I’m willing to supervise student projects in the area of my research interests. Below is a list of indicative project topics. Undergraduate and postgraduate students are encouraged to email me to arrange an appointment in order to discuss projects of the list, or their own proposal.
1. Neural networks for time-series analysis and forecasting
In many scientific, economic and engineering applications there arises the problem of predicting the future on the basis of some collected historical data. The most powerful approach to the problem of prediction is to find a law underlying the given dynamic process or phenomenon. An alternative approach is to discover some strong empirical regularities in the observation of the time series. Unfortunately, the information about the dynamic process under investigation is often partial and incomplete and regularities, such as periodicity, are usually masked by noise. This project will provide a review of neural network-based methods for time-series analysis and forecasting. In the implementation phase of the project the use of a neural network in a practical prediction task will be investigated.
Indicative literature
1. Gately, E. J, (1996) Neural Networks for Financial Forecasting, John Wiley & Sons, United States of America.
2. Gilbert, J. (1995) Artificial Intelligence on Wall Street: An Overview and Critique of Applications in the Finance Industry’, http://gryphon.ccs.brandeis.edu/~grath/brandeis/ai-paper/
3. Giles, C. L., Lawrence, S., Tsoi C. A., (1997), ‘Rule Inference for Financial Prediction using Recurrent Neural Networks’, Proceedings of IEEE/IAFE Conference on Computational Intelligence for Financial Engineering (CIFEr), IEEE, Piscataway, NJ, pp. 253–259.
4. Kaashoek, J. F., (1998) ‘A Simple Strategy to Prune Neural Networks with an Application to Economic Time Series’, Paper provided by Erasmus University Rotterdam, Econometric Institute in its series Econometric Institute Report, No 103.
5. Wong, C. C., Chan, M. C., Lam, C., (2000) ‘Financial Time Series Forecasting By Neural Network Using Conjugate Gradient Learning Algorithm And Multiple Linear Regression Weight Initialisation’ Paper provided by Society for Computational Economics in its series Computing in Economics and Finance 2000, No 61.
Search keywords: prediction, forecasting, multivariate time series
2. Backpropagation training and local minima
The commonly used training methods are gradient-based algorithms such as the widely used backpropagation. These are local minimisation methods and have no mechanism that allows them to escape the influence of a local minimum. In this project global search methods for feedforward neural network batch training will be investigated. These methods are expected to lead to "optimal" or near-optimal weight configurations by allowing the network to escape from local minima during training. Specifically, the practical part of the project will be focused on the software implementation of a method of this type and its application on a notorious for its local minima learning problem.
Indicative literature
1. Gori, IEEE Trans.Pattern Analysis and Machine Intelligence, vol. 14, pp. 76-85, 1992
2. Magoulas G.D., Vrahatis M.N. and Androulakis G.S., On the alleviation of the problem of local minima in back-propagation, Nonlinear Analysis: Theory, Methods and Applications, 30, 4545-4550, 1997.
3. Parsopoulos K.E. , Plagianakos V.P. , Magoulas G.D. and Vrahatis M.N., Objective function ``stretching'' to alleviate convergence to local minima, Nonlinear Analysis: Theory, Methods and Applications, vol. 47, 3419-3424, 2001.
Search keywords: local minima, global search. global optimisation, backpropagation
3. The use of neural networks in medical imaging
Intelligent systems, particularly those for medical imaging, cover a major application area providing significant assistance in medical diagnosis. In most cases, the development of these systems leads to valuable diagnostic tools that may largely assist physicians in the identification of tumours or malignant formations by means of non-invasive or minimally invasive imaging procedures (e.g., computed tomography, ultrasonography, endoscopy, confocal microscopy, computed radiography, and magnetic resonance imaging). The aim of the project is to provide a survey of neural network-based intelligent systems in this area and to implement a neural network-based system for identifying abnormal tissue regions in endoscopic images.
Indicative literature
1. Pouloudi A. and Magoulas G.D. , Neural Expert Systems in Medical Image Interpretation: Development, Use and Ethical Issues, Journal of Intelligent Systems, vol.10, No. 5-6, 451-471, 2000.
2. Karkanis S., Magoulas G.D. and Theofanous N., Image Recognition and Neuronal Networks: Intelligent Systems for the Improvement of Imaging Information, Minimally Invasive Therapy and Allied Technologies, vol. 9, No. 3-4, 225-230, August 2000.
Relevant articles in journals as IEEE Transactions on Medical Imaging, IEEE Transactions on Biomedical Engineering, Computers in Biology and Medicine, Computer Methods and Programs in Biomedicine.
