I implement the same markov logic network in two different ways using the Alchemy2 software. They should yield to the same result smoking-out.mln and smoking.result, but that's not the case. Why?
I use an example from the Alchemy tutorial.
First way:
smoking-train.db:
Friends(Anna, Bob)
Friends(Bob, Anna)
Friends(Anna, Edward)
Friends(Edward, Anna)
Friends(Anna, Frank)
Friends(Frank, Anna)
Friends(Bob, Chris)
Friends(Chris, Bob)
Friends(Chris, Daniel)
Friends(Daniel, Chris)
Friends(Edward, Frank)
Friends(Frank, Edward)
Friends(Gary, Helen)
Friends(Helen, Gary)
Friends(Gary, Anna)
Friends(Anna, Gary)
Smokes(Anna)
Smokes(Edward)
Smokes(Frank)
Smokes(Gary)
Cancer(Anna)
Cancer(Edward)
smoking.mln:
// MLN for social networks section in tutorial
// Evidence
Friends(person, person)
// Some evidence, some query
Smokes(person)
// Query
Cancer(person)
// Rules
// If you smoke, you get cancer
Smokes(x) => Cancer(x)
// People with friends who smoke, also smoke
// and those with friends who don't smoke, don't smoke
Friends(x, y) => (Smokes(x) <=> Smokes(y))
And I run:
$ learnwts -d -i smoking.mln -o smoking-out.mln -t smoking-train.db -ne Smokes,Cancer -queryEvidence
$ infer -ms -i smoking-out.mln -r smoking.result -e smoking-test.db -q Smokes,Cancer -ow Smokes,Cancer -cw Friends
Notice that I use the -queryEvidence option, which means "all the groundings of query preds not in db are assumed false evidence.".
I get this result: smoking-out.mln:
//predicate declarations
Cancer(person)
Friends(person,person)
Smokes(person)
// 0.112381 Smokes(x) => Cancer(x)
0.112381 !Smokes(a1) v Cancer(a1)
// 1.02505 Friends(x,y) => (Smokes(x) <=> Smokes(y))
0.512526 !Friends(a1,a2) v Smokes(a1) v !Smokes(a2)
0.512526 !Friends(a1,a2) v Smokes(a2) v !Smokes(a1)
// 0 Friends(a1,a2)
0 Friends(a1,a2)
// 1.62975 Smokes(a1)
1.62975 Smokes(a1)
// 1.45472 Cancer(a1)
1.45472 Cancer(a1)
smoking.result:
Smokes(John) 0.930957
Smokes(Katherine) 0.916958
Smokes(Lars) 0.907959
Smokes(Michael) 0.977952
Cancer(Ivan) 0.820968
Cancer(John) 0.847965
Cancer(Katherine) 0.80497
Cancer(Lars) 0.831967
Cancer(Michael) 0.883962
Cancer(Nick) 0.849965
Now, the second way:
In the second way, I simulate the -queryEvidence option by adding the false evidence directly in the training database (and then, I don't use the -queryEvidence option).
smoking-train.db:
Friends(Anna, Bob)
Friends(Bob, Anna)
Friends(Anna, Edward)
Friends(Edward, Anna)
Friends(Anna, Frank)
Friends(Frank, Anna)
Friends(Bob, Chris)
Friends(Chris, Bob)
Friends(Chris, Daniel)
Friends(Daniel, Chris)
Friends(Edward, Frank)
Friends(Frank, Edward)
Friends(Gary, Helen)
Friends(Helen, Gary)
Friends(Gary, Anna)
Friends(Anna, Gary)
Smokes(Anna)
Smokes(Edward)
Smokes(Frank)
Smokes(Gary)
Cancer(Anna)
Cancer(Edward)
!Smokes(Bob)
!Smokes(Chris)
!Smokes(Daniel)
!Smokes(Helen)
!Cancer(Bob)
!Cancer(Frank)
!Cancer(Chris)
!Cancer(Daniel)
!Cancer(Gary)
!Cancer(Helen)
!Friends(Anna, Chris)
!Friends(Anna, Daniel)
!Friends(Anna, Helen)
!Friends(Bob, Edward)
!Friends(Bob, Frank)
!Friends(Bob, Daniel)
!Friends(Bob, Gary)
!Friends(Bob, Helen)
!Friends(Edward, Bob)
!Friends(Edward, Chris)
!Friends(Edward, Daniel)
!Friends(Edward, Gary)
!Friends(Edward, Helen)
!Friends(Frank, Bob)
!Friends(Frank, Chris)
!Friends(Frank, Daniel)
!Friends(Frank, Gary)
!Friends(Frank, Helen)
!Friends(Chris, Anna)
!Friends(Chris, Edward)
!Friends(Chris, Frank)
!Friends(Chris, Gary)
!Friends(Chris, Helen)
!Friends(Daniel, Anna)
!Friends(Daniel, Bob)
!Friends(Daniel, Edward)
!Friends(Daniel, Frank)
!Friends(Daniel, Gary)
!Friends(Daniel, Helen)
!Friends(Gary, Bob)
!Friends(Gary, Edward)
!Friends(Gary, Frank)
!Friends(Gary, Chris)
!Friends(Gary, Daniel)
!Friends(Helen, Anna)
!Friends(Helen, Bob)
!Friends(Helen, Edward)
!Friends(Helen, Frank)
!Friends(Helen, Chris)
!Friends(Helen, Daniel)
I use the same smoking.mln as before, and now I run learnwts without the -queryEvidenceoption:
$ learnwts -d -i smoking.mln -o smoking-out.mln -t smoking-train.db -ne Smokes,Cancer
$ infer -ms -i smoking-out.mln -r smoking.result -e smoking-test.db -q Smokes,Cancer -ow Smokes,Cancer -cw Friends
And I get a very different result: smoking-out.mln:
//predicate declarations
Cancer(person)
Friends(person,person)
Smokes(person)
// 0.934545 Smokes(x) => Cancer(x)
0.934545 !Smokes(a1) v Cancer(a1)
// 0.886065 Friends(x,y) => (Smokes(x) <=> Smokes(y))
0.443032 !Friends(a1,a2) v Smokes(a1) v !Smokes(a2)
0.443032 !Friends(a1,a2) v Smokes(a2) v !Smokes(a1)
// 0 Friends(a1,a2)
0 Friends(a1,a2)
// 0.583657 Smokes(a1)
0.583657 Smokes(a1)
// -1.3644 Cancer(a1)
-1.3644 Cancer(a1)
smoking.result:
Smokes(John) 0.686981
Smokes(Katherine) 0.485001
Smokes(Lars) 0.475002
Smokes(Michael) 0.848965
Cancer(Ivan) 0.384012
Cancer(John) 0.318018
Cancer(Katherine) 0.257024
Cancer(Lars) 0.308019
Cancer(Michael) 0.353015
Cancer(Nick) 0.40001
So, why is this?
I should get the exact same smoking-out.mln and smoking.result, isn't it?
