arXiv subjects

Tshilidzi Marwala

Publications and source records attributed to Tshilidzi Marwala.

At least 73 records · Page 4Linked to original sources

arXiv2007

Evolving Classifiers: Methods for Incremental Learning

The ability of a classifier to take on new information and classes by evolving the classifier without it having to be fully retrained is known as incremental learning. Incremental learning has been successfully applied to many classification problems, where the data is changing and is not all available at once. In this paper there is a comparison between Learn++, which is one of the most recent incremental learning algorithms, and the new proposed method of Incremental Learning Using Genetic Algorithm (ILUGA). Learn++ has shown good incremental learning capabilities on benchmark datasets on which the new ILUGA method has been tested. ILUGA has also shown good incremental learning ability using only a few classifiers and does not suffer from catastrophic forgetting. The results obtained for ILUGA on the Optical Character Recognition (OCR) and Wine datasets are good, with an overall accuracy of 93% and 94% respectively showing a 4% improvement over Learn++.MT for the difficult multi-class OCR dataset.

cs.LG↗

arXiv2007

Autoencoder, Principal Component Analysis and Support Vector Regression for Data Imputation

Data collection often results in records that have missing values or variables. This investigation compares 3 different data imputation models and identifies their merits by using accuracy measures. Autoencoder Neural Networks, Principal components and Support Vector regression are used for prediction and combined with a genetic algorithm to then impute missing variables. The use of PCA improves the overall performance of the autoencoder network while the use of support vector regression shows promising potential for future investigation. Accuracies of up to 97.4 % on imputation of some of the variables were achieved.

cs.AI↗

arXiv2007

Estimation of Missing Data Using Computational Intelligence and Decision Trees

This paper introduces a novel paradigm to impute missing data that combines a decision tree with an auto-associative neural network (AANN) based model and a principal component analysis-neural network (PCA-NN) based model. For each model, the decision tree is used to predict search bounds for a genetic algorithm that minimize an error function derived from the respective model. The models' ability to impute missing data is tested and compared using HIV sero-prevalance data. Results indicate an average increase in accuracy of 13% with the AANN based model's average accuracy increasing from 75.8% to 86.3% while that of the PCA-NN based model increasing from 66.1% to 81.6%.

stat.AP↗

arXiv2007

Bayesian Approach to Neuro-Rough Models

This paper proposes a neuro-rough model based on multi-layered perceptron and rough set. The neuro-rough model is then tested on modelling the risk of HIV from demographic data. The model is formulated using Bayesian framework and trained using Monte Carlo method and Metropolis criterion. When the model was tested to estimate the risk of HIV infection given the demographic data it was found to give the accuracy of 62%. The proposed model is able to combine the accuracy of the Bayesian MLP model and the transparency of Bayesian rough set model.

cs.AI↗

arXiv2007

Automatic Detection of Pulmonary Embolism using Computational Intelligence

This article describes the implementation of a system designed to automatically detect the presence of pulmonary embolism in lung scans. These images are firstly segmented, before alignment and feature extraction using PCA. The neural network was trained using the Hybrid Monte Carlo method, resulting in a committee of 250 neural networks and good results are obtained.

cs.CV↗

arXiv2007

Multi-Agent Modeling Using Intelligent Agents in the Game of Lerpa

Game theory has many limitations implicit in its application. By utilizing multiagent modeling, it is possible to solve a number of problems that are unsolvable using traditional game theory. In this paper reinforcement learning is applied to neural networks to create intelligent agents

cs.MA↗

arXiv2007

Using Genetic Algorithms to Optimise Rough Set Partition Sizes for HIV Data Analysis

In this paper, we present a method to optimise rough set partition sizes, to which rule extraction is performed on HIV data. The genetic algorithm optimisation technique is used to determine the partition sizes of a rough set in order to maximise the rough sets prediction accuracy. The proposed method is tested on a set of demographic properties of individuals obtained from the South African antenatal survey. Six demographic variables were used in the analysis, these variables are; race, age of mother, education, gravidity, parity, and age of father, with the outcome or decision being either HIV positive or negative. Rough set theory is chosen based on the fact that it is easy to interpret the extracted rules. The prediction accuracy of equal width bin partitioning is 57.7% while the accuracy achieved after optimising the partitions is 72.8%. Several other methods have been used to analyse the HIV data and their results are stated and compared to that of rough set theory (RST).

