as discussed with ferra i moved the entire oat code to a new project (joone_oat), made a jar of it and included it in the joone build path (oat_ferra.jar).
let me know if you encounter any problem building joone/joone_oat. for those how develop on the joone_oat project i suggest cross linking the the joone and joone_oat project (for eclipse users: don't forget to uncheck the error message for circular dependencies).
Just checking on some networks, will check the new CVS out later, but thanks a *lot* from my side. CVS is not my strong point.....
PS : Used the additional time the weekend gives to test some ideas on a moderate sided network. Tried three or four OAT optimisation strategies on it - systematic differential, differential evolution and some others that worked well before on smaller networks. Also, since it is FFN, tried the EKF FFN on it since I've never really tested that algorithm. I have to trim the network radically to push just 20 cycles through if I use EKF, but the diagnostics (MSE, ARV etc) are miles and miles ahead of a bigger network trained with e.g. 10,000,000 differential evolution iterations. While de is a lot faster, EKF is a lot better for these sized networks. Can probably improve OAT's performance with better initial guesses, but maybe the time is better spent on another few EKF cycles.
damn, the stuff looks so interesting. i wish i had more time :-( ...
Log in to post a comment.