Daniel T. Halstead

IJCAI Conference 2007 Conference Paper

• Daniel T. Halstead
• Kenneth D. Forbus

A key issue in artificial intelligence lies in finding the amount of input detail needed to do successful learning. Too much detail causes overhead and makes learning prone to over-fitting. Too little detail and it may not be possible to learn anything at all. The issue is particularly relevant when the inputs are relational case descriptions, and a very expressive vocabulary may also lead to inconsistent representations. For example, in the Whodunit Problem, the task is to form hypotheses about the identity of the perpetrator of an event described using relational propositions. The training data consists of arbitrary relational descriptions of many other similar cases. In this paper, we examine the possibility of translating the case descriptions into an alternative vocabulary which has a reduced number of predicates and therefore produces more consistent case descriptions. We compare how the reduced vocabulary affects three different learning algorithms: exemplar-based analogy, prototype-based analogy, and association rule learning. We find that it has a positive effect on some algorithms and a negative effect on others, which gives us insight into all three algorithms and indicates when reduced vocabularies might be appropriate.

AAAI Conference 2005 Conference Paper

Transforming between Propositions and Features: Bridging the Gap

• Daniel T. Halstead

It is notoriously difficult to simultaneously deal with both probabilistic and structural representations in A. I. , particularly because probability necessitates a uniform representation of the training examples. In this paper, we show how to build fully-specified probabilistic models from arbitrary propositional case descriptions about terrorist activities. Our method facilitates both reasoning and learning. Our solution is to use structural analogy to build probabilistic generalizations about those cases. We use these generalizations as a framework for mapping the structural representations, which are well-suited for reasoning, into features, which are well-suited for learning, and back again. Finally, we demonstrate how probabilistic generalizations are an excellent bridge for joining reasoning and learning by using them to perform a traditional machine learning technique, Bayesian network modeling, over arbitrarily high order structural data about terrorist actions, and further, we discuss how this might be used to facilitate automatic knowledge acquisition.