cs.NE↗

arXiv2007

Finite Element Model Updating Using Bayesian Approach

This paper compares the Maximum-likelihood method and Bayesian method for finite element model updating. The Maximum-likelihood method was implemented using genetic algorithm while the Bayesian method was implemented using the Markov Chain Monte Carlo. These methods were tested on a simple beam and an unsymmetrical H-shaped structure. The results show that the Bayesian method gave updated finite element models that predicted more accurate modal properties than the updated finite element models obtained through the use of the Maximum-likelihood method. Furthermore, both these methods were found to require the same levels of computational loads.

stat.AP↗

arXiv2007

Computational Intelligence for Condition Monitoring

Condition monitoring techniques are described in this chapter. Two aspects of condition monitoring process are considered: (1) feature extraction; and (2) condition classification. Feature extraction methods described and implemented are fractals, Kurtosis and Mel-frequency Cepstral Coefficients. Classification methods described and implemented are support vector machines (SVM), hidden Markov models (HMM), Gaussian mixture models (GMM) and extension neural networks (ENN). The effectiveness of these features were tested using SVM, HMM, GMM and ENN on condition monitoring of bearings and are found to give good results.

cs.CE↗

arXiv2007

Fuzzy and Multilayer Perceptron for Evaluation of HV Bushings

The work proposes the application of fuzzy set theory (FST) to diagnose the condition of high voltage bushings. The diagnosis uses dissolved gas analysis (DGA) data from bushings based on IEC60599 and IEEE C57-104 criteria for oil impregnated paper (OIP) bushings. FST and neural networks are compared in terms of accuracy and computational efficiency. Both FST and NN simulations were able to diagnose the bushings condition with 10% error. By using fuzzy theory, the maintenance department can classify bushings and know the extent of degradation in the component.

cs.AI↗

arXiv2007

A Study in a Hybrid Centralised-Swarm Agent Community

This paper describes a systems architecture for a hybrid Centralised/Swarm based multi-agent system. The issue of local goal assignment for agents is investigated through the use of a global agent which teaches the agents responses to given situations. We implement a test problem in the form of a Pursuit game, where the Multi-Agent system is a set of captor agents. The agents learn solutions to certain board positions from the global agent if they are unable to find a solution. The captor agents learn through the use of multi-layer perceptron neural networks. The global agent is able to solve board positions through the use of a Genetic Algorithm. The cooperation between agents and the results of the simulation are discussed here. .

cs.NE↗

arXiv2007

Fault Classification using Pseudomodal Energies and Neuro-fuzzy modelling

This paper presents a fault classification method which makes use of a Takagi-Sugeno neuro-fuzzy model and Pseudomodal energies calculated from the vibration signals of cylindrical shells. The calculation of Pseudomodal Energies, for the purposes of condition monitoring, has previously been found to be an accurate method of extracting features from vibration signals. This calculation is therefore used to extract features from vibration signals obtained from a diverse population of cylindrical shells. Some of the cylinders in the population have faults in different substructures. The pseudomodal energies calculated from the vibration signals are then used as inputs to a neuro-fuzzy model. A leave-one-out cross-validation process is used to test the performance of the model. It is found that the neuro-fuzzy model is able to classify faults with an accuracy of 91.62%, which is higher than the previously used multilayer perceptron.

