[Computer-go] Some personal thoughts on 9x9 computer go
crescentmoon24 at gmail.com
Fri Mar 2 21:18:26 PST 2018
I was informed on the thread of the topic 9x9 go as Mangus had stated some good points that were similar to mine. Mine are based on my experience of correspondence go on OGS and GoQuest.
I have been playing online go 2011 with Deep Scholar which originally was Fuego with large GoQuest opening database the current version uses 2013 CrazyStone with this same database, as I am more of a Centaur Chess player and not a programmer I am more used to this type of playing where a large database can make a difference with along side a decent Computer Program.
How I operate mine is evaluate each position with a value of 0 if it a draw position or give a equal score, then with black advantages it gets a positive score and white positions get negative score to distinguish the scores and as the position got deeper and gave a value of +4 for black but previous positon had +2 then I would go back and change it till a refute tells otherwise. Originally I would do the same of searching in which then decide a move after a while and record the scores to the database which is all stored on a Multigo single sgf file.
Currently Deep Scholar runs on GoQuest with rating of about 2550 Elo.
It does well against programs like Aya about 90% win rate and some of the stronger 7-8 Dan players. Thankfully to Alderi the man behind the program MinusGo on OGS it has helped me realize how flawed the method that I was using for future growth of the database as mine. The database has some lines that can accurately play lines that are very deep about 40-50 moves and others may vary to about 10-20 moves. Now when a new position comes to play I search nearly even move in the next position as the MCTS tends to not see clearly enough now this doesn’t fix all just some of the short range issues, as Magnus had stated endgames, forcing moves, and ko fights tend to be problem that will take a while to deal with.
The current database version that I have built up over the years has dramatically changed thanks to the network programs such as MinusGo and MiniGo and open new scores of positions and help re-evaluate my position on 9x9 perfect play time and time again as the database grows. Currently I consider that 7 komi looks to be the accurate compensation for the first move as currently I have found a refute to previous positions that I went thoroughly in my games that I have lost then I test the line if it stands true or not.
I have done 9x9 computer go for quite a while now and 9x9 go is opening new horizons to how one should approach the game. It will be interesting to see what new development will become of the game for perfect play.
P.S- this is my first message on the mailing list so hopefully it comes out okay and thank you Magnus for bringing up the topic.
Sent from Mail for Windows 10
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Computer-go