Artificial intelligence and inteligent systems

The members of the group carry out research in the following fields: • machine learning and knowledge discovery; • genetic algorithms and search algorithms; • constraint programming and combinatorial optimization; • qualitative reasoning methods; • machine learning in biomedical informatics; • agent technologies and semantic web. The research is constantly motivated with practical uses and concrete applications. Selected scientific achievements for the year 2005: • Analysis and reconstruction of genetic networks on the basis of microarray profiles, published also in van Driessche, Demšar, Juvan, Zupan, et al. Epistasis analysis with global transcriptional phenotypes. Nature Genetics, May 2005, impact factor 24.69. • Intelligent visualization of data using machine learning, published in (Leban et al., VizRank: finding informative data projections in functional genomics by machine learning; and in Curk et al., Microarray data mining with visual programming) Bioinformatics, #1 ranking journal in computer science. • Argument-based machine learning: a new approach to machine learning where the expert in the domain being learned can comment the learning examples with his or her arguments (explaining selected details in the learning examples). • Genetic algorithm for the analysis of complex biological systems with EPR spectroscopy (two papers in leading journals), and a genetic algorithm for steel production optimization introduced into use in steel mills Acroni Jesenice (Slovenia) and Ruuki Steel (Finland). • Machine learning of qualitative models from numerical data with our original method Q2 with application in ecological domains (predicting ozone concentration and Savinja river flooding). • System for automatic discovery of abnormalities in scintigraphic images of skeletons; in collaboration with the Clinic for Nuclear Medicine in Ljubljana. • Contributions to technology of Slovenian language: automatic accentuation of Slovenian words, use of ontologies for interpretation, continuing development of system “GOVOREC”. International awards in 2005: • ECCAI award for the best doctoral dissertation in artificial intelligence in Europe in 2005: A. Jakulin, Machine Learning Based on Attribute Interactions. • First award in NIST (USA) competition of e-mail spam filters (A. Bratko and B. Filipič). Fields of applicative research: • medicine; • bioinformatics and functional genomics; • system identification and management; • environmental issues; • engineering applications (textile industry, mechanical engineering); • internet applications (e-mail spam filtering, intelligent browsers); • marketing and economy.

The research program Artificial Intelligence and Intelligent Systems contributes to a number of areas of Artificial Intelligence: machine learning and knowledge discovery in data, qualitative modelling, heuristic search, evolutionary computation, and intelligent tutoring systems. Contributions are of various types: theoretical results, methods and techniques for solving open problems in these areas, freely available implementations of developed methods, and applications in chosen application areas. Several methods and techniques are also of general interest to science outside Artificial Intelligence. These include methods for machine learning and knowledge discovery in data which have become very important part of infrastructure for research in other areas of science like biology, genetics, medicine etc. Results of this program that are of interest from this more general point of view include the following: the popular Orange system for machine learning and data analysis that is continuously being upgraded within this program; data fusion techniques from various data sources with matrix factorization; methods for explanation of classifications produced by induced classifiers, Argument-based machine learning (ABML) that enables standard machine learning approaches to take into account existing knowledge in the form of expert-provided arguments about concrete learning examples; ABML knowledge acquisition loop. The work in qualitative modelling enables a non-traditional way of modelling in science and engineering, that is modelling without numbers, but with qualitative relations instead. Often, only such qualitative relations are known in the problem domain. An approach called Padé to learning qualitative models was developed in this program that enables automatic acquisition of qualitative models from data. Concrete important results of this program include: • A method for data fusion with simultaneous matrix tri-factorisation that produced excellent results in genetic applications • Empirical demonstration that explanations based on attribute interactions are correct and intuitive for any model of learning; a new, better measure of success of classification was introduced • Argument-based machine learning (ABML) and a knowledge acquisition procedure “ABML loop” • A new approach to discovery of abstract concepts in robotic domains • Approaches to robot learning from sensory data • Padé, a new method for qualitative learning based on computing qualitative partial derivatives from numerical data; experimental applications of Padé to real problems like: bacterial infections, robot learning, analysis of wind noise in cars • An approach to robot programming with qualitative planning • Analysis of factors that effect chances of pathology in heuristic search when paradoxically more thorough search that requires more work produces worse results • The speeding up of evolutionary algorithm for multi criteria optimisation DEMO by a parallel implementation and use of substitute models of search space • A new approach that automatically finds patterns, sub-goals and stages of search for a given combinatorial problems; this enables a human to deal with otherwise non-tractable search spaces Some scientific results of this program are of foundational character: the introduction of research into the understanding of how the difficulty of problems for humans can be automatically assessed; automatic discovery of abstract concepts from data, that is concepts that have never been explicitly observed in measured data.

This research program is particularly important for the development of Slovenia because of its application oriented research and transfer of scientific achievements to practice in many application areas. Also, the program ensures the basis for good quality education in artificial intelligence at the university study level, including doctoral study. Many researchers participating in this program carry out applications in collaboration with users in many, very diverse areas of applications like: medicine, bioinformatics, industrial process control, ambient intelligence, robotics, ecology, support for cultural heritage, computer processing of Slovene language. Examples include concrete realised applications of AI methods in the study and discovery of biomarkers in cancer diseases (lung cancer for example), diagnosis of tremors in neurology, diagnosis of heart disorders, intelligent buildings (for example, building management with optimisation with respect to energy consumption and comfort, ensuring good quality living conditions for the elderly), optimisation of manufacturing processes in high quality steel production, control and optimisation of warehouses, systems for recommendation of films, synthesis and analysis of Slovene speech and computer understanding of Slovene text, support to preserving cultural heritage (for example reassembly of large sets of fragments of old paintings found at historical exploration sites). There are continuously appearing new potential application possibilities, such as development of robotic applications in unstructured environments – automated programming of robots in changing environment, demanding intelligent tutoring systems, such as intelligent programming tutor, etc. As an example of surprising diversity of AI applications we here mention in more detail a current application of of machine learning that enables automatic discovery of long term spatial and temporal regularities from data, either measured data or data produced through simulation models. This enables long term validation of numerical models (that otherwise typically suffice for short term predictions only), as well as the discovery of yet unknown laws in processes in nature, such as the movement of water masses in the Adriatic Sea. Currently planned applications, in collaboration with the Piran biological station, include the prediction of emergence and movement of jellyfish which affects the fishing and tourism in the Northern Adriatic Sea.