Hi Oliver,Oliver wrote:Hi Gerd,
let me first say what i would do. I would simply export my trainingsdata into a giant csv-file (or a custom MySQL-Table) and then use prefabricated tools to solve this problem.
If you like to fiddle with Neural Networks you might consider SNNS
(http://www-ra.informatik.uni-tuebingen.de/SNNS/).
If you are open to other statistical tools, you may choose R
(http://www.r-project.org/).
But if i evaluate you correctly, you are not the person to like that. You like to program every thing out for yourself. (Thats a good thing).
don't know, I need to go the "hard" way to understand things
No idea, if unsupervised is sufficient, fine. Otoh I would like to supervise or inspect the "learning" by computer aided tools.Oliver wrote:Regarding the sampleproblem you gave, it is not clearly defined!
One must distinguish supervised and unsupervised learning.
See e.g. here:
http://en.wikipedia.org/wiki/Unsupervised_learning
So the question is: Do you have already high quality scores or not?
Please answers this first, it is important. If not, you have really a problem with it. Otherwise, your problem is conceptual trivial.
Oliver
Don't know how you define high quality scores.
I tried and will try my best. All three mentioned scores, king-safety, king-activity and pawn-scores may be winning / losing scores and may somehow volatile. One tempo or side to move may be decisive.
Transitions are always critical - often subject of a tactical (q)search. For instance a huge "winning" score of an outside passed pawn in a pawn-ending, which must be captured by the opponent king in the very last moment. Leaving the center and allowing a path for the own king to pick all opponent rammed or backward pawns on the other wing after some quite moves. Ideally, to avoid oscillations, the "winning" score of an outside passed pawn should transpose to a winning king-activity score, even actually with the former "winning" pawn down.
Assuming the scores are high quality, the issue is to scale those values by gamestate.
Of course the more advanced a passer is, the higher the score.
Occupancy and control of the stop- or advanced-stop-square is an issue as well. Supporting pawns, indirect defence/attack aka rook from behind etc.
Scaling passers by gamestate usually considers, that it is harder or more committed, to defend or occupy a stop-square or an advanced stop-square. Which usually makes passers more valuable, the less pieces are available. The more one already considers the ability to block or control stop-squares independently of the gamephase (which is implicitly harder in late endings), the lower the amplitude of the heuristic sigmoid curve.
Gerd