecture. The method is particularly useful for manufacturing processes, in cases where on-line sensors for classifying the product quality are not available. More specifically, the fuzzy means algorithm is employed on a set of training data, where the input data refer to variables that are measured on-line and the output data correspond to quality variables that are classified by human experts. The produced neural network model acts as an artificial sensor that is able to classify the product quality in real time. The proposed method is illustrated through an application to real data collected from a paper machine. The method produces successful results and outperforms a number of classifiers, which are based on the feedforward neural network (FNN) architecture.

A Novel RBF Neural Network Training Methodology to Predict Toxicity to Vibrio Fischeri

SummaryThis work introduces a neural network methodology for developing QSTR predictors of toxicity to Vibrio fischeri. The method adopts the Radial Basis Function (RBF) architecture and the fuzzy means training strategy, which is fast and repetitive, in contrast to most traditional training techniques. The data set that was utilized consisted of 39 organic compounds and their corresponding toxicity values to Vibrio fischeri, while lipophilicity, equalized electronegativity and one topological index were used to provide input information to the models. The performance and predictive ability of the RBF model were illustrated through external validation and various statistical tests. The proposed methodology can be used to successfully model toxicity to Vibrio fischerifor a heterogeneous set of compounds.

Modelling of nonlinear process dynamics using Kohonen's neural networks, fuzzy systems and Chebyshev series

Publisher Summary This chapter presents a new systematic methodology that facilitates the development of both linguistic and nonlinear analytical models with the aid of Kohonens self-organizing neural networks, fuzzy logic, and truncated Chebyshev series. The result is a model structure, which contains only few essential polynomial terms that are suitable to capture both the qualitative and the quantitative characteristics of the system dynamics. The methodology is applied to the identification of a continuous stirred tank reactor (CSTR), which, depending on the operating conditions may exhibit multiple steady states and limit cycles. The results demonstrate that the produced model captures the qualitative dynamic behavior of the process and offers a satisfactory quantitative approximation. For comparison purposes, the performance of the proposed approach is compared with two other identification methods: one based on feed-forward neural networks and one based on normal form theory.

Development of nonlinear quantitative structure-activity relationships using RBF networks and evolutionary computing

Quantitative Structure Activity Relationships (QSARs) are mathematical models that correlate structural or property descriptions of compounds (hydrophobicity, topology, electronic properties etc.) with activities, such as chemical measurements and biological assays. In this paper we propose a modeling methodology suitable for QSAR studies which selects the proper descriptors based on evolutionary computing and finally produces Radial Basis Function (RBF) neural network models. The method is successfully applied to the benchmark Selwood data set.

MODELING AND CONTROL OF CONTINUOUS DIGESTERS USING THE PLS METHODOLOGY

The partial least squares (PLS) methodology is a regression tool able to deal with two important problems in system identification: the noise contained in the industrial data and the correlations that are observed among the input variables. In this article, the PLS technique is applied on real industrial data to produce dynamical multi-input single-output (MISO) models of the behavior of the kappa number in a continuous digester with respect to the important input variables that can be measured on-line. The ultimate goal is the development of a model predictive control (MPC) scheme that can be used for keeping the kappa number of the produced pulp close to the desired set point. We show that a modification in the standard MPC algorithm is needed in order to take into account the correlations among the input variables.

Adaptive control of continuous pulp digesters based on radial basis function neural network models

Abstract In this paper, an adaptive nonlinear Model Predictive Control (MPC) configuration is proposed, where the model used for predicting the future behavior of the process is a discrete dynamic Radial Basis Function (RBF) network. The innovative fuzzy clustering algorithm allows the continuous and easy adaptation of the model, thus making it suitable for controlling time-varying processes, such as the continuous pulp digester.