When porting to python the SumOfSquaredErrors.java file, you only loop through the output size and don't look at the rest of the examples. This doesn't seem like correct behavior.

https://github.com/pushkar/ABAGAIL/blob/master/src/shared/SumOfSquaresError.java#L22

I'm not certain but I think this might be the right way to do it
for (int i = 0; i < output.size(); i++) {
for (int j = 0; i < label.size(); i++) {
sum += (output.getContinuous(i) - label.getContinuous(j))
* (output.getContinuous(i) - label.getContinuous(j))
* example.getWeight(); // Not sure if weight should change or move out of the whole thing (as in multiply by weight at the end).
}

I just wanted to get your thoughts on this.

Is it possible to remove this line or specify a newer version?

Line 40 and 41 of build.xml should be:

    <target name="javadoc" depends="prepare" >
        <javadoc sourcepath="${src.dir}" destdir="${jdocs.dir}" additionalparam="subpackages"/>

the current version fails when run ant.

predicted = instance.getLabel().getContinuous() print "predicted ", predicted actual = networks[i].getOutputValues().get(0) print "actual ", actual
When I run that code, you see that predicted is either 0 or 1, and the actual is shown as a continuous value.

Since the data is loaded with all values less than 15 "washed" as a 1 and 0 if otherwise, these values are swapped in the code.

Not a big deal since the code is measuring absolute value of difference, but can be confusing

Any chance you understand what's happening on this line?
https://github.com/pushkar/ABAGAIL/blob/master/src/util/graph/KruskalsMST.java#L29

From what I can tell we are getting a cop of data (edges) then clearing it, but it appears that it is only clearing the returned list, which doesn't modify anything, nor does it save off the state in any way.

(I'm trying to convert to Python, but this is one of the very few lines that has me utterly baffled why it's there).

The ArffDataSetReader initializes all instances with null labels when reading an arff file.

The error ends up looking like this

Running Random Hill ClimbingException in thread "main" java.lang.NullPointerException
	at shared.SumOfSquaresError.value(Unknown Source)
	at opt.example.NeuralNetworkEvaluationFunction.value(Unknown Source)
	at opt.example.NeuralNetworkOptimizationProblem.value(Unknown Source)
	at opt.RandomizedHillClimbing.<init>(Unknown Source)
	at tests.OptdigitsTest.main(Unknown Source)

We need this pretty badly.

Is this library meant to be released under an open source style license? Without one it's unclear what uses of this code are acceptable - can this be used in a commercial context, for example? Strictly speaking, without a license, modifying and redistributing the code are copyright infringement.

https://github.com/pushkar/ABAGAIL/blob/master/jython/knapsack.py

Somehow the GA and MIMIC functions get scores that exceed the knapsack volume. This shouldn't happen if we have the correct checks in place, right?

If I run the example in the wiki:

cd ABAGAIL
ant
java -cp ABAGAIL.jar opt.test.XORTest

I get
Error: Could not find or load main class opt.test.XORTest

Instead of taking a Class as an argument for the constructor of the AdaBoostClassifier class, perhaps take in a factory that creates FunctionApproximaters. This way the user has the ability to customize the internals.

https://github.com/pushkar/ABAGAIL/blob/master/src/func/AdaBoostClassifier.java#L46

@pushkar
Some of these will make our lives a lot easier. Perhaps have Maven manage dependencies?

The search method in ABAGAIL.arrays does not return a valid index when the value found is at the top end of the distribution. The start 'high' value is the length of the array instead of length - 1, which results in the 'high' value sometimes returning an index that is out of bounds from the array.