README

This directory contains ANSI C source code for performing exact
probabilistic inference on multiply-connected Bayesian belief
networks using the modified join tree algorithm (to be published).
Some code has been borrowed from Adam Galper's implementation of
the Lauritzen-Spiegelhalter algorithm.

Currently five programs are provided:

	1.   infer <ergo_network_file> [<# procs>]

		A command line interface for performing basic functions,
		(computing marginal distributions for uninstantiated
		nodes, computing the probability of observed
		evidence, etc.) on belief networks. Takes a network
		file in ergo format.

	2.   econvert < <netview_network_file> > <ergo_network_file>

		A filter that converts netview network files to
		ergo network files. Currently, there are only two
		supported formats.

	3.   process < <netview_case_file>

		Process case files in netview format.  The file includes
		commands in ascii to perform inference given evidence.

	4.   decompose <ergo_network_file> <max_parents> <ergo_network_file>

		Decompose noisy-OR in a la-Heckerman fashion.  Needs to be
		optimized further to show useful results.

	5.   sample <network> <numberCases>

		A program that generates a database of cases from the
		specified network. Current output format includes data
		useful for a belief-network learning algorithm like
		K2.

Source code for these programs is provided to demonstrate
calls to the network inference and management routines. Use as
you wish FOR RESEARCH PURPOSES ONLY; if you add substantive
functionality or fix bugs, please let me know
(alexvk@cs.stanford.edu).  Likewise, if you have any
suggestions on style, let me know.

Notice that the core inference functionality resides in the
libinfer.a library, created with "make libinfer". By linking
this library into your executable, you'll have access to the
full range of routines.  Examples of routine calls can be
found in file process.l or infer.c. Check out the Makefile for
more details; note that Makefile is used to generate makefile.

Part of this code and of this file has been extracted from Adam
Galper's (galper@camis.stanford.edu) original code for the
Lauritzen-Spiegelhalter algorithm implementation.  The code is
optimized for multiprocessor as well as uniprocessor execution by
enchancing data locality and removing redundant computations.  It
runs hundred times faster then the old code for some networks and
requires substantially less memory.

For the multiprocessor execution you need to have ANL macros for
your machine and set environment variables:

MACROM4F -- the m4 ANL macros file
MACROLIB -- the object library file for multiprocessor execution

For the uniprocesor execution you do not need comppar.* files.
Remove any instance of it from the Makefile and run make as
usual.

To start, type:

% make -f Makefile.in depend; make infer; infer Asia

I hope the menu is self-explainable.  Please let me know if you have
any questions.

Alexander V. Kozlov
alexvk@cs.stanford.edu
February 9, 1996