The Department of intelligent systems was founded in 1995 by joining two independent laboratories: Laboratory of artificial intelligence and Laboratory of language and speech technology. Basic activities cover research and development in the field of artificial intelligence and intelligent computer systems. Head: Prof. Dr. Ivanko Bratko In 1995, research activity was focused on the following topics: machine learning, inductive logic programming, multistrategy learning, evolutionary algorithms and knowledge synthesis using artificial intelligence methods for difficult problem domains (e.g. ecology, drug design, medicine and engineering). Research into language and speech has moved from speech synthesis for isolated words to pronouncing sentences, and developing computerised corpora. The department closely cooperates with the Laboratory of Artificial Intelligence at the Faculty of Computer Science and Information in Ljubljana. Some work has also been done in collaboration with the Faculties of Mechanical Engineering in Ljubljana and Maribor, the Hydrometeorological Institute of Slovenia, the Faculty of Civil Engineering and Geodesy - Department of Hydroengineering in Ljubljana, and other institutions. The department has well-established international links, many of them through participation in the following European projects and networks of excellence: - Esprit III "Network of Excellence in Computational Logic", - Esprit III "Network of Excellence in Machine Learning", - Esprit III Basic Research Action 6020 "Inductive Logic Programming", - Tempus project "Ecological Modelling and Water Resources Management", - PECO92 "Inductive Logic Programming Pan-European Scientific Network" (coordination of the project), - COST 233 project "Prosody in synthetic speech", - COPERNICUS project MULTEXT-EAST (Multilingual Texts and Corpora for Eastern and Central European Languages), - COPERNICUS project TELRI (Trans-European Language Infrastructure). BASIC RESEARCH PROJECTS Machine learning is one of the most active areas in artificial intelligence. It is concerned with extracting regularities and general laws from specific training examples, i.e., data from a given problem domain. The project "Knowledge synthesis from data" aims to improve current and develop new machine learning techniques that will achieve better classification accuracy and generate meaningful symbolic concept descriptions that will be accepted as new elements of knowledge by experts in demanding real-life domains. Within the project, special attention is devoted to the techniques of inductive logic programming (ILP) and multi-strategy learning. The techniques developed are being applied in a variety of problem domains, such as system control, the analysis of influence of physical and chemical parameters on river organisms, biological classification of river water quality, modelling algal growth in stagnant waters, modelling structure-activity relationships for drugs, finite element mesh design, modelling diffusion processed in thin layers and control of machining processes. A variety of applications is envisaged for the methodology of behaviour cloning of human operators that control dynamic systems, which is also being developed within the project. Inductive logic programming, compared to propositional learning, enables us to solve a broader class of problems, because it uses more expressive language for concept description and also utilises background knowledge. However, most existing ILP systems do not have adequately developed mechanisms for dealing with real numbers and with incomplete data. Therefore we developed the system FORS (First Order Regression System). FORS can induce concepts described in first order logic which incorporate continuous (real-valued) variables, makes use of background knowledge in intentional form, models dynamic systems and comprises mechanisms for noise handling. The ILP system DINUS was developed, which transforms ILP problems into attribute-value form and allows for the introduction of new determinate variables in the learned clauses. This approach allows efficient learning of a restricted class of logic programs, even in the presence of class noise. The attribute-value problems generated by DINUS typically have a large number of attributes: this problem was also addressed within the project through an algorithm for eliminating irrelevant attributes. In the alternative, non-monotonic semantics of ILP, we treated problems that arise from the closed world assumption and proposed the removal of this assumption. In the ILP system CLAUDIEN, developed in Leuven (Belgium), we added the facility of treating linear equations in the learned clauses in a manner similar to that of the machine discovery system LAGRANGE, developed in Ljubljana. The GOLDHORN system incorporates several improvements to LAGRANGE which are important for handling real data. GOLDHORN has been successfully used to model algal growth in the Lagoon of Venice and diffusion processes in thin layers. The GOLDHORN and LAGRANGE systems attracted wide attention from researchers in other fields, since they provide automatic structure synthesis of differential equations from measured data. FORS was successfully used in synthetic and real-life domains, where all improvements incorporated in the system, proved to be useful. In all real-life domains, linear regression improved the expressive power of the induced models. In experiments with modelling water behaviour in a surge tank, FORS proved that it can successfully handle time series. In the domain of steel grinding, the results of learning enabled the expert to grasp some additional properties of the process, which were not revealed by classical statistical methods. Figure 1: In the electrical discharge machining process, the workpiece surface is machined by electrical discharges occurring in the gap between two electrodes - the tool and the workpiece. The gap is continuously flushed by the third element, the dielectricum. One of the most promising FORS applications was modelling the operator behaviour in an electrical discharge machining process. In cooperation with the Faculty of Mechanical Engineering of Ljubljana Universitya data acquisition environment which enabled us to monitor and record crucial process parameters, as well as the operator's control actions. Models were induced which, according to the expert, capture the main behaviour patterns of the operator and can serve as a basis for automatic process control. During the knowledge acquisition process, several important guidelines emerged, mainly concerning the interaction with the domain expert. They confirm that comprehensibility of the induced models plays an important role in the process of behaviour cloning, and that, for successful modelling of the expert behaviour, one can not rely solely on interviews with an expert. The machine learning