On Feb 2, 2009, at 12:09 PM, "Isaac Deutsch" <i...@gmx.ch> wrote:
Wow, thanks for all the answers! You're being really helpful.
"Do you use UCT with a too large exploration term?"
That's a good idea. I actually use a rather big value for c=0.5. I
might try
lowering it. Thanks! (Precisely, the formula is c*sqrt(log(p)/c).)
That formula is wrong. Did you mean to have a denominator of c?
"My first (braindead) multi-threaded UCT played weaker with two cores
than one core. How do you combine search trees/results? How do you
pick a move to play?"
I have the threads write to a global array (of course one after
another
that stores the values and the numbers of games played (each thread
adds to
these values) of the first ply. I then pick the move with the most
games
played through that node. The value of this best move (sum/numplayed)
determines if the bot should resign.
That sounds a lot like what I did (two independent searches that are
combined when genmove is called). I believe that strategy had a 100+
ELO loss in strength for me. I'm not sure why that was the case (but
have some theories). You may want to experiment with one search thread
to see if it hurts your performance too.
"Yes, and you should go by the bayeselo page which is a better
picture of
what is going on."
The ELOs posted here refer to these values.
"I don't think many people realize that you have to play hundreds of
games just to be within 40 or 50 ELO with much certainty. If you
play
less than 100 games you could easily be off by over 100 ELO."
Maybe I'm a bit (a lot :) impatient, but I try to make a rough guess
after
about 150-200 games, which is about 1-2 days of letting the program
run on
the server. I think it's possible to say after 150 games that RAVE
did not
give me a 300 ELO boost.
Ofline testing should be faster...
Thanks again,
Greetings,
Isaac
--
Jetzt 1 Monat kostenlos! GMX FreeDSL - Telefonanschluss + DSL
für nur 17,95 Euro/mtl.!* http://dsl.gmx.de/?ac=OM.AD.PD003K11308T45
69a
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
