1

I'm interested in AI and 2 days ago I found an interesting recent development in this area, it's called ES-HyperNEAT, first there was NEAT, then HyperNEAT then ES-HyperNEAT.

Here are some links to the topic :

http://eplex.cs.ucf.edu/hyperNEATpage/

http://eplex.cs.ucf.edu/ESHyperNEAT/

So I've downloaded the Java version of AHNI, but there is no tutorial, I guess the developers took for granted that it's easy to use, but to me, I don't know how to implement a solution to the following problem, doesn't seem very hard, but could someone show me how to get started ?

Input looks like this :

   Date   ,  A  ,  B  ,  C  ,  D
2013-07-26,18.94,19.06,18.50,18.63
2013-07-25,18.85,19.26,18.55,19.04
2013-07-24,19.32,19.40,18.47,18.99
2013-07-23,20.15,20.30,19.16,19.22   <-- Predict it ? [ Output ]
2013-07-22,20.09,20.23,19.80,20.03   <-- Start Date
2013-07-19,20.08,20.48,19.76,20.02
2013-07-18,19.88,20.68,19.64,20.12
2013-07-17,19.98,20.07,19.69,19.83
2013-07-16,20.38,20.49,19.51,19.92
  ......
2013-07-02,18.19,18.20,17.32,17.69
2013-07-01,18.38,18.96,17.95,18.15   <-- End Date

The program should read the above data from Start Date counting back n days to End Date, train on those data and the correct output will always be the next day's D value, I wonder how this can be implemented with ES-HyperNEAT ?

Specifically :

[1] Which classes to call to start the process ?
[2] How to tell it which fields in the input file to gather data, in this case it can ignore the Date field, and gather data from A,B,C,D [ not normalized to 0,1 ]
[3] How to tell it the correct result is the next day's D value ?
[4] How to specify the program should start from line x at the Start Date, and get data through line y at the End Date ?

Is there something like : myProgram.start(FilePath,Delimiliter,Filed2,Field3,..,Line_X,Line_Y,...) ?

Frank
  • 30,590
  • 58
  • 161
  • 244

1 Answers1

4

The readme.txt (which you can see at https://github.com/OliverColeman/ahni) contains some info about getting started with your own experiments, specifically see the DEVELOPMENT AND CREATING NEW EXPERIMENTS section. There is currently no code specific to performing time-series prediction in AHNI, so you would have to extend one of the base fitness function classes (see the readme). Your code would need to do the things you ask about (points 2-4), but you could create a fairly generic time-series prediction class which can be configured via the .properties file to specify the things in points 2-4. If you do do this then feel free to contribute it and we'll add it to the AHNI software on github :).

AHNI is intended as a research platform to support my own research (and hopefully others along the way), rather than an "easy to use, throw generic machine learning problem X at it" kind of software package (depending on your definition of "easy to use"). I try to keep the code clean, well-organised and the API well-documented so that others may use it, but creating a full-blown tutorial (and functionality) for the many possible use-cases is beyond the scope of the project (though of course I'd gladly include tutorials written by others).

Before going further I recommend considering the below:

When googling around for previous research on using HyperNEAT for time-series prediction I came across a question I asked several years ago that is similar to yours that I had completely forgotten about (I was surprised to see my name attached to the question! :)) http://tech.groups.yahoo.com/group/neat/message/5470 The reply to this question is good food for thought on the matter. Additionally:

(ES-)HyperNEAT is designed to exploit geometric regularities (patterns, correlations) in the input or output (see http://eplex.cs.ucf.edu/papers/gauci_nc10.pdf), so one question that might be worth exploring is whether the data contains regularities that can be represented geometrically (in my question I suggested plotting some window of the time-series on a 2D plane, which the 2D input layer of the network "sees", similar to the approach used in http://eplex.cs.ucf.edu/papers/verbancsics_gecco10.pdf. However, it sounds like NEAT, using a recurrent network, might be just as good if not better than HyperNEAT for this kind of problem.

Oliver Coleman
  • 734
  • 7
  • 13