top of page

See also tech report (currently password protected).

​

 

Network format The network format is a simplified* version of the .net format used by by LibB (Friedman and Elidan). Alarm comes from the Bayesian Network Repository(See also Friedman et al. Challenge: Where is the impact of Bayesian networks in learning? IJCAI-97). Parameters are scaled by 100 to yield Dirichlet row distributions. This doesn't affect the samples, which are generated from a network with mean parameters. A network starts by defining variables, for example,

​

var HISTORY FALSE TRUE

​

defines variable HISTORY taking on two values. Then the conditional probability tables are defined, for example,

 

parents HISTORY LVFAILURE

FALSE 20.000000 5.000000

TRUE 8.000000 471.000000

​

The first variable is the child, and the following variables are parents. In this case P(HISTORY|LVFAILURE) where the first row is the distribution P(HISTORY|LVFAILURE=FALSE) and the second row is P(HISTORY|LVFAILURE=TRUE). The first value of each row corresponds to HISTORY=FALSE; the second value of each row corresponds to HISTORY=TRUE.

 

* The difference between LibB's format and ours is that we ignore lisp quotes and brackets, using linebreaks and spacing instead.

 

Query format

 

The following is an example of a query

​

q STROKEVOLUME HIGH

o HYPOVOLEMIA TRUE

o ERRCAUTER TRUE

​

Each line that starts with 'q' is a query variable, each line that starts with 'o' is an evidence variable. This query is

​

P(STROKEVOLUME = HIGH | HYPOVOLEMIA = TRUE, ERRCAUTER = TRUE)

Data Associated

With The Experiments

alarm.cmdline

[./aij-experiment_fast, sampling, net, alarm.net, query_root, queries/hc91/alarm.q, num_queries, 100, data_samples, 500, query_samples, 10000, output_root, results/data_samples_500_query_samples_10K/]

loq1
27453.837u 14.561s 7:38:18.79 99.8%0    0+0k 0+0io 25pf+0w

Anchor 2
bottom of page