2106-Wuillemin-BayesianNetwork

PyAgrum cheatsheet

Useful links :

aGrUM/pyAgrum's website

pyAgrum's documentation

aGrUM's gitlab

Import

Import pyAgrum

import pyagrum as gum

Import notebook specific library

import pyAgrum.lib.notebook as gnb

In/Out

Load a BN

bn = gum.loadBN({path})

Save a BN

gum.saveBN(bn,"file.bif")

Basic Manipulation

BN generation

bn = gum.BayesNet("Your BN name")

Fast BN generation

Arcs, number of modalities and default number of modalities can be directly specified

bn = gum.fastBN("a[2]->b<-c",3)

Labels can be directly specified

bn = gum.fastBN('A->B[1,3]<-C{yes|No}->D[2,4]<-E[1,2.5,3.9]',6)

Random variable generation

LabelizedVariable is a discrete random variable with a customizable sequence of labels.

va = gum.LabelizedVariable('a','a labelized variable',3)

RangeVariable represents a variable with a range of integers as domain.

vi = gum.gum.RangeVariable('I','I in [4,10]',4,10)

DiscretizedVariable is a discrete random variable with a set of ticks defining intervalls.

vX=gum.DiscretizedVariable('X','X has been discretized')
vX.addTick(1).addTick(2).addTick(3).addTick(3.1415)

CPT generation

Randomly generate CPT for a given structure or for a given node in a given structure.

bn.generateCPTs()

bn.generateCPT({var_id|var_name})

Adding/removing random variables

bn = gum.BayesNet('my BayesNet')
va = gum.LabelizedVariable('a','a labelized variable',3)
vb = gum.LabelizedVariable('b','another labelized variable',3)
vc = gum.LabelizedVariable('c','a third labelized variable',3)
bn.add(va)
bn.add(vb)
bn.add(vc)

bn.erase('c')

Arc adding

bn.addArc('a','b')

Arc reversing

bn.reverseArc('a','b')

Arc removing

bn.eraseArc('b','a')

CPT definition

Once the network topology is constructed, we must initialize the conditional probability tables (CPT) distributions. Each CPT is considered as a Potential object in pyAgrum. There are several ways to fill such an object.

Consider the following BN

bn=gum.fastBN("c->r->w<-s<-c")

low-level approach

bn.cpt('s').fillWith([0.5,0.5,0.9,0.1])

Using the order of variable

bn.cpt("s")[:]=[[0.5,0.5],[0.9,0.1]]

Then $P(S|C=0)=[0.5,0.5]$ and $P(S|C=1)=[0.9,0.1]$.

using a dict

bn.cpt("s")[{'c':0}]=[0.5,0.5]
bn.cpt("s")[{'c':1}]=[0.9,0.1]

Visualization

Consider the following BN :

bn = gum.fastBN("a->b<-c",3)
cpt = bn.cpt("a")

Show BN

gnb.showBN(bn)

Show CPT

gnb.showPotential(cpt)

Show posterior with or without evidence

gnb.showPosterior(bn,target="b",evs={})
gnb.showPosterior(bn,target="b",evs={'a':1,'c':2})

Show multiple informations at the same time

gnb.sideBySide(bn,cpt)

Inference

Consider the following BN :

bn = gum.fastBN("a->b<-c",3)

Exact Inference

ie = gum.LazyPropagation(bn)
ie.addEvidence('b',1)
ie.posterior('b')

Approximate Inference

ie = gum.LoopyBeliefPropagation(bn)
ie.setEpsilon(1e-7)
ie.setMaxIter(1000)
ie.setMaxTime(10)
ie.addEvidence('b',1)
ie.posterior('b')

Generator

Consider the following BN :

bn = gum.fastBN("a->b<-c",3)

CSV generation

gum.generateCSV(bn,"database.csv",50000)

Learning

Consider the following BN :

bn = gum.fastBN("a->b<-c",3)

BN learner initialization :

learner=gum.BNLearner("database.csv",bn)

Parameter learning

bn2=learner.learnParameters()

Structural learning

learner.useLocalSearchWithTabuList() ### Or any other algorithm
bn2=learner.learnBN()

Add prior knowledge about the structure

learner.setMaxIndegree(1)
learner.addMandatoryArc('a','b')

Change score

learner.useScoreLog2Likelihood() ### Or other scores