On 10/29/07, Christoph Birk <[EMAIL PROTECTED]> wrote: > > On Oct 29, 2007, at 8:39 AM, Jason House wrote: > > For all of us in the bot-making kiddie pool, it's exceptionally > > helpful to have reference implementations of basic algorithms > > running on the server. When playing with AMAF, I found the > > reference AMAF bots very helpful. Now that I'm playing with UCT, > > references for UCT would be helpful. > > 'myCtest-V-0003' is running 50k UCT. Pure random playouts guided > by a UCT search with theses parameters: > # playouts before expanding = 50 > node-score = win_ratio + 0.5 * sqrt(log(N)/n); > > I will start it under the nam 'myCtest-50k-UCT' later today running > 24/7.
I think I've gotten my big UCT bugs worked out. Thanks a lot for the reference. For any who are interested, hb-672-UCT has the following configuration: # playouts per move = variable (should be in the ballpark of 10k) # playouts before expanding = 10 node-score = win_ratio + tuned_standard_deviation * sqrt(0.8*log(N)/n); tuned_standard_deviation = sqrt(min(0.25 ,win_ratio*(1-win_ratio)+sqrt(2*ln(N)/n))) The 0.8 factor is carry over from initially following http://senseis.xmp.net/?UCT The tuning was based on http://hal.inria.fr/inria-00117266 and is supposedly superior to a flat 0.5 I'll likely try variants to better match Ctest: * No 0.8 factor * 50 playouts before expansion * No tuning * True 10k simulations per move
_______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
