Ivan Bratko#

Laudatio by Krzysztof R. Apt#


Prof Bratko is one of the world leading experts in machine learning. His earlier contributions concerned improved search in game playing, such as chess. Over the past 25 years he developed a number of methods and novel systems that employ various forms of learning in several domains. Here is the list of his main scientific achievements.

KARDIO (1984-87), and expert system for diagnosing of cardiac arrhythmias from the ECG signal (in cooperation with I. Mozetic and N. Lavrac, and a medical team). At that time "surface" knowledge and manual knowledge acquisition were the dominating ways of development of expert systems. The KARDIO approach was a major innovation in the use of “deep" and qualitative knowledge, and of machine learning. KARDIO contained a complete qualitative model of the electrical activity of the heart (represented in logic) which guaranteed the correctness of the diagnostic rules w.r.t. the deep model. This work was published, among other publications, as book "KARDIO: A Study in Deep and Qualitative Knowledge for Expert Systems", by I. Bratko, I. Mozetic and N. Lavrac (MIT Press 1989).

ABML approach to machine learning (Argument Based Machine Learning; M. Mozina, I. Bratko, J. Zabkar, Argument based machine learning. Artificial Intelligence. Vol. 171, pp. 922-937, 2007). In this approach to machine learning, a learning program accepts, in addition to the usual kind of learning data, expert-supplied arguments relevant to some of the given learning examples. For example, a medical doctor may annotate selected medical cases by adding arguments that partially explain these cases. These arguments enable an ABML learning algorithm to narrow its search space. As a result, the learning is faster, it typically achieves better classification accuracy, and learned theories are easier to interpret by experts.

Q2 learning (D. Suc, D. Vladusic, I. Bratko, Qualitatively faithful quantitative prediction. Artificial Intelligence, 2004, Vol. 158, no. 2, pp. 189-214). This was a new approach to learning from numerical data where the learning algorithm respects qualitative relations in the learning data. The qualitative relations can be automatically induced from the data, or they can be provided by a domain expert. Several advantages of this approach in comparison with pure quantitative methods of machine learning were experimentally demonstrated: induced theories typically offer better intuitive understanding of how the modelled system works, and they also enable better numerical predictive accuracy because qualitative constraints alleviate the effects of noise in learning data. These advantages have also been demonstrated in applications of Q2 learning to various domains of ecological and industrial modeling, and the modelling of human skill.

Contributions to the modeling with machine learning of human skill. A key idea there was the decomposition of the learning of system control into the learning of human’s qualitative strategies and the learning of system’s dynamics. These contributions were published in a number of papers, including: D. Suc, I. Bratko, Skill modeling through symbolic reconstruction of operator's trajectories. IEEE Trans. Syst. Man and Cybernetics, Part A, Syst. humans, Vol. 30 (2000), no. 6, pp. 617-624; I. Bratko, T. Urbancic, Control skill, machine learning and hand-crafting in control1er design. In: Furukawa, K., Michie, D., Muggleton, S. Machine Intelligence I 5 : Intelligent Agents. Oxford: Oxford University Press, 1999, pp. 130-153.

Mathematical analysis of minimax search pathology (I. Bratko, M. Gams, Error analysis of the minimax principle. In: Clarke, M.R.B. (ed.). Advances in Computer Chess. 3. Oxford: Pergamon Press, 1982, pp. 1-15. This analysis explained a challenging inconsistency between the observed benefits of deep minimax search in the practice of game playing, and earlier mathematical analyses that predicted exactly the opposite result. A similar explanation of this phenomenon was found independently at the same time by D. Beal, D. Nan and J. Pearl. These works were consistently referenced in further studies of minimax pathology. A much more recent contribution to this topic is the paper M. Lustrek, M. Gams, I. Bratko, Is real-valued minimax patho1ogical?, Artificial Intelligence. Vol. 170, pp. 620-642 (2006).

A method for tree pruning in learning of decision trees from noisy data with T. Niblett 1986, and later improved by introducing m-estimates in 1991 with B. Cestnik (B. Cestnik, I. Bratko, On estimating probabilities in tree pruning. Proc. EWSL’91, Porto, Portugal, March 6-8, 1991 (Lecture notes in artificial intelligence, 482). Berlin: Springer-Verlag, l99l, pp. 138-150.). The method later became known as "minimal-error pruning” or "Nib1ett-Bratko pr1ming" and was regularly referenced.

Any further pages in alphabetic order of their title as created by you.
#

Just click at "Create new page", then type a short title and click OK, then add information on the empty page presented to you (including maybe a picture from your harddisk or a pdf-file by using the "Upload" Button) and finally click at "Save".
...no Data available yet!

Imprint Privacy policy « This page (revision-4) was last updated on Wednesday, 30. March 2011, 11:10 by Kaiser Dana
  • operated by