I am more than willing to provide a binary version of my program if someone
else provides the operator and (refundable) entrance fee. While I doubt my
weak program would get best of the year, if it did win, I'd be fine with
giving the extra $$ to whoever paid the entrance fee.
On 9/13/07, Ray Tay
I know I'm only wading in the kiddie pool of computer go with my 1-ply bots,
but I think I may have found a useful enhancement to monte carlo.
HouseBot supports three 1-ply search modes:
1plyMC - Uniform sampling
1plyShuffle - Uniform sampling with monte carlo transposition reuse
1plyUCT - N
Does it defeat it based on number of samples taken or time allotted per turn?
On 9/14/07, Jason House <[EMAIL PROTECTED]> wrote:
> I know I'm only wading in the kiddie pool of computer go with my 1-ply bots,
> but I think I may have found a useful enhancement to monte carlo.
>
> HouseBot supports
Time management is identical. Based on quick profiling, the MCTR version
does about 1/6 of the simulations. Actually, since the MCTR does extra
tracking of info, it can reuse some old simulations, so it may be more like
1/4 or 1/5 of the simulations. It's just using the results more
efficiently.
Can you explain a bit more about how 1plyShuffle works?
On 9/14/07, Jason House <[EMAIL PROTECTED]> wrote:
> Time management is identical. Based on quick profiling, the MCTR version
> does about 1/6 of the simulations. Actually, since the MCTR does extra
> tracking of info, it can reuse some old
Sure.
I'm not sure how best to answer your question, so I'll respond with mostly
pseudo code. I can answer questions on theory, but I expect that to go back
and forth a bit since it's typically tough to explain.
Where there's grey areas in implementation choice, I have text describing it
(includi
On 9/14/07, Jason House <[EMAIL PROTECTED]> wrote:
> // Note that variance of the estimate increases by (0.5 * fractional
> game)^2
I should say that the variance of "wins" increases by that amount. The
estimate of winning percentage will be computed as wins / sims. The
variance for that
I'm curious - What is ReadyFreddy?
http://senseis.xmp.net/?ComputerGoServerdoes not list it. I do see
ReadyToGo. If it is, the description says 512
simulations. Over 81 moves, that seems like almost nothing... 6 sims per
move. Does it instead mean 512 sims per candidate move? That seems to mak
I'm not sure how many simulations ReadyToGo does but if it's on Senesis
that is probably correct.
ReadyFreddy always plays on CGOS and runs from the server. It DOES DO
512 simulations but they are all-as-first simulations. There is also
a strong fixed discouragement for playing self-atari in l
It's interesting that you say that it's just 512 all as first simulations...
especially when I look at the performance of ego110_allfirst. Anyone know
how ego110_allfirst was implemented?
I'd be really curious to see how 512 AMAF simulations would do without those
extra enhancements. If ego110_a
On Fri, 2007-09-14 at 15:15 -0400, Jason House wrote:
> I'd expect your enhancements are worth significantly more than 100
> ELO.
I can't really remember except that it was a non-trivial pleasing amount
that surprised me. Sorry I can't be more accurate.
Also, it may not scale the same at every
On Fri, 2007-09-14 at 15:15 -0400, Jason House wrote:
> It's interesting that you say that it's just 512 all as first
> simulations... especially when I look at the performance of
> ego110_allfirst. Anyone know how ego110_allfirst was implemented?
Does 110 mean 110 simulations?
Here are some m
12 matches
Mail list logo
