Hey Jim,
So I started playing around with
clojure-encog<https://github.com/jimpil/clojure-encog>,
and I'm pretty excited about it so far. Again, I'm trying to make a
financial series predictor. And I'm trying to go through the steps of 1)
nomalizing / preparing the data 2) creating a feed-forward neural network
with back-prop (I'll try sigmoid & gaussian activations). Then I'll 3)
train and 4) run the network.
*A)* The first problem I'm having is a library one. I'm trying to normalize
the data with the (*prepare* ...) function, but the
*normalization*namespace isn't in
*[clojure-encog "0.4.0-SNAPSHOT"]*. Here, we see that the *nnets* and *
training* namespaces are in the snapshot jar, but not the
*normalization*namespace. So I don't know how easy it is to update the
snapshot jar. But
in the meantime, I'll see if I can use the github version.
webkell@ubuntu:~/Projects/nn$ jar tvf
lib/clojure-encog-0.4.0-20120518.170223-1.jar
72 Fri May 18 17:58:04 PDT 2012 META-INF/MANIFEST.MF
1961 Fri May 18 17:58:04 PDT 2012
META-INF/maven/clojure-encog/clojure-encog/pom.xml
111 Fri May 18 17:58:04 PDT 2012
META-INF/maven/clojure-encog/clojure-encog/pom.properties
584 Fri May 18 17:00:30 PDT 2012 project.clj
* 9839 Fri May 18 17:01:38 PDT 2012 clojure_encog/nnets.clj*
11532 Fri May 18 17:57:20 PDT 2012 clojure_encog/examples.clj
* 10144 Fri May 18 17:43:58 PDT 2012 clojure_encog/training.clj*
2177 Mon May 14 21:57:20 PDT 2012 java/NeuralPilot.java
7574 Wed May 16 20:34:30 PDT 2012 java/PredictSunspotSVM.java
2338 Mon May 14 21:56:42 PDT 2012 java/LanderSimulator.java
1794 Fri May 18 16:02:22 PDT 2012 java/XORNEAT.java
1672 Fri May 18 16:04:14 PDT 2012 java/XORNEAT.class
1872 Mon May 14 14:53:26 PDT 2012 java/LanderSimulator.class
1943 Mon May 14 14:53:26 PDT 2012 java/NeuralPilot.class
7357 Wed May 16 20:37:20 PDT 2012 java/PredictSunspotSVM.class
*B)* The second problem I see is when trying to deal with the input data.
The example in
clojure-encog<https://github.com/jimpil/clojure-encog/blob/master/src/clojure_encog/examples.clj#L107>,
has just an array of doubles. But my input data is slightly different in
that I'm dealing with a LazySeq of arrays. Each of those arrays contain
tick data, Time, Ask, Bid, AskVolume and BidVolume:
(["01.05.2012 20:00:00.676" "1.32390" "1.32379" "3000000.00" "2250000.00"]
["01.05.2012 20:00:00.888" "1.32390" "1.33238" "3000000.10" "2200000.00"]
...)
So of course a call to ((*make-data* ...) , fails with the error
"*clojure.lang.LazySeq
cannot be cast to [Double..*". So I need to figure out 1) a way to get each
one of those input data points , into an input-layer neuron. I've started
to think about that when I was dabbling with
code<https://github.com/twashing/nn/blob/master/src/nn/neuralnet.clj>.
If you like, I can look into trying to jerry-rig these kinds of tick data
mappings into (
training/make-data<https://github.com/jimpil/clojure-encog/blob/master/src/clojure_encog/training.clj#L43>
).
But I need a better understanding of the concept of a Temporalwindow. The
other thing is 2) to figure out how to transform the time field into data
the nn can use. I've been spitting the Datetime object out to longs.
Thanks
Tim Washington
Interruptsoftware.ca
416.843.9060
On Sun, Jul 29, 2012 at 11:35 AM, Dimitrios Jim Piliouras <
jimpil1...@gmail.com> wrote:
> Hi Tim,
>
> According to :
>
> http://www.heatonresearch.com/content/encog-30-article-2-design-goals-overview
>
>
> encog 3 should have descent support for any temporal (time-series) based
> prediction support in particular for financial predictions...I'm afraid
> however that the only example that I've ported to clojure-encog which uses
> temporal data is the sunspot example (SVM not NN).
>
> Also, you shouldn't have any problems with the data (most likely you need
> to normalize them - I usually find (-1 1) or (0 1) to work best.
> for an example of how exactly you would do it look for
> "PREDICT-SUNSPOT-SVM" here:
>
> https://github.com/jimpil/clojure-encog/blob/master/src/clojure_encog/examples.clj
>
>
> these 2 lines do all the job with regards to your input data:
>
> normalizedSunspots (prepare :array-range nil nil :raw-seq spots :ceiling 0.9
> :floor 0.1)
>
> train-set ((make-data :temporal-window normalizedSunspots) window-size 1
> )
>
>
> As far as algorimthmic problems go encog has been around for quite a
> while...even though I don't necessarily agree with all the design decisions
> made along the way I find it is a rather mature lib...of course it is
> written in Java so being large means it is a bit of a mess! also there is a
> lot of duplication in random places...anyways, what I'm trying to say is:
>
> if you've got a specific example in mind, (like the financial prediction)
> maybe it's worth trying it out using clojure-encog or the encog-workbench
> (the gui) or any other already-made lib and see how it goes...writing your
> own will certainly teach you loads but it might take a while until you
> actually test what you want to test...
>
> Normalisation, randomisation or both are almost always needed...
>
> Hope that helps...
>
> Jim
>
>
>
> On Sun, Jul 29, 2012 at 5:41 PM, Timothy Washington <twash...@gmail.com>wrote:
>
>> Hey Ben,
>>
>> It's the same problem.
>>
>> user> (incanter/exp (incanter/minus 3254604.9658621363))
>> 0.0
>>
>>
>> But it's not the functions. It's the math. Euler's number 2.71828...
>> raised to the power of 3254604.9658621363, gives Infinity. So for my neural
>> net's activation func, either i) I shouldn't used a sigmoid, or ii) my
>> linear combiner needs to keep values within a certain bound. My neuron
>> inputs are below. And it's the bid and sk volumes and the long time value
>> that's giving me such a large number.
>>
>> - 1.3239 (bid price)
>> - 1.32379 (ask price)
>> - 3000000.0 (bid volume)
>> - 2250000.0 (ask volume)
>> - 1335902400676 ( #<DateTime 2012-05-01T20:00:00.676Z> long value)
>>
>>
>> I just had the idea to try a Gaussian or tanh activation function. I
>> think this is the point where I'll give
>> clojure-encog<https://github.com/jimpil/clojure-encog>a whirl. I have a
>> feeling I'll be running into a lot of these data and
>> other algorithmic problems. And it'd be good to work with something that
>> has already dealt with these issues. I still don't know if I need to
>> normalize my input data, how to untangle the activation result for
>> back propagation, etc. Any insights are welcome.
>>
>>
>> Tim Washington
>> Interruptsoftware.ca
>> 416.843.9060
>>
>>
--
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
