Hey Jim, Encog does look very interesting. Right now, I'm trying (and failing) to implement the sigmoid function. I'm using wikipedia's reference<http://en.wikipedia.org/wiki/Sigmoid_function>, and trying to use Incanter's (incanter/exp) function, but Incanter's function doesn't seem to work:
user> user> (incanter/exp -3254604.9658621363) 0.0 user> user> (incanter/exp 3254604.9658621363) Infinity I took a peek at encog's java sigmoid implementation<https://github.com/encog/encog-java-core/blob/master/src/main/java/org/encog/engine/network/activation/ActivationSigmoid.java>, and while my calculus is correct, they are using their own exponential function. *i)* Is there another Incanter function I should be using? *ii)*Maybe the number going into (incanter/exp) has to be within a certain range? I'm getting that *3254604.9658621363* number as a linear combination of these inputs (i took the long value of time and randomized the weights). *iii)* Is there a different way to sum these value? Is this where input data massaging comes in? ({:key :time, :value #<DateTime 2012-05-01T20:00:00.676Z>, :weight 0.22072475374809264} {:key :bid, :value 1.3239, :weight 0.5831561982324751} {:key :ask, :value 1.32379, :weight 0.9364679759677283} {:key :bvolume, :value 3000000.0, :weight 0.5956072849191396} {:key :avolume, :value 2250000.0, :weight 0.07837823837657176}), There's a few other issues I'm still trying to understand. I outlined them in this quant.stackexchange<http://quant.stackexchange.com/questions/3814/multilayer-perceptron-neural-network-for-time-series-prediction>post. But primarily: - Algorithmically, I don't understand how to achieve both bid and ask predictions with each tick? the i) linear combiner and ii) activation functions, coupled with the bias, act to produce 1 value - In Back Propagation, is it just the weights that are adjusted, or the biases as well? There are other things like how to decide the number of neurons in the hidden layer. But once I get a very simplistic code example working, I'll feel more confidence in understanding the mechanics of this neural network works. And I'd start using your clojure-encog library as it's probably much more complete than anything I could build. Any insights or examples you could provide would be great. Cheers Tim Washington Interruptsoftware.ca 416.843.9060 On Sat, Jul 28, 2012 at 8:51 AM, Dimitrios Jim Piliouras < jimpil1...@gmail.com> wrote: > If you're doing anything related with neural nets I think clojure-encog is > indeed the place to start... I've wrapped most of the original java encog > v3 so far and I am actively using it for my own projects...I'd be very > happy to hear your feedback if you decide to use it after all...The truth > is I've not provided a wiki but I have provided an examples.clj where I've > ported some of the most famous encog examples using my wrapper...hope you > find it useful! any help you need along the way don't hesitate to ask! > predicting financial series has been done in the past by heaton-research > but unfortunately I've not ported that example! maybe it is time to do so > as soon as I return from holidays... > > take care... :) > > Jim > > On Fri, Jul 27, 2012 at 10:03 PM, Timothy Washington > <twash...@gmail.com>wrote: > >> I'm working on a Neural Network application to tech myself machine >> learning and AI. >> >> - https://github.com/twashing/nn >> >> There's nothing there right now. But I'm keenly interested in the field, >> and getting a working project, predicting financial time series. I'm keen >> to check out clojure-encog <https://github.com/jimpil/clojure-encog>, >> and any other AI library. I just want to make sure I understand the >> mechanics of the algorithms that are being used. >> >> -- You received this message because you are subscribed to the Google Groups "Clojure" group. To post to this group, send email to clojure@googlegroups.com Note that posts from new members are moderated - please be patient with your first post. To unsubscribe from this group, send email to clojure+unsubscr...@googlegroups.com For more options, visit this group at http://groups.google.com/group/clojure?hl=en
