Adaptive Probabilistic Networks with Hidden Variables

  • Authors:
  • John Binder;Daphne Koller;Stuart Russell;Keiji Kanazawa

  • Affiliations:
  • Computer Science Division, University of California, Berkeley, CA 94720-1776. E-mail: binder@cs.berkeley.edu;Computer Science Department, Stanford University, Stanford, CA 94305-9010. E-mail: koller@cs.stanford.edu;Computer Science Division, University of California, Berkeley, CA 94720-1776. E-mail: russell@cs.berkeley.edu;Microsoft Corporation, One Microsoft Way, Redmond, WA 98052-6399. E-mail: keijik@microsoft.com

  • Venue:
  • Machine Learning - Special issue on learning with probabilistic representations
  • Year:
  • 1997

Quantified Score

Hi-index 0.01

Visualization

Abstract

Probabilistic networks (also known as Bayesian belief networks)allow a compact description of complex stochastic relationships amongseveral random variables. They are used widely for uncertain reasoning inartificial intelligence. In this paper, we investigate the problem oflearning probabilistic networks with known structure and hidden variables.This is an important problem, because structure is much easier to elicitfrom experts than numbers, and the world is rarely fully observable. Wepresent a gradient-based algorithm and show that the gradient can becomputed locally, using information that is available as a byproduct ofstandard inference algorithms for probabilistic networks. Our experimentalresults demonstrate that using prior knowledge about the structure, evenwith hidden variables, can significantly improve the learning rate ofprobabilistic networks. We extend the method to networks in which theconditional probability tables are described using a small number ofparameters. Examples include noisy-OR nodes and dynamic probabilisticnetworks. We show how this additional structure can be exploited by ouralgorithm to speed up the learning even further. We also outline anextension to hybrid networks, in which some of the nodestake on values in a continuous domain.