I’m currently working through my second assignment for CS7641 at Georgia Tech and I thought I’ll record some of my personal notes and coments on the ABAGAIL library. Since I’m not primarily a Java developer, working of this code obviously took longer than a seasoned Java developer. As such I wouldn’t be surprised if the methods and approaches I used here are:

Not efficient
Outright wrong!

Nevertheless, here are my attempts to extend the functionality in ABAGAIL.

The Setup

The first thing I tried (and failed) was to see if there was a good Jython IDE editor. There isn’t one that I found that was easy enough to use. I ended up just coding in notepad which was infact more pleasant than trying to dive through confusing IDE settings.

With the Java IDE I settled on IntelliJ Community edition, simply because I have used Android studio before and they are quite similar. Building the *.jar files for ABAGAIL were rather trivial and I had no hiccups there. However by now I had already (unfortunately) wasted three nights fiddling around with settings and playing around with the sample jython code which was provided. In terms of actually making progress on the assignment, I can say I made none.

Implementing my own `EvaluationFunction`

There isn’t much fun just using what was given in the examples. It is far more exciting and rewarding implementing something that you created yourself. So I began with the rather trivial example, maximising a parabola of the form:

$$ y = -(x-m)^2 + b $$

In order to accomplish this, we would quite naturally have to write some Java. Mirroring the other functionality in the opt.example, we can quite readily create such an EvaluationFunction.

public class ParabolaEvaluationFunction implements EvaluationFunction {

    private double xAxis;
    private double intercept;

    public ParabolaEvaluationFunction(double m, double b){
        xAxis = m;
        intercept = b;
    }

    public double value(Instance d) {
        Vector data = d.getData();
        return (-((data.get(0)-xAxis)*(data.get(0)-xAxis))) + intercept;

    }
}

Although it felt extremely difficult before, most of it was getting my head around what exactly Instance did. But other than that it was quite simple to implement. The next part was a little more tricky, since I (still) don’t completely understand what the Jython code was doing in the setup of the reinforcement learning file. But nevertheless following the same vein as the examples, we can quite quickly and easily create the correct script:

m = 6
b = 7
N=1
fill = [100] * N
ranges = array('i', fill)

ef = ParabolaEvaluationFunction(m, b)

odd = DiscreteUniformDistribution(ranges)
nf = DiscreteChangeOneNeighbor(ranges)
mf = DiscreteChangeOneMutation(ranges)
cf = SingleCrossOver()
df = DiscreteDependencyTree(.1, ranges)
hcp = GenericHillClimbingProblem(ef, odd, nf)
gap = GenericGeneticAlgorithmProblem(ef, odd, mf, cf)
pop = GenericProbabilisticOptimizationProblem(ef, odd, df)

rhc = RandomizedHillClimbing(hcp)
fit = FixedIterationTrainer(rhc, 200000)
fitness = fit.train()
print "RHC fitness : ", fitness
print "RHC optimal : ", rhc.getOptimal()
print "RHC: " + str(ef.value(rhc.getOptimal()))

sa = SimulatedAnnealing(1E11, .95, hcp)
fit = FixedIterationTrainer(sa, 200000)
fitness = fit.train()
print "SA fitness : ", fitness
print "SA optimal : ", sa.getOptimal()
print "SA: " + str(ef.value(sa.getOptimal()))

ga = StandardGeneticAlgorithm(200, 100, 10, gap)
fit = FixedIterationTrainer(ga, 1000)
fitness = fit.train()
print "GA fitness : ", fitness
print "GA optimal : ", ga.getOptimal()
print "GA: " + str(ef.value(ga.getOptimal()))

mimic = MIMIC(200, 20, pop)
fit = FixedIterationTrainer(mimic, 1000)
fitness = fit.train()
print "MIMIC fitness : ", fitness
print "MIMIC optimal : ", mimic.getOptimal()
print "MIMIC: " + str(ef.value(mimic.getOptimal()))

Which unsurprisingly comes to the right solution:

RHC fitness :  6.996995
RHC optimal :  6.000000
RHC: 7.0
SA fitness :  3.183305
SA optimal :  6.000000
SA: 7.0
GA fitness :  -8001.420405
GA optimal :  8.000000
GA: 3.0
MIMIC fitness :  6.983
MIMIC optimal :  6.000000
MIMIC: 7.0

Hopefully its an interesting take on how you can combine Python and Java in a machine learning setting.

The Setup

Implementing my own EvaluationFunction

Implementing my own `EvaluationFunction`