mbeMPE
Approximates the Most Probable Explanation in Bayesian networks.
Author
Description
The library contains the implementation of the Mini-Bucket Elimination algorithm for Bayesian MPE.
Input format
The belief network and the evidence need to be specified in the Ergo_file_format.
For some example problems please see our Repository.
Usage
The algorithm is invoked with three (if no evidence present) or four (if evidence present) arguments:
mbeMPE <networkFile> [<evidenceFile>] <parameterFile> <outputFile>
with the following meaning:
<networkFile> specifies the path to the network specification in Ergo_file_format.
<evidenceFile> (optional) specifies the path to the evidence specification according to the Ergo_file_format.
<parameterFile> specifies the path to the file containing custom parameters for the algorithm. See the example file that comes with the binary program file for syntax reference.
<outputFile> specifies the path to the file to which (the logarithm of) the probability of the most probable explanation is written.
Detailed step-by-step instructions are available here.
Parameters
Parameters for the algorithm, which can be specified within the parameter file:
h: (string) the heuristic to use for finding an variable elimination order by which to construct the AND/OR search space. The following values can be used:
minfill: to indicate the min-fill heuristic
hypergraph: to indicate the hypergraph partitioning heuristic
a: (integer) specifies which algorithm to run. The following values can be used:
- 2: the Mini-Bucket algorithm
ib: (integer) specifies the i-bound of MBE(i).
Download
A 32-bit Linux binary and an example parameter will be available soon
Get the 32-bit Windows binary and an example parameter file here.
If you are interested in obtaining the source code, please contact the program author.
References
[1] Mini-Buckets: A General Scheme For Generating Approximations In Automated Reasoning. Rina Dechter. In Fifteenth International Joint Conference of Artificial Intelligence (IJCAI97), Japan, 1997. Link