Knowledge Integration Project [1989-1991]

by Pavel Brazdil and Luís Torgo


The main goal of this project was to try to integrate the predictions of different, possibly conflicting sources into one better knowledge base. Our setup includes a set of possibly different learning systems learning with different data sets (again with possibly overlapping data).These learning systems learn without any communication among them. After this initial learning stage our INTEG system tries to qualify their resulting theories (by means of an independent testing set) and with this characterization builds a knowledge base that is supoused to include the better parts of the individualy learned theories.

In our experiments we have used two different learning systems, a decision tree and a rule learning system. We have used four individual learning agents. Two used as learning engine the decision tree and the others used the rule learner. Our results indicated a clear advantage of the resulting integrated theory when compared to each of the individual theories.

This research project is strongly related to a recent flood of research on combining multiple versions of unstable classifiers (like boosting, arcing and bagging). This relation was even more obvious in a learning system developped by Luís Torgo called SuperIDx [1990-1991]. This system internally performed a process of knowledge integration as described above. However the theories being integrated where all developed using the same algorithm (an ID3-like decision tree) although with different random samples of the given learning data set. The system SuperIDx is briefly described in Torgo & Kubat, 1991 (section 5.1).

These ideas of knowledge integration are also related to the topic of redundancy. In effect there is some redundancy on the individual theories learned by the different agents (both by repeated data as well as similar learned rules). Our results proved that this was benefic. This idea was followed further on in the learning system YAILS which learned a set of possibly redundant rules as a method of dealing with different kinds of noise.

Some references about this research line :

  • Brazdil, P.; Torgo, L. (1990) : Knowledge Acquisition via Knowledge Integration, in Current Trends in Knowledge Acquisition, Wielinga, B. et al (eds.), IOS Press.
    (Abstract)( 59201 bytes in format ".ps.gz") (The following internal report extends this paper)
  • Brazdil, P., Torgo, L.(1991): Knowledge Integration and Learning - LIACC, Machine Learning Group, Technical Report-91.1
    ( 64075 bytes in format ".ps.gz")(HTML version)
  • Brazdil, P.; Gams, M.; Sian, S.; Torgo, L.; Van de Velde,W. (1991): Learning in Distributed Systems and Multi-Agent Environments, in Machine Learning: EWSL-91 (European Working Session on Learning), Y. Kodratoff (Ed.), Lecture Notes in Artificial Intelligence, Springer-Verlag
    (Abstract)(42096 bytes in format ".ps.gz")(HTML version)
  • Torgo,L.; Kubat,M. (1991) : Knowledge Integration and Forgetting in Proceedings of the Checoslovak AI Conference in 1991, Prague.
    (Abstract) (59391 bytes in format ".ps.gz")(HTML version)