Demand-Driven Concept Formation

  • Stefan Wrobel


With the discussion of KRT knowledge revision in the preceding chapter, we have finally assembled the three pillars upon which our model of computational concept formation is to be built:
  • In chapter 2, we have discussed a number of psychological results as hints to the requirements for concept representation, and the possible mechanisms for concept formation.

  • In chapter 3, we developed a logical representation language that meets most of the psychological requirements on concept representation, and showed that it has a well-defined semantics and is tractable.

  • In chapter 4 finally, we discussed MOBAL’s knowledge revision activities, which are to provide the context for concept formation.


Knowledge Base Concept Formation Inductive Logic Programming Predicate Variable Rule Schema 
These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.


Unable to display preview. Download preview PDF.

Unable to display preview. Download preview PDF.


  1. 1.
    In fact, within MOBAL, the use of RDT on higher levels is currently not supported. 2 Further restrictions of the hypothesis space are defined by the predicate topology, see [Kietz and Wrobel, 1992]Google Scholar
  2. 4.
    Care must be taken, however, to avoid overinstantiating rule schemata that literally contain the target concept.Google Scholar
  3. 5.
    This way of using RDT for CLT characterization was designed by J.U. Kietz, K. Morik, and the author, and implemented by J.U. Kietz, see the acknowledgments.Google Scholar
  4. 6.
    Before these criteria are applied, CLT also checks whether the new concept has turned out to be a synonym of an existing concept.Google Scholar
  5. 7.
    Without any loss in inferential power, that is. There might be a loss of efficiency because the “compiled” one-step rule may execute faster than the two rules that replace it. Due to the cost of indexing, however, the overall effect of removing a compiled rule could also be a speedup. This is known as the utility-problem of explanation-based learning [Minton, 1990].Google Scholar
  6. 8.
    This is true only if inference depth limits are not used or set high enough.Google Scholar
  7. 9.
    The relation of this method to human concept formation certainly warrants a deeper examination, e.g. with respect to its power of predicting empirically observable psychological effects. Such studies, however, are outside the scope of this work.Google Scholar
  8. 10.
    We are thankful to R. Quinlan for making FOIL available to other researchers.Google Scholar
  9. 11.
    In [Stahl, 1993b], necessary terms according to our definition are referred to as “useful”, and useful terms according to our definition are not examined.Google Scholar
  10. 12.
    See [Stahl, 1993a] for an excellent overview of other first-order constructive induction operators.Google Scholar

Copyright information

© Springer Science+Business Media Dordrecht 1994

Authors and Affiliations

  • Stefan Wrobel
    • 1
  1. 1.GMDSankt AugustinGermany

Personalised recommendations