UAI 2016 Inference Evaluation

Sparse Model Format

We use the .suai suffix to specify sparse networks.

The sparse format uses sparse tables to specify the tables. Its structure is same as the UAI model format, except the following two differences:

  • The preamble network type will say SPARSEMARKOV (or SPARSEBAYES) instead of MARKOV (or BAYES).

  • Each potential/function table is specified in a special way, as described below.

Sparse Function Tables

Consider a function f defined over X and Y. X has two values in its domain {0,1} while Y has three values in its domain: {0,1,2}.

X Y f(X,Y)
0 0 3.4
0 1 3.6
0 2 4.7
1 0 4.7
1 1 4.7
1 2 4.7

The function will be specified as follows:

4.7 2
 0 0 3.4
 0 1 3.6

The first line specifies that the default-value of the function is 4.7 and it has 2 entries that do not equal the default-value. The next two lines specify the non-default entries, followed by the corresponding potential value.

(Note that line breaks and empty lines are effectively just a whitespace, exactly like plain spaces “ ”. They are used here to improve readability.)