Weight Learning via Pseudo Log Likelihood

somdeb commented 9 years ago

Weight Learning using Pseudo Log Likelihood. Inference is not needed. Only Count is needed.

somdeb commented 9 years ago

Issue #1 and Issue #6 needs to be resolved

tuanh118 commented 9 years ago

[x] Functionality: Count number of true groundings in a particular formula given a database (world), before and after flipping the value of a ground atom. WClause's ID?
[x] LogDouble explanation and usage. Why are formula weights LogDoubles?
[x] PredicateSymbol -> List of WClause
[x] flipAtom(PredicateSymbol, list or array of term values)
[x] The gradient can be negative -> cannot be added to weight.

somdeb commented 9 years ago

LogDouble: When multiplying a bunch of small numbers (or even large numbers) and storing them in computer the data type can underflow (or overflow). To avoid this we store the numbers in log space.
GrindingMill: Creates ground clauses by (partially) grounding a formula.
Negative Gradient: New Issue (#9) added
GroundAtom: In the GroundStore all groundings of a Predicate is stored in a one dimensional array. A new issue (#10) is added to provide a factory method as well as access method

tuanh118 commented 9 years ago

Exception thrown while running WebKB.

Exception in thread "main" java.lang.ArrayIndexOutOfBoundsException: 2
    at starlib.mln.store.internal.jt.JoinTreeInt.<init>(JoinTreeInt.java:94)
    at starlib.mln.store.GraphModBasedGroundStore.setUpInternalClauseStore(GraphModBasedGroundStore.java:192)
    at starlib.mln.store.GraphModBasedGroundStore.init(GraphModBasedGroundStore.java:170)
    at starlib.mln.store.GroundStoreFactory.createGraphModBasedGroundStore(GroundStoreFactory.java:13)
    at starlib.mln.store.GroundStoreFactory.createGraphModBasedGroundStore(GroundStoreFactory.java:26)
    at starlib.mln.learn.weight.PseudoLogLikelihoodBasedLearning.computeCounts(PseudoLogLikelihoodBasedLearning.java:65)
    at starlib.mln.learn.weight.PseudoLogLikelihoodBasedLearning.main(PseudoLogLikelihoodBasedLearning.java:193)

somdeb commented 9 years ago

Hi @tuanh118 I have fixed it. But it's taking too long in my mac to run. Please run it on a server.

-Somdeb

tuanh118 commented 9 years ago

Exception thrown after running webkb-0.txt for roughly 8 hours and 50 minutes:

Exception in thread "main" java.lang.reflect.InvocationTargetException
    at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
    at sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:57)
    at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
    at java.lang.reflect.Method.invoke(Method.java:606)
    at org.eclipse.jdt.internal.jarinjarloader.JarRsrcLoader.main(JarRsrcLoader.java:58)
Caused by: java.lang.OutOfMemoryError: Java heap space
    at starlib.mln.learn.weight.FormulaCounts.<init>(FormulaCounts.java:41)
    at starlib.mln.learn.weight.PseudoLogLikelihoodBasedLearning.computeCounts(PseudoLogLikelihoodBasedLearning.java:82)
    at starlib.mln.learn.weight.PseudoLogLikelihoodBasedLearning.main(PseudoLogLikelihoodBasedLearning.java:196)
    ... 5 more

The program seemed to not be able to store all the counts required.

somdeb / magician

Weight Learning via Pseudo Log Likelihood #4