Motion Research Progress Report: 2010-04-11

I said I would use Daniel Shiffman's XOR code in Processing if I decided against using Weka in Processing. I skipped that and went directly to using Weka in NetBeans. I did this partly out of instinct and partly because I reasoned that if it turns out that a single-hidden-layer feedforward network isn't the best learning algorithm, I could always switch to another one using Weka, easily.

~~~~~~
A small victory! I managed to get a tiny piece of Java code to build fine. The code snippets from the Weka website oversimplifies: their classes throw a lot of exceptions, which I suppose is a good thing.

Ok, now I need to make sure that my classifier is working. The next thing i have to do is write some test data.
~~~~~~~
Good! I was able to get a Naive Bayes classifier to work. Whatever a Naive Bayes classifier is. (Wikipedia'd it. Thanks, World Wide Web.)

Even better! I got a MultilayerPerceptron (which has built in backpropagation) to work! I don't have as fine a control over it as if I had built it. It doesn't look like I can add a bias node, for example. But it gives me wonderful statistics to work with.

Ok, I guess the next step now is to train the neural network in advance, and then add data in real time and make it spit out classifications. I'm working with the Labor.arff file (really interesting, btw) which shows labor options and the decision of the union (I think it's the union; it's unlikely that it's the management) whether to accept the offer (neural network output: "good") or not (neural network output: "bad").

So try this out, maybe I'll see if I can import in Processing and create a gui.
~~~~~~~~~~
Um, porting it to Processing was ridiculously easy. I'm kind of shocked.
~~~~~~~~~~

Labels: metacreation, motion research