Home | Publications | CS Home

Learning Novel Domains Through Curiosity and Conjecture


Paul Scott and Shaul Markovitch. Learning Novel Domains Through Curiosity and Conjecture. In Proceedings of International Joint Conference for Artificial Intelligence, 669-674 Detroit, Michigan, 1989.


Abstract

This paper describes DIDO, a system we have developed to carry out exploratory learning of unfamiliar domains without assistance from an external teacher. The program incorporates novel approaches to experience generation and representation generation. The experience generator uses a heuristic based on Shannon's uncertainty function to find informative examples. The representation generator makes conjectures on the basis of small amounts of evidence and retracts them if they prove to be wrong or useless. A number of experiments are described which demonstrate that the system can distribute its learning resources to steadily acquire a good representation of the whole of a domain, and that the system can readily acquire both disjunctive and conjunctive concepts even in the presence of noise.


Keywords: Active Learning, Relational Reinforcement Learning, Exploration
Secondary Keywords:
Online version:
Bibtex entry:
 @inproceedings{Scott:1989:LND,
  Author = {Paul Scott and Shaul Markovitch},
  Title = {Learning Novel Domains Through Curiosity and Conjecture},
  Year = {1989},
  Booktitle = {Proceedings of International Joint Conference for Artificial Intelligence},
  Pages = {669--674},
  Address = {Detroit, Michigan},
  Url = {http://www.cs.technion.ac.il/~shaulm/papers/pdf/Scott-Markovitch-ijcai1989.pdf},
  Keywords = {Active Learning, Relational Reinforcement Learning, Exploration},
  Secondary-keywords = {Exploratory Learning},
  Abstract = {
    This paper describes DIDO, a system we have developed to carry out
    exploratory learning of unfamiliar domains without assistance from
    an external teacher. The program incorporates novel approaches to
    experience generation and representation generation. The
    experience generator uses a heuristic based on Shannon's
    uncertainty function to find informative examples. The
    representation generator makes conjectures on the basis of small
    amounts of evidence and retracts them if they prove to be wrong or
    useless. A number of experiments are described which demonstrate
    that the system can distribute its learning resources to steadily
    acquire a good representation of the whole of a domain, and that
    the system can readily acquire both disjunctive and conjunctive
    concepts even in the presence of noise.
  }

  }