programods

A Python package for representing and querying Probabilistic Graphical Models.

GitHub repository:

https://github.com/thalespaiva/programods

Core Philosophy

• Code that is easy to read is favored over efficient code
• It should help students to visualize and understand Probabilistic Models concepts
• The dificulties of implementations should be clear, and not delegated to another Python library
• No use of other libraries for probablistic graphical models (a corolary of the above, but individually important)

Supported Probabilistic Models

Currently the following are implemented:

• Probabilistic Logic
• Bayesian Networks
• Markov Networks

Examples

Bayesian Networks

We can load the asia network from a BIF file with:

asia = BayesNet.init_from_bif_file('examples/bayesnet/asia/asia.bif')

To see the network structure, we can call

asia.draw('asia.png')

which creates the following PNG image:

For each node in the above Digraph, there is a variable with the same name. For example, the variable associated with the lung node can be accessed by asia['lung'], which is an instance of Variable.

>>> asia['lung']
Variable<lung>

>>> print(asia['lung'])
[V] Name : lung
    Dom  : ['yes', 'no']

The local probabilities are indexed by their main varible name. So, to print the local probability of the dysp variable, one can call:

>>> print(asia.local_probs['lung'])
[+] LocalProbability(dysp|bronc,either)
[ ] Scope: dysp,bronc,either
[ ] yes,yes,yes |  0.9000 
[ ] yes,yes,no  |  0.8000 
[ ] yes,no,yes  |  0.7000 
[ ]  yes,no,no  |  0.1000 
[ ] no,yes,yes  |  0.1000 
[ ]  no,yes,no  |  0.2000 
[ ]  no,no,yes  |  0.3000 
[ ]   no,no,no  |  0.9000 
[.]

Additionally, we can generate the markdown table for that local probability:

asia.local_probs['dysp'].get_markdown_table()

Which would give us the code that generated the following table:

Local Probability (dysp|bronc,either)

bronc	either	dysp = yes	dysp = no
yes	yes	0.9000	0.1000
yes	no	0.8000	0.2000
no	yes	0.7000	0.3000
no	no	0.1000	0.9000

Here we have the tables of all the asia network local probabilities.

To perform a conjunctive query:

>>> asia.conjunctive_query(xray='yes', dysp='no')
0.0396199356

We can check if two sets of variables are d-separated:

>>> asia.is_d_separated(['asia', 'tub'], ['smoke', 'lung', 'bronc'])
True

And also, we can pass a set of observed variables:

>>> asia.is_d_separated(['asia', 'tub'], ['smoke', 'lung', 'bronc'], observed_set=['either'])
False

To explore d-separation over a network more deeply, it's nice to see some shiny and colorful graphs. This can be accomplished by calling:

asia.draw_reachable_via_active_trails('.../reachable.png', ['asia', 'tub'], observed_set=['either'])

which produces the graph below. It's like in the poem

Unreachables are red,
Source variables are blue.
Grey variables are sad,
Because of what we saw them do.
A v-structure makes them mad,
They turn green some that were red,
And red some that were green, too.