Just very short report (I am still working, changing everything and don't have established results so far). I have almost 15 mln of positions at master level and 15 mln of positions from FISC database now. Each evaluated with ~5 sec Stockfish evaluations (thanks to BOINC community!). I introduced 66 basic numerical features from position. I divided my collection into 2 classes {safe, unsafe}. unsafe - positions where side on move is most probably losing (< -100 cp) and safe - positions where side on move is not losing (>= -100 cp). Then I applied some known methods of classification. The best results I received with neural networks with 2 hidden layers 33,10 (area under ROC 0.8755) relatively good results are also with simple logistic regression (area under ROC 0.8528). I developed small chess engine and put there this logistic regression as position evaluation function and played some games. Engine is weak for now - but it seems to play reasonable with some visible plan. That is good enough to continue. When I implement more features and game will be more pleasant to see I will present it here. This is the end of this short report. Please wait for more detailed reports later. Thanks.

More...