It really depends on the amount of time put into the monte carlo algorithm. It was 3 or 4 months of near full time work to get monte carlo stronger than the old Many Faces, and it was probably only about 3 kyu when it won the 2008 world championship (about 5 more months of part time work). It took another year or so to get it to one dan, and another year or so to get to 3 dan.
This is a graph of win rate (5000 playouts, 1000 test games) on 9x9 vs Gnugo by my monte carlo version number. In 2008 during this development I was working on it 4 days a week, and doing about 2 test versions a day. You can see that most of the things I tried didn’t work but every once in a while I had a breakthrough. Version 499 was the 2008 world championship entry. I’m on version 1242 now, so development slowed down quite a bit after 2009. > -----Original Message----- > From: [email protected] [mailto:computer-go- > [email protected]] On Behalf Of "Ingo Althöfer" > Sent: Saturday, September 27, 2014 8:09 AM > To: [email protected] > Subject: Re: [Computer-go] History of program strength > > Hi Peter, > > this is not exactly what you have been asking for, but the operating mask > of "Many Faces of Go"-bot gives an interesting judgement by senoir > programmer David Fotland. > > The levels of play say about strength on 19x19 board: > > * 1-dan (full Monte Carlo, without artificial time limit) > * 3-kyu (Monte Carlo with 3,000 playouts per move) > - - - - - - - - - - - - - - - > * 6-kyu (best level of traditional Many Faces - without Monte Carlo > techniques) > > So, this would mean a difference of 6 grades between MC and non-MC for > Many Faces. > > See also the screen shot in the appendix of this mail > > Cheers, Ingo.
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
