Data steal
Moderators: hgm, Rebel, chrisw
-
- Posts: 7217
- Joined: Mon May 27, 2013 10:31 am
Data steal
When using MCTS and a neural network maybe it is best to store the training examples that are gathered during search. So if you later find out that the configuration of neural network is not that great and you chose to use another one then you can still use these training examples for tuning your new network.
-
- Posts: 2645
- Joined: Wed Mar 10, 2010 10:18 pm
- Location: Hamburg, Germany
- Full name: Srdja Matovic
Re: Data steal - games
I suggest to use games for supervised learning first,
to make sure the NN behaves like expected,
then start the Zero approach with selfplay.
There are some games on CCRL,
http://www.computerchess.org.uk/ccrl/404/games.html
With pgn extract
https://www.cs.kent.ac.uk/people/staff/djb/pgn-extract/
and the option -Wuci
you can extract pgn games to moves in can notation and final result,
so there is no need to implement an own pgn parser.
--
Srdja
to make sure the NN behaves like expected,
then start the Zero approach with selfplay.
There are some games on CCRL,
http://www.computerchess.org.uk/ccrl/404/games.html
With pgn extract
https://www.cs.kent.ac.uk/people/staff/djb/pgn-extract/
and the option -Wuci
you can extract pgn games to moves in can notation and final result,
so there is no need to implement an own pgn parser.
--
Srdja
-
- Posts: 931
- Joined: Tue Mar 09, 2010 3:46 pm
- Location: New York
- Full name: Álvaro Begué (RuyDos)
Re: Data steal - games
I am generating a database of games precisely for this purpose. I don't want to deal with differences in strength of the participants, so I am using Stockfish-vs-Stockfish at 1 second + 0.1 seconds/move. So far I have about 2.5 million games, and I intend to use 3 million.smatovic wrote:I suggest to use games for supervised learning first,
to make sure the NN behaves like expected,
then start the Zero approach with selfplay.
There are some games on CCRL,
http://www.computerchess.org.uk/ccrl/404/games.html
With pgn extract
https://www.cs.kent.ac.uk/people/staff/djb/pgn-extract/
and the option -Wuci
you can extract pgn games to moves in can notation and final result,
so there is no need to implement an own pgn parser.
--
Srdja
If anyone is interested, I'll upload them somewhere so we don't all have to spend a couple of months generating games.
-
- Posts: 2645
- Joined: Wed Mar 10, 2010 10:18 pm
- Location: Hamburg, Germany
- Full name: Srdja Matovic
Re: Data steal - games
I would appreciate it, thanks.If anyone is interested, I'll upload them somewhere so we don't all have to spend a couple of months generating games.
--
Srdja
-
- Posts: 931
- Joined: Tue Mar 09, 2010 3:46 pm
- Location: New York
- Full name: Álvaro Begué (RuyDos)
Re: Data steal - games
I was planning on publishing 3 million games in about two weeks. But I can publish 2 million right away if you are interested.smatovic wrote:I would appreciate it, thanks.If anyone is interested, I'll upload them somewhere so we don't all have to spend a couple of months generating games.
--
Srdja
-
- Posts: 2645
- Joined: Wed Mar 10, 2010 10:18 pm
- Location: Hamburg, Germany
- Full name: Srdja Matovic
Re: Data steal - games
I am in no hurry,I was planning on publishing 3 million games in about two weeks. But I can publish 2 million right away if you are interested.
two weeks are fine,
still assembling my gpu machine for training
and doing some research on NN.
--
Srdja