There might be a bug in jtree_inf_engine and var_elim_inf_engine

[yzhang2016]

Hi,

I am not sure whether it is a bug in junction tree inference and variable elimination.

For the following BN model, when computing marginal for each node given node 1, I have tested Junction tree inference , variable elimination and pear_inf_engine. I found that the results by Junction tree and variable elimination are incorrect while pear_inf_engine is correct. I have no idea what happens in the Junition tree and the variable elimination engine .

I compute the marginal given node 1 for each other node manually. For node 8 and node 9, only pearl_infer_engine gets the right results.

I put the learned bnet model (matlab format) in the attachment.
Model configuration :

• all nodes are discrete
• node 1 has 8 states and others are binary

## Results : P(A8 = 2 | A1) P(A9 = 2 | A1)
Result : jtree

Result : var_elim

Result : pearl

## *Manual *:

code：
---------------------------　
jtree_Engine = jtree_inf_engine(bnet) ;
var_Engine = var_elim_inf_engine(bnet) ;
pearl_Engine = pearl_inf_engine(bnet) ;

jtree_conMarg = compMarginal (bnet,jtree_Engine) ;
var_conMarg_ = compMarginal (bnet,var_Engine) ;
pearl_conMarg_ = compMarginal (bnet,pearl_Engine) ;

---

function [conMarg] = compMarginal (bnet,engine)
%% First node : exp, 8 states, others : AUs, binary
numNodes = size(bnet.dag,1) ;
numAUs = numNodes - 1;
numCls = bnet.node_sizes(1) ;
evidence = cell(1,numNodes) ;
conMarg = zeros(numCls,numAUs) ;
for i = 1 : numCls
evidence{1} = i ;
engine = enter_evidence(engine,evidence) ;
for j = 1 : numAUs
temNodeInd = j + 1 ;
temMarg = marginal_nodes(engine,temNodeInd) ;
conMarg(i,j) = temMarg.T(2) ;
end
end

bnet_learnedModel.zip