Check parameter space definition

[larskotthoff]

There are some errors that suggest that the parameter space is not correctly specified wrt to some conditional dependencies, e.g.

[WARN ] [PROCESS-ERR] java.lang.RuntimeException: Failed to set classifier options [-I, 8, -H, 0.442016595562837, -R, 5, -F, 4, -Q, -P, 100, -L, 1e50, -S, 1, -W, weka.classifiers.bayes.BayesNet, --, -D, -Q, weka.classifiers.bayes.net.search.local.SimulatedAnnealing] for classifier weka.classifiers.meta.LogitBoost: Weight pruning with resampling not allowed.

[larskotthoff]

Also check definitions for attribute selection. Conditional dependencies do not appear to be specified completely there.

[larskotthoff]

Certain attribute selection search methods disallow certain evaluation methods. There's currently no way to specify these conditional dependencies as each class (i.e. each evaluation/search method) is only considered in isolation.

The proper way to fix this would be to have a unified parameter space for attribute selection where the top-level parameter determines the algorithm to be run and everything else hangs off of that with conditionals. This will require a major redesign of AutoWEKA's internals.

[larskotthoff]

Upon closer inspection it turns out that there are some such dependencies with respect to classifiers as well, so really the only proper way would be to have a grand unified parameter space definition for everything.

[larskotthoff]

Examples of things that need to be addressed in the parameter space definition but can't be at the moment:

Training classifier (weka.classifiers.lazy.LWL [-U, 3, -A, weka.core.neighboursearch.LinearNNSearch, -W, weka.classifiers.rules.OneR , --, -B, 3]) failed: Classifier must be a WeightedInstancesHandler!
java.lang.IllegalArgumentException: Classifier must be a WeightedInstancesHandler!
    at weka.classifiers.lazy.LWL.buildClassifier(LWL.java:507)
    at autoweka.ClassifierRunner$BuilderThread.doWork(ClassifierRunner.java:399)
    at autoweka.WorkerThread.run(WorkerThread.java:26)


Training classifier (weka.classifiers.meta.RandomCommittee [-I, 10, -S, 1, -W, weka.classifiers.bayes.BayesNet, --, -Q, weka.classifiers.bayes.net.search.local.TAN]) failed: Base learner must implement Randomizable!
java.lang.IllegalArgumentException: Base learner must implement Randomizable!
    at weka.classifiers.meta.RandomCommittee.buildClassifier(RandomCommittee.java:150)
    at autoweka.ClassifierRunner$BuilderThread.doWork(ClassifierRunner.java:399)
    at autoweka.WorkerThread.run(WorkerThread.java:26)