Hi All.
Unfortunately, I have a problem using Jointree in the folowing context:

from OpenBayes import MCMCEngine, JoinTree, learning, bayesnet
from copy import deepcopy
from time import time

cases = [{'x1':1,'x2':0,'x3':1,'x4':0, },
{'x1':1,'x2':1,'x3':0,'x4':0, },
{'x1':0,'x2':1,'x3':0,'x4':0,}]

# Create a new BNet with no edges
G2 = bayesnet.BNet('test BN')
x1,x2,x3,x4 = [G2.add_v(bayesnet.BVertex(nm,True,2)) for nm in 'x1 x2 x3 x4'.split()]
G2.InitDistributions()

x1.setDistributionParameters([0.5 ,0.5 ])
x2.setDistributionParameters([0.7 ,0.3 ])
x3.setDistributionParameters([0.5 ,0.5 ])
x4.setDistributionParameters([0.35,0.65])

# Learn the structure
struct_engine = learning.GreedyStructLearningEngine(G2)
struct_engine.StructLearning(cases)

G2=struct_engine.BNet

# Inference
ie=JoinTree(G2)
ie.SetObs({'x1':1})
print ie.Marginalise('x3')

# this results to:
#
#Traceback (most recent call last):
# File "testbayes1.py", line 24, in
# ie=JoinTree(G2)
# File "C:\Python25\Lib\site-packages\OpenBayes\inference.py", line 426, in __init__
# self.GlobalPropagation()
# File "C:\Python25\Lib\site-packages\OpenBayes\inference.py", line 514, in GlobalPropagation
# start.CollectEvidence()
# File "C:\Python25\Lib\site-packages\OpenBayes\inference.py", line 187, in CollectEvidence
# if not v.marked: v.CollectEvidence(self)
# File "C:\Python25\Lib\site-packages\OpenBayes\inference.py", line 189, in CollectEvidence
# if not X == None: self.MessagePass(X)
# File "C:\Python25\Lib\site-packages\OpenBayes\inference.py", line 168, in MessagePass
# newphiR = self.potential + e.potential # phiR = sum(X/R)phiX
# File "C:\Python25\lib\site-packages\OpenBayes\potentials.py", line 156, in __add__
# return self.Marginalise(var)
# File "C:\Python25\lib\site-packages\OpenBayes\potentials.py", line 119, in Marginalise
# return self.__class__(newnames,temp.shape,temp)
#AttributeError: 'float' object has no attribute 'shape'

The main idea is just that i want to structlearn a network and then do some inferences with it.

Hi thomas,

it seems like the ConnexeInferenceEngine does not work correctly or is not used at all...

The ConnexeInferenceEngine was created to treat problems like yours. It separates the BNet into independant compounds that are not connected between them, and creates a different inference engine for each one of them. Viewed from the user, it should be completely transparent. For example when you choose to marginalise a node, the corresponding inference engine (inside the ConnexeInference) that contains that node is used and the nodes that are not connected to that subset are not used at all.

For the moment I'm working on the Gaussian Distributions so I'll check your problem after that. If all goes well, probably next week I'll start working on your problem.

However, if you wish to participate to OpenBayes, you could try to solve this problem by yourself. In that case, I'll be glad to help you track down the bug and solve it.

Please tell me your plans

Kosta

Hello,

I think I fixed the bug. I have sent my code to Kosta. He will post it after he checked it.

François

Hello,

If anybody needs François code for SEM learning for not fully connected networks (using the ConnexeInferenceEngine) you can grab it here :

http://svn.berlios.de/wsvn/pybayes/branches/Learning/?rev=0&sc=0

Learningwanna.py should be used instead of learning.py

This will be included in the new version of OpenBayes.

Since I'm working on the new version for the moment, I have decided to stop correcting bugs in the old version, since the structure of OpenBayes is under modification and this would mean double work.

I'm sorry for any inconvenience, and I assure you I'll do my best for this new version to come out as soon as possible including all the features and bug fixes that you have requested.

kosta