def conditionalModel(bn, evs):
"""
create a new condtional bn from a bn and a instanticiation of some variable
:param bn: a bayesian network
:param evs: map of evidence
:return: a bayesian network
"""
newbn = gum.BayesNet(bn)
newevs = dict(evs)
for name in evs:
nid = newbn.idFromName(name)
for ch in newbn.children(nid):
# create the new cpt
q = newbn.cpt(ch) \
.extract({name: evs[name]}) \
.reorganize([v.name() for v in newbn.cpt(ch).variablesSequence() if v.name() != name])
# erase arc
newbn.eraseArc(nid, ch)
# update cpt
# todo : add a Potential::fillWithParamOF(Potential) in agrum
newbn.cpt(ch)[:] = q[:]
# remove evidence without parent
if len(newbn.parents(nid)) == 0:
newbn.erase(nid)
newevs.pop(name)
return newbn, newevs
def testWithInstantiation(self):
bn = gum.BayesNet()
id_list = []
self.fillBN(bn, id_list)
list3 = bn.cpt(id_list[3])
list3[:] = [[[1, 0], [0.1, 0.9]],
[[0.1, 0.9], [0.01, 0.99]]]
i = gum.Instantiation(list3)
list3.set(i, 0)
i.inc()
list3.set(i, 1)
self.assertListsAlmostEqual(list3[:],
[[[0, 1], [0.1, 0.9]],
[[0.1, 0.9], [0.01, 0.99]]])
self.assertListsAlmostEqual(list3[:], bn.cpt(id_list[3])[:])
def unsharpenedModel(bn, epsilon=1e-2):
"""
Modify the cpts of a BN in order to tend to uniform distributions
:param bn: a bayesian netwok
:param epsilon: a value that will be added every where before normalization
:return: the newBN
"""
if bn.minNonZeroParam() < epsilon:
newbn = gum.BayesNet(bn)
for k in newbn.ids():
# adding epsilon on all non-zero value, and normalize as CPT again
newbn.cpt(k)[:] = (newbn.cpt(k).isNonZeroMap().scale(epsilon) + newbn.cpt(k)).normalizeAsCPT()[:]
else:
newbn = bn
return newbn
