Chess@home/en

Aus BC-Wiki
Zur Navigation springen Zur Suche springen
The purpose of this project is to construct a chess winnig possition classifier (or score estimator) based on some simple features (e.g. the number of pieces of each kind, number of squares controled by given side, numbers of checks that you can give in a possition and so on) The first goal of the project is to establish how far we can get with relatively cheap methods of position evaluation and simple data mining methods (eg. logistic regression, k-nearest neighbors, SVM, decision trees). The second goal is to use more advance machine learning techniques for recognition of winning position to emulate human behaviour of various levels of chess proficiency. This means we need good heuristic which will allow to narrow the range of positions to calculate in depth. There is a hope that modifying this heuristic from “more learnt” to “less learnt” we will obtain something similar to natural human levels.

To achieve this goals, in a first phase, I am going to build a vast database of chess positions (I collected already 25 mln and growing) and evaluate them with chess engine - my choice is open source Stockfish. For each position I will keep the score (in centipawns or moves to mate if discovered) and the best move. I assumed that for the purposes of this project it will be enough to calculate 5 000 000 nodes per position which requires an average 5 - 6 secs per position for one core (~2.5 GHz). To evaluate 25 mln possition on one core I would need 6 x 25 mln = 150 mln sec, which means 41 666 hours, 1736 full days. And 25 mln is only the number of possitions in games of known chess masters (that I have), which means games on a good level. I also plan to add more games from less advanced players from FISC game database to have also relatively weaker positions. So the overall number of positions can reach easily hundreds of millions. It is impossible to build desired database within single computer, so the only chance for me to complete my project is to use generosity of BOINC community. Later, after first phase of data collecting, BOINC will be still used for some part of machine learning phase.

Chess@home
Start 2014
End 2014
Status finished
Admin Michal Stanislaw Wojcik
Institution Poland BOINC Foundation
Country Poland
Area Games
Apps
Win chessathome 0.03
Linux
Mac
64bit
PS3
ATI
CUDA
Intel
Android
RPi
NCI
System-Specs
VRAM SP DP
RAM 138MB
Runtime 5-9min
HDD 1,4MB
Traffic dl/ul kb / kb
Deadline 2 days
Checkpoints