approach to human skill reconstruction was compared to the approach based on human introspection. Container crane control and inverted pendulum control served as testing domains. In both domains, control rules obtained by introspection shared some structural characteristics, showing goals, subgoals, task decomposition and control loops. These elements can provide useful guidelines for automatic skill reconstruction. The expert system for control of rolling mill emulsion in the Sendzimir rolling mill at Acroni Jesenice Steel Works was further modified. The system integrated four single systems and is the major industrial AI application in Slovenia. It has been in regular use for around 5 years. The GOLDING system for combining several single machine learning systems was developed. It enables easy combination of several systems at different layers. Theoretical models for combining different systems were further analysed. All results, theoretical as well as practical, strongly indicate that, in order to achieve good results in real-life circumstances, several systems have to be combined or integrated together. In 1995, multistrategy-learning research was extended into analysis of differences between humans and computers. New understanding in this field could greatly enhance knowledge about ourselves and computers. In addition, more advanced and more user-friendly computer systems could be designed. One of the major events in this area was the publication of a special issue of the Informatica journal entitled Mind<>Computer; Were Dreyfus and Winograd right? with papers written by Dreyfus, Winograd, Agre, Michie etc. Research in the project "Analysis of clinical data-bases and synthesis of medical knowledge" focused on: (1) Development of analytical methods for clinical data-bases. The methods were based on inductive learning, clustering, fuzzy sets, Bayesian classifiers, causal networks, genetic algorithms, and discrimination and regression analysis techniques. These techniques were then used in several medical domains. In addition to the standard ones, we analysed new data from new medical domains that were obtained through cooperation with local specialists. Among these domains was post head-injury analysis. (2) Development of specialised methods for modelling and acquisition of medical knowledge and their use for medical decision making support in several medical domains. These included: 3D reconstruction and visualisation of brain blood vessels, single nerve fibre models, estimation of the compatibility of stress and strains at work, and decision support in cases of hospital infections. As part of this project, we organised a workshop on ``Computer-Aided Data Analysis in Medicine'', CADAM-95. The workshop was organised in Bled on 27.- 28.11.1995 and was attended by 50 researchers and medical experts. Within the project "Evolutionary Computing in Optimisation and System Identification", we implemented and tested a machine discovery system based on genetic programming, and applied a genetic algorithm in the development of a signal interpretation method for gas-liquid flow measurements. The proposed machine discovery system employs genetic programming as search heuristics to generate equations from experimentally obtained numerical data. The goal is to minimise both the error measure and the complexity of the generated equations. Using this approach, a number of models of test dynamic systems were reconstructed from data obtained through simulation. The proposed system improves existing machine discovery techniques in two respects: it exhibits higher expressive power and allows for the application of background knowledge. In cooperation with the Laboratory for Fluid Dynamics and Thermodynamics at the Faculty of Mechanical Engineering, University of Ljubljana, we developed a new procedure to interpret probe signals detected in gas- liquid flow. The approach is based on acquisition of expert skill and tuning of procedure parameters via genetic algorithms. It has been applied in a study of flow regimes in a trickle-bed reactor under laboratory conditions. DEVELOPMENT AND APPLICATIONS A project "Integrated Information Systems for Employment Optimisation in Slovenia" was prepared in cooperation with the Employment Office of Slovenia. Within the project, an internet server for the Office was installed and home pages presenting the activities of the Office were prepared. We also installed a searchable database of current jobs available in Slovenia on the Internet server. Also in preparation is a database of both employers and employees. Towards the end of the project, an intelligent agent for employment tasks will be implemented, performing more advanced and user-friendly functions. The Internet address for job positions is http://www-ai.ijs.si/RZZ_doc/pd/rzzservy.html. In the development project "Computer-based Information System for Supporting Medical Therapeutic Activities", which takes place in collaboration with Infonet, d.o.o., Kranj, we developed a prototype information system for supporting nosocomial infection therapies. The prototype supports various analyses related to the effectiveness of antibiotics and bacteria resistance to antibiotics. We tested the prototype on a database of nosocomial infections made available for this purpose by the General Hospital in Jesenice. A complete version of the system will be implemented for MS Windows in the first half of 1996. Work on synthesising Slovene speech mainly consisted of improving the synthesis process by going from phoneme-based to diphone-based synthesis; upgrading the pronunciation of words to pronouncing texts by considering syntactic information is now in the experimental stage. In the area of computational morphology of Slovene, the system ALE-RA was developed for formal, computational descriptions of the morpho-phonology of verbs, upgrading the ALE system (Attribute Logic Engine) with realisational morphology and one-level phonology. Work on language corpora within the Multext-East project produced morpho-syntactic specifications of the lexicon for the corpus of Slovene texts now being built. Work on finite constraint networks was directed towards analysing the decomposition of hyper-graphs and the structure of the relation of semantic consequence, while work on formal self-reference proceeded to develop the idea of reflexive sequences of theories. In collaboration with other electronics and information technology departments at the Institute, we participated in the project "Establishing continuing education in the field of information technologies". The project was financed by the Institute in order to investigate relevant educational models in other countries, and to perform a feasibility study based on data from Slovenian enterprises. To evaluate the idea, several seminars and workshops were also prepared in the framework of the project.