4. Texture recognition and classification using neural networks
Texture plays an important role for the characterisation of regions in digital images. Texture carries information about the micro-structure of the regions and the distribution of the grey levels. A scheme for the recognition of regions based on the texture information should be capable of encoding the properties of the texture using a number of parameters, named descriptors. These descriptors are usually represented by sets of statistical measures defining by this way the vectors to be used, consequently, for the recognition and can be very useful for recognition and classification. Usually, the approach followed has two major processing stages. The first stage consists of all the processing procedures that will be performed on an image to extract all the identifiable features, which will form the feature vectors. The second processing stage decides how to incorporate obtained from the first stage together with background and prior information, such as temporal data, relationships about features, etc., in order to draw inferences. The project will focus on the second stage of processing by investigating the use of neural networks in the recognition and classification of images by texture. A review of neural network models for texture recognition and classification should be provided and the performance of a neural network will be investigated by means of a software implementation.
Indicative literature
1. Panden, IEEE Tr. Pattern Analysis and Machine Intelligence, vol. 24, pp. 291-310, 1999; http://www.ux.his.no/~tranden/.
2. Karkanis S., Magoulas G.D. and Theofanous N., Image Recognition and Neuronal Networks: Intelligent Systems for the Improvement of Imaging Information, Minimally Invasive Therapy and Allied Technologies, vol. 9, No. 3-4, 225-230, August 2000.
Search keywords: texture, backpropagation, neural networks, feature extraction
5. Genetic algorithms for simulation optimisation
An effective way to analyse complex systems is to devise an abstract model, simplify the model in such a way that superfluous systems details are removed without loosing validity, and observe a simulation of the simplified model under the desired sets of experimental conditions. Evolutionary algorithms can be used as a means to optimise discrete-event simulation models. These methods do not require derivative-related information and are characterised by good global convergence properties. For example, Evolutionary algorithms can aid in model input parameter optimisation, which appears to be a traditional weakness of computer simulation [3-4], and in the building of auxiliary models (metamodels) for different analysis goals [5].
1. Abdurahiman V. and Paul R. (1994). Machine learning and simulation model specification, Simulation Practice and Theory, 2, 1-15.
2. Bratko I., Paul R. et al. (1993). Using machine learning techniques to interpret results from discrete event simulation. Proc. 15th Int. Conf. Information Technology, Croatia, 401-406.
3. Paul R. and Chanev T. (1998). Simulation optimisation using a genetic algorithm, Simulation Practice and Theory, 6, 601-611.
4. Paul R. and Chanev T. (1997). Optimising a complex discrete event simulations model using a genetic algorithm, Neural Computing and Applications, 6, 229-237.
5. Hurrion R. (2000). A sequential method for the development of visual interactive meta-simulation models using neural networks, J. Operational Research Society, 5, 712-719.
6. Magoulas, G.D., Eldabi, T., and Paul R.J., Adaptive Stochastic Search Methods for Parameter Adaptation of Simulation Models, in Proceedings of the IEEE International Symposium on Intelligent Systems, Varna, Bulgaria, Sept. 10-12, 2002, vol. 2, 22-26.
Relevant articles in journals such as Simulation, ACM transactions on modeling and computer simulation, The Journal of the Operational Research Society, European journal of operational research.
6. Knowledge-based neurocomputing in user-adaptive systems
In Neural expert systems or knowledge-based neurocomputing, as it is the name that is used now, the emphasis is on the use and representation of knowledge about an application within the neurocomputing paradigm. Despite the powerful processing capabilities of a neurocomputing system, explicit modelling of the knowledge represented by that system remains a major research topic. The aim of this project is to address this issue from various perspectives, present state-of-the-art in knowledge-based neurocomputing in an easily accessible form, and implement a knowledge-based neural network for user modelling.
Search keywords: hybrid systems, neuro-fuzzy systems, knowledge-based neurocmputing
Indicative literature
1. Stathacopoulou R., Magoulas G. D., Grigoriadou M. and Samarakou M., Neuro-fuzzy knowledge processing in intelligent learning environments for improved student diagnosis, Information Sciences, 170, 2, 273-307, 2005.
2. Frias-Martinez E., Magoulas G.D., Chen S., and Macredie R. Recent Soft Computing Approaches to User Modeling in Adaptive Hypermedia. In Paul De Bra, Wolfgang Nejdl (eds), Adaptive Hypermedia and adaptive web-based systems, Proceedings of 3rd Int Conf Adaptive Hypermedia-AH 2004, Eindhoven, The Netherlands, Aug. 2004, Lecture Notes in Computer Science, vol. 3137, Springer, 104-113.
3. Magoulas G. D. , Papanikolaou K. and Grigoriadou M., Towards a computationally intelligent lesson adaptation for a distance learning course, in Proceedings of the IEEE International Conference on Tools with Artificial Intelligence, 5-11, Chicago, November 1999.
4. Magoulas G.D. , Papanikolaou K.A., and Grigoriadou M. Neuro-fuzzy Synergism for Planning the Content in a Web-based Course, Informatica, vol. 25, 39-48, 2001.
Relevant articles in the journals IEEE Tr. Neural Networks, Neurocomputing, Neural Computing and Applications, Neural Networks, and in the ACM Digital Library.
