Abstract
Consider a doctor with a knowledge base KB consisting of first-order information (such as “All patients with hepatitis have jaun- dice”), statistical information (such as “80have hepatitis”), and default information (such as “patients with pneumonia typically have fever”). The doctor may want to make decisions regarding a particular patient, using the KB in some principled way. To do this, it is often useful for the doctor to assign a numerical “degree of belief” to measure the strength of her belief in a given statement A. I focus on one principled method for doing so. The method, called the random worlds method, is a natu- ral one: For any given domain size N, we can look at the proportion of models satisfying A among models of size N satisfying KB. If we don’t know the domain size N, but know that it is large, we can approximate the degree of belief in A given KB by taking the limit of this fraction as N goes to infinity. In many cases that arise in practice, the answers we get using this method can be shown to match heuristic assumptions made in many standard AI systems. I also show that when the language is restricted to unary predicates (for example, symptoms and diseases, but not relations such as “Taller than”), the answer provided by the random worlds method can often be computed using maximum entropy. On the other hand, if the language includes binary predicates, all connections to maximum entropy seem to disappear. Moreover, almost all the questions one might want to ask can be shown to be highly undecidable. I conclude with some general discussion of the problem of finding reason- able methods to do inductive reasoning of the sort considered here, and the relevance of these ideas to data mining and knowledge discovery. The talk covers joint work with Fahiem Bacchus, Adam Grove and Daphne Koller [1][2].
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
References
F. Bacchus, A. J. Grove, J. Y. Halpern, and D. Koller, From statistical knowledge bases to degrees of belief, Artificial Intelligence 87:1-2, 1996, pp. 75–143.
F. Bacchus, A. J. Grove, J. Y. Halpern, and D. Koller, From statistics to belief, Proceedings of AAAI-92 (Proceedings of the Tenth National Conference on Artificial Intelligence, 1992, pp. 602–608.
A. J. Grove, J. Y. Halpern, and D. Koller, Random worlds and maximum entropy, Journal of AI Research 2, 1994, pp. 33–88.
A. J. Grove, J. Y. Halpern, and D. Koller, Asymptotic conditional probabilities: the unary case, SIAM Journal on Computing, 25:1, pp. 1–51, 1996.
A. J. Grove, J. Y. Halpern, and D. Koller,Asymptotic conditional probabilities: the non-unary case, Journal of Symbolic Logic, 61:1, 1996, pp. 250–275.
Author information
Authors and Affiliations
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2000 Springer-Verlag Berlin Heidelberg
About this paper
Cite this paper
Halpern, J.Y. (2000). Degrees of belief, random worlds, and maximum entropy. In: Arikawa, S., Morishita, S. (eds) Discovery Science. DS 2000. Lecture Notes in Computer Science(), vol 1967. Springer, Berlin, Heidelberg. https://doi.org/10.1007/3-540-44418-1_2
Download citation
DOI: https://doi.org/10.1007/3-540-44418-1_2
Published:
Publisher Name: Springer, Berlin, Heidelberg
Print ISBN: 978-3-540-41352-3
Online ISBN: 978-3-540-44418-3
eBook Packages: Springer Book Archive