UAI model files and solutions
These spreadsheets contain links to UAI files used in past UAI exact and approximate inference competitions,
along with many solutions and other statistics and information. Use the tabs at the bottom to move between
different problem groups.
If you have any problems you would like to donate, please email Alex Ihler (ihler@ics.uci.edu).
For summation inference solutions, we used a basic conditioned variable elimination or junction tree algorithm;
a linux binary implementation can be downloaded here.
No warrantees, express or implied; use at your own risk, etc.
An example execution is, e.g.
./exact -f $1 -e $1.evid -t100 -o1000 --order-rand 3 --memory 20000 -T PR
and you can add --order-file $1.ord
to use a pre-identified elimination order (or save one once found).
Statistics listed in the spreadsheet include:
- Name:
- The instance name, along with a link to the file
- nEvid:
- The number of evidence variables, along with a link to the evidence file (if nonzero)
- Width:
- The induced width of the file for a (linked) elimination ordering
- Depth:
- The depth of the pseudo-tree for the given elimination ordering
- logSS:
- State space size (storage size of the junction tree), log base 10 (so, 1GB of doubles is ~ 8.1, while 100GB ~ 10.1)
- # var:
- Number of variables in the model
- d-avg:
- Average cardinality (# of states) of the variables
- d-max:
- Largest cardinality (# of states) of any variable
- # factor:
- Number of factors in the model
- s-avg:
- Average scope size (# of arguments) of the factors
- s-max:
- Largest scope size (# of arguments) of any factor
- 0-avg:
- Average fraction of zeros (forbidden configurations) per factor; a loose measure of sparsity of the state space
- MPE:
- Value of the MPE/MAP solution (log10), along with a link to the configuration.
- MPLP:
- MPE/MAP upper bound given by a basic MPLP implementation
- PR:
- Value of the log partition function (log10)
- LBP:
- Log partition function estimate given by a basic Loopy BP implementation