7. Hybrid Genetic Algorithms in User-adaptive systems
GAs may be crossed with various problem-specific search techniques that exploit the global perspective of the GA and the convergence of the problem-specific technique. There are a number of ways to hybridise GAs and still maintain a fairly modular program structure. The aim of this project is to address this issue and to present state-of-the-art in hybrid GAs in an easily accessible form.
Search keywords: global search, global optimisation, Darwinian strategies, Lamarkian strategies,
Indicative literature
1. Zacharis N, and Panayiotopoulos T. (2001). Web search using genetic algorithms, IEEE Internet Computing, March-April, 18-26.
3. Frias-Martinez E., Magoulas G.D., Chen S., and Macredie R. Recent Soft Computing Approaches to User Modeling in Adaptive Hypermedia. In Paul De Bra, Wolfgang Nejdl (eds), Adaptive Hypermedia and adaptive web-based systems, Proceedings of 3rd Int Conf Adaptive Hypermedia-AH 2004, Eindhoven, The Netherlands, Aug. 2004, Lecture Notes in Computer Science, vol. 3137, Springer, 104-113.
Relevant articles in the journals IEEE Tr. Neural Networks, IEEE Tr. Evolutionary Computation, Neurocomputing, Neural Computing and Applications, Neural Networks, Natural Computing, IEEE Intelligent Systems, IEEE Internet Computing, and in the ACM Digital Library.
8. Methods for improving the generalisation
The ability to extrapolate the good performance from the training set to the test set is called generalisation. Improving generalisation performance of intelligent systems is a subject of considerable ongoing research and various methods have been proposed to this end, such as early stopping, cross validation, pruning, mixture of experts etc. This project can take different forms. For example, students may decide to provide a detailed review of methods belonging to one category, e.g. mixture or experts, or cross validation methods, or decide to cover two or more classes of methods. In any case the report must be within the word limits defined in the Study Guide.
Search keywords: early stopping with cross-validation or cross-validation, pruning methods, such as weight elimination and optimal brain damage, committee of networks or classifier ensembles, modular networks.
Indicative literature
1. Hassoun M., Fundamentals of Artificial Neural Networks, MIT Press, ISBN: 0-262-08239-X, 1995.
2. Hagan M. Demuth H., and Beale M., Neural Network Design, PWS Publishing Company, ISBN: 0-534-94332-2, 1996.
Relevant articles in the journals IEEE Tr. Neural Networks, Neurocomputing, Neural Computing and Applications, Neural Networks
9. Data clustering with graph theory
So far many clustering methods have been developed in the fields of psychology, statistics, machine learning and recently in molecular biology. The choice of which clustering technique to use for a given data set, is often very difficult. Most techniques require the user to define the number of clusters in advance, and those that do not, often require tuning of various parameters. Most clustering methods also require well-formed, convex clusters if they are to do a good job. Clustering with minimum spanning has been proposed as a promising alternative because MST-clustering is intuitive and easy to implement, and they seem to to work well on a variety of distributions. The project will focus on MST clustering algorithms and discuss their applications.
Search keywords: graph theory, clustering, minimum spanning trees.
10.Data Mining in Personalisation
Personalised web-based systems make use of intelligent techniques to provide individual users with content tailored to their needs. A number of data mining techniques can be used to support personalisation and they can be classified into supervised learning (i.e. classification) and unsupervised learning (i.e. clustering). These techniques can increase the effectiveness of personalisation, but they still have some limitations. The project can take different forms. For example, it can focus on either (a) the comparison between supervised learning and unsupervised learning, or (b) the differences between data mining and traditional techniques in the support of personalisation, or (c) the effectiveness and limitations of a particular technique in personalisation, e.g. neural network, genetic algorithms.
Search keywords: adaptive web-based systems, adaptive hypermedia, personalisation, data mining, machine learning, user modelling, neural networks, genetic algorithms.
Indicative literature
1. Adomavicius G. and Tuzhilin A. (2001). Using data mining methods to build customer profiles, Computer, February, 74-82.
2. Changchien S and Lu T (2001). Mining association rules procedure to support online recommendation by customers and product fragmentation. Expert Systems with application, 20, 4, 325-335.
3. Hui S., and Jha G. (2000). Data mining for customer service support, Information and Management, 38, 1-13.
11.Personalisation in e-commerce, e-health, digital libraries, e-learning, e-museums, TV, mobile computing (several projects)
The use of the Web has proliferated in businesses, libraries, and schools. Research and development on the semantic web indicates that we must direct web technologies towards developing relevant, and, to the extent possible, complete personal information spaces, visualization methods that enable users to process vast quantities of information, and interaction paradigms that facilitate human-computer communication. The proposed set of projects aim at the development of personalised information spaces in which both e-content and navigation as well as the user interface are adapted according to the individual information and navigation requirements based on a user model, which records user actions and changes in observed user navigational and interactive behaviour depending on the task. The projects require significant programming work (usually this includes Java or ASP, Javascript, Access or SQL but I’m open to suggestions) and their output will be the development and evaluation of proof-of-concept applications.
No comments:
Post a Comment