notes.txt

2003_12_12:
I implemented the auto-correlation tool (AutoWKineticEulerEval) which modified W to meet its needs.  In general, doing an evaluation of this nature takes about twice as long as a normal evaluation, but it is more likely to be guaranteed to get a result.  Definite reason to implement less setback (i.e., < 1000).  We shall see tomorrow.  

Additional ideas to implement/do:
- implement more nodes during parallel annealing to get right (doing, but using 100 iterations, only)
- a quicker method for doing the training would be to do an inital scan at a large resolution, and then starting the gradient descent from the lowest point, starting at the next highest resolution.  Still may be problematic,  . . . but . 
- use a proper gradient-descent algorithm for speed (currentlly takes 20-50 iteartions to get point right)
- implement learning algorithm (Alopex?) :  Alopex-B: A New, Simpler, but yet Faster Version of the Alopex Training Algorithm (A. Bia) through  ACM


Things I need to do:
- make sure that the binning works (reproduce the same data-set, test2)
- can I train to the binned-training? (kineticanneal)
- run the real data
- ADD ALL OF THIS TO CODE FOR DOXYGEN (for todo-part)
- UPDATE THIS IN BLOG 
- add xml reader
- do gui
- add on-line writer for networks and other progress