Iterated phantom induction: A little knowledge can go a long way

Mark Brodie, Gerald DeJong

Research output: Contribution to conferencePaperpeer-review

Abstract

We advance a knowledge-based learning method that augments conventional generalization to permit concept acquisition in failure domains. These are domains in which learning must proceed exclusively with failure examples that are relatively uninformative for conventional methods. A domain theory is used to explain and then systematically perturb the observed failures so that they can be treated as if they were positive training examples. The concept induced from these `phantom' examples is exercised in the world, yielding additional observations, and the process repeats. Surprisingly, an accurate concept can often be learned even if the phantom examples are themselves failures and the domain theory is only imprecise and approximate. We investigate the behavior of the method in a stylized air-hockey domain which demands a nonlinear decision concept. Learning is shown empirically to be robust in the face of degraded domain knowledge. An interpretation is advanced which indicates that the information available from a plausible qualitative domain theory is sufficient for robust successful learning.

Original languageEnglish (US)
Pages665-670
Number of pages6
StatePublished - 1998
EventProceedings of the 1998 15th National Conference on Artificial Intelligence, AAAI - Madison, WI, USA
Duration: Jul 26 1998Jul 30 1998

Other

OtherProceedings of the 1998 15th National Conference on Artificial Intelligence, AAAI
CityMadison, WI, USA
Period7/26/987/30/98

ASJC Scopus subject areas

  • Software
  • Artificial Intelligence

Fingerprint

Dive into the research topics of 'Iterated phantom induction: A little knowledge can go a long way'. Together they form a unique fingerprint.

Cite this