A Network Representation of the Multiprocess Dynamic Linear Model,
Abstract
Dempster 1 has characterized the dynamic linear model (DLM) as a probabilistic belief network, showing that recent algorithms for propagation of information in such networks generalize Kalman filtering, prediction and smoothing algorithms for the DLM. Recently the Bayesian network technology has been extended to model mixed discrete and continuous random variables using conditional Gaussian (CG) distributions 5 with analogous propagation schemes 6. This paper applies the theory of CG probability networks to characterize the multiprocess dynamic linear model (MPDLM) and its requisite computations in a unified way. The complexity of exact computations is determined and approximate methods are proposed.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 1992
- Accession Number
- ADP007127
Entities
People
- David Tritchler
- Sharon-lise Normand
Organizations
- Harvard Medical School