Teaching Learners with Restricted Mind Changes
Within learning theory teaching has been studied in various ways. In a common variant the teacher has to teach all learners that are restricted to output only consistent hypotheses. The complexity of teaching is then measured by the maximum number of mistakes a consistent learner can make until successful learning. This is equivalent to the so-called teaching dimension. However, many interesting concept classes have an exponential teaching dimension and it is only meaningful to consider the teachability of finite concept classes.
A refined approach of teaching is proposed by introducing a neighborhood relation over all possible hypotheses. The learners are then restricted to choose a new hypothesis from the neighborhood of their current one. Teachers are either required to teach finitely or in the limit. Moreover, the variant that the teacher receives the current hypothesis of the learner as feedback is considered.
The new models are compared to existing ones and to one another in dependence of the neighborhood relations given. In particular, it is shown that feedback can be very helpful. Moreover, within the new model one can also study the teachability of infinite concept classes with potentially infinite concepts such as languages. Finally, it is shown that in our model teachability and learnability can be rather different.
KeywordsBoolean Function Concept Class Target Concept Teaching Time Neighborhood Relation
Unable to display preview. Download preview PDF.
- 1.Angluin, D.: Queries and concept learning. Machine Learning 2(4), 319–342 (1988)Google Scholar
- 6.Balbach, F.J., Zeugmann, T.: Teaching Learners that can only Perform Restricted Mind Changes, TCS Technical Report, Series A, TCS-TR-A-05-5, Division of Computer Science, Hokkaido University, July 18 (2005)Google Scholar
- 8.Freivalds, R., Kinber, E.B., Wiehagen, R.: Learning from good examples. In: Lange, S., Jantke, K.P. (eds.) GOSLER 1994. LNCS (LNAI), vol. 961, pp. 49–62. Springer, Heidelberg (1995)Google Scholar
- 12.Goldman, S.A., Sloan, R.H.: The power of self-directed learning. Machine Learning 14(3), 271–294 (1994)Google Scholar
- 14.Jain, S., Lange, S., Nessel, J.: Learning of r.e. languages from good examples. In: Li, M. (ed.) ALT 1997. LNCS (LNAI), vol. 1316, pp. 32–47. Springer, Heidelberg (1997)Google Scholar