(including the research of my PhD and MSc students)
Estimation of Distribution Algorithms in Time Series Analysis
This topic concerns Estimation of Distribution Algorithms, such as BOA, EBNA, ECGA, and their applications in time series analysis, especially in real-time systems. It focuses on adapting these algorithms to fast and robust computing, which is necessary for providing accurate and on-time expertises. Main issues addressed in the research are analysing and improving methods of building probability models, such as linkage learning or bayesian networks constructing. So far, two practical applications were considered: stock market data analysis and electric power prices and volumes analysis.
Dependency Mining and Dimensionality Reduction in Evolutionary Computation
This topic aims at improving evolution processes by dependency mining and dimensionality reduction. It focuses not only on the data preprocessing before the algorithm starts, but also on the data restructuring while the algorithm is running. A few kinds of dependencies are studied, linear and non-linear, using a variety of methods from simple principal component analyses to advanced generative topographic mappings.
Distributed Evolutionary Computation
In this project, a platform for distributed evolutionary computation, called wEvo, is developed. It provides a support for running a range of algorithms from classic Genetic Algorithms and Evolution Strategies to advanced Estimation of Distribution Algorithms. wEvo enables to solve problems with various objective functions over various search spaces with various representations. A prototype version of the platform was developed by 14 research students and researchers. Currently, it is used in time series analysis, biological and genetic simulations, and image processing. wEvo is available on the SourceForge site: wevo.sourceforge.net (still in the prototype version).
Decision Support Systems
Real-Time Evolutionary Decision Support Systems
This topic concerns applying evolutionary algorithms for constructing efficient real-time experts built on the basis of a set of specific expert rules, which analyse time series of recent data and propose advices for users. So far, applications focus on financial data analysis, where the expert rules come from practictioner knowledge and are frequently applied by financial analysts or market investors, and the experts built in this approach enables to combine these rules together to achieve much better performance.
Early Warning Systems based on Evolutionary Algorithms
This topic aims at proposing a new functionality of early warning for decision support systems, especially these concerning financial applications. Warnings help to focus the trader's attention on specific situations on the financial market. These situations relate to the rare circumstances where the trader should be alerted e.g. by exceptional raises or drops of stock prices, volatilities and market index.
Discovering Expert Rules using Neural Networks
This topic concerns an approach to extracting stock market trading rules directly from stock market data. Such trading rules are based on multi-layer perceptrons fed with values of technical indicators computed on historical stock quotations. Extracted trading rules may be applied in decision support systems, such as described above.
Performance Measures in Financial Decision Support Systems
In financial decison support systems, performance measures are applied for assessing and comparing different investment possibilities and for selecting the best ones. As comparing is usually performed on past period data, and the chosen solution is applied for future period data, performance measures are extremely important for the overall efficiency of such systems. There are a number of well-known performance measures coming from a few financial models and theories, such as the Sharpe ratio, the Sortino ratio or the Sterling ratio. However, in practice, no one of them is perfect, and each of them may be overperformed by other in some particular circumstances. Therefore, this project focuses on analysing these performance measures and constructing new measures as combinations of the original ones.
Portfolio Optimization using Evolutionary Algorithms
Portfolio optimization is a well-known financial problem consisting in finding optimal proportions of capital to invest in individual financial assets in order to maximize expected returns or minimize expected risks. When the risk is defined by the variance of returns, the problem can be solved using well-known classic models, such as the Markowitz's portfolio theory. However, other risk measures, such as the semivariance or the downside risk, lead to difficult optimization problems, which cannot be solved using analytical methods.
Intelligent Agents and Multi-Agent Systems
Evolutionary Multi-Agent Systems
This project combines multi-agent systems with some evolution mechanisms aiming at improving and learning agents. Evolutionary algorithms are used to watch the performance of individual agents, eliminate weak agents and create new efficient agents by a kind of replication of existing ones.
Financial Data Mining based on Intelligent Agents
In this project, a financial data-mining system, consisting of a number of independent and heteregonous intelligent agents, is developed. Each agent aims at analysing financial data using a different method. Combining various agents, enabling their communication and cooperation, leads to higher performance in data analysis.
Image Retrieval using Evolutionary Computation
This topic concerns image classification and comparison methods, which could be applied to retrieve images from large image databases. Main issues addressed are image preprocessing and representing in the database aiming at robust comparison and finding similarities between images. Although there are some popular methods for such problems coming from computer graphics, from simple ones based on statistics, to more advanced methods based on the Fourier analysis or the wavelet theory, they are either too simple to give significant results or too complex to be computed in a reasonable time. Therefore, a new evolutionary approach is proposed. It uses evolutionary algorithms to discover the most characteristic regions of the image, and then, it calculates some statistics for each of these regions to describe the image. Using such descriptions, various images could be compared in a reasonable time. In the project, practical research focuses on developing a mechanism for searching a large image database for images similar to a given one.
Image Classification and Object Recognition using Evolutionary Computation
This topic concerns a couple of issues related to content-based image classification. The first part of the study refers to image segmentation and contour detection. The second part concerns image classification and discovering optimal classification rules. So far, a few practical problems are under scrutiny, such as satellite image classification, medical images analysis, as well as biological and ecological images interpretation.
Text and Web Mining
Association Rules Discovering in Web-Mining
This topic concerns human-computer interaction and web usage mining, especially the analysis of user activities in large web services. It focuses on discovering typical sequences of user activities and associations among them. Such research aims at improving the structure of large web services by information rearranging and optimizing executions of most frequent tasks.
Data-Mining Techniques in Text Classification
This topic concerns text mining and information retrieval, especially the content-based text classification. It focuses on numerical text representation, where each message is represented by a vector of term frequencies, and on clustering of such vectors.