Quoting David Ongaro <david.ong...@hamburg.de>:
Some say the discovery of MCTS had to come this late, because it
needs computer power to be efficient. But I think this misses a
great deal of the story. I'm sure todays MCTS programs can beat the
hell out of expert system Go programs even on year 2000's hardware.
Even though the time was ripe in 2006, I think with a more
systematic approach in Go we could have get there much earlier.
Therefore Rémis achievement cannot be overestimated.
I am running Valkyria3.5.9 on a Pentium4 2.8Ghz single core processor
which is about 10 slower than a modern i7 running 4 threads.
From the current Bayes Elo rating
72 Valkyria3.5.9_P4Bx 2576 8 8 17664 77% 2173
Gnugo as the reference of the state of the art
374 gnugo-3.8-l10F 1839 7 6 31622 42% 1880
It took me a long time to beat gnugo on 9x9 with the help of
MC-evaluation. The problem was I never discovered to do MC *Tree
Search* in the way it is done today. This I have to thank Crazystone
and Mogo for.
And with a modern standard i7 CPU
9 Valkyria3.5.17_4cx 2748 21 20 2696 85% 2261
one gets 200 Elo.
The thing was when I started doing MC-evaluation with the "wrong" tree
search method it was extremely inefficient. I really needed a lot of
faith in MC to continue working on. The idea was already there. But it
took a lot of work on the simulations to get to a point where it
showed how good the idea can be.
The latest version of Valkyria is almost rated 1000 Elo higher than
gnugo. 200 Elo is hardware development during these years. Reaching
gnugo strength is due to the search algorithm but the rest is to all
the tewas patterns and heuristics in the heavy playouts.
Best
Magnus
_______________________________________________
Computer-go mailing list
Computer-go@dvandva.org
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