cs.AI↗

arXiv2007

Finite Element Model Updating Using Response Surface Method

This paper proposes the response surface method for finite element model updating. The response surface method is implemented by approximating the finite element model surface response equation by a multi-layer perceptron. The updated parameters of the finite element model were calculated using genetic algorithm by optimizing the surface response equation. The proposed method was compared to the existing methods that use simulated annealing or genetic algorithm together with a full finite element model for finite element model updating. The proposed method was tested on an unsymmetri-cal H-shaped structure. It was observed that the proposed method gave the updated natural frequen-cies and mode shapes that were of the same order of accuracy as those given by simulated annealing and genetic algorithm. Furthermore, it was observed that the response surface method achieved these results at a computational speed that was more than 2.5 times as fast as the genetic algorithm and a full finite element model and 24 times faster than the simulated annealing.

cs.CE↗

arXiv2007

Dynamic Model Updating Using Particle Swarm Optimization Method

This paper proposes the use of particle swarm optimization method (PSO) for finite element (FE) model updating. The PSO method is compared to the existing methods that use simulated annealing (SA) or genetic algorithms (GA) for FE model for model updating. The proposed method is tested on an unsymmetrical H-shaped structure. It is observed that the proposed method gives updated natural frequencies the most accurate and followed by those given by an updated model that was obtained using the GA and a full FE model. It is also observed that the proposed method gives updated mode shapes that are best correlated to the measured ones, followed by those given by an updated model that was obtained using the SA and a full FE model. Furthermore, it is observed that the PSO achieves this accuracy at a computational speed that is faster than that by the GA and a full FE model which is faster than the SA and a full FE model.

cs.CE↗

arXiv2007

Modeling and Controlling Interstate Conflict

Bayesian neural networks were used to model the relationship between input parameters, Democracy, Allies, Contingency, Distance, Capability, Dependency and Major Power, and the output parameter which is either peace or conflict. The automatic relevance determination was used to rank the importance of input variables. Control theory approach was used to identify input variables that would give a peaceful outcome. It was found that using all four controllable variables Democracy, Allies, Capability and Dependency; or using only Dependency or only Capabilities avoids all the predicted conflicts.

cs.CY↗

arXiv2007

HMM Speaker Identification Using Linear and Non-linear Merging Techniques

Speaker identification is a powerful, non-invasive and in-expensive biometric technique. The recognition accuracy, however, deteriorates when noise levels affect a specific band of frequency. In this paper, we present a sub-band based speaker identification that intends to improve the live testing performance. Each frequency sub-band is processed and classified independently. We also compare the linear and non-linear merging techniques for the sub-bands recognizer. Support vector machines and Gaussian Mixture models are the non-linear merging techniques that are investigated. Results showed that the sub-band based method used with linear merging techniques enormously improved the performance of the speaker identification over the performance of wide-band recognizers when tested live. A live testing improvement of 9.78% was achieved

cs.LG↗

arXiv2007

Evolutionary Optimisation Methods for Template Based Image Registration

This paper investigates the use of evolutionary optimisation techniques to register a template with a scene image. An error function is created to measure the correspondence of the template to the image. The problem presented here is to optimise the horizontal, vertical and scaling parameters that register the template with the scene. The Genetic Algorithm, Simulated Annealing and Particle Swarm Optimisations are compared to a Nelder-Mead Simplex optimisation with starting points chosen in a pre-processing stage. The paper investigates the precision and accuracy of each method and shows that all four methods perform favourably for image registration. SA is the most precise, GA is the most accurate. PSO is a good mix of both and the Simplex method returns local minima the most. A pre-processing stage should be investigated for the evolutionary methods in order to improve performance. Discrete versions of the optimisation methods should be investigated to further improve computational performance.

cs.CE↗

arXiv2007

Option Pricing Using Bayesian Neural Networks

Options have provided a field of much study because of the complexity involved in pricing them. The Black-Scholes equations were developed to price options but they are only valid for European styled options. There is added complexity when trying to price American styled options and this is why the use of neural networks has been proposed. Neural Networks are able to predict outcomes based on past data. The inputs to the networks here are stock volatility, strike price and time to maturity with the output of the network being the call option price. There are two techniques for Bayesian neural networks used. One is Automatic Relevance Determination (for Gaussian Approximation) and one is a Hybrid Monte Carlo method, both used with Multi-Layer Perceptrons.

cs.CE↗