|Subject:||Re: [Bug-gnubg] bug when stopping ROs at a JSD difference.|
|Date:||Mon, 22 Dec 2008 08:25:36 +0000|
Jim Segrave wrote:|
> On Sat 20 Dec 2008 (00:02 -0500), bob koca wrote:
>> I've been setting rollouts to stop at a minimum of 324 trials when the best play is 3.2 JSD better than any of the rolloed out alternatives.
>> What has often been happening in this case is that the RO finishes and a number of games of eiterh 325 or 326 is reported where the JSD is much larger than 3.2 so it should have stopped at the 324
> Do you have a dual core processor? It could be that there are separate
> rollouts in different threads, so when game 324 starts rolling out in
> one thread, another begins rolling out (giving 325 games) or even
> rolls out game 325 and begins on 326 while 324 is not yet finished.
> This is speculation, not from checking the code, but I've got a strong
> feeling that it's the case.
You're right Jim, the different numbers are caused by the fact that the jsd and
end checking is separate from the rollout code. So in this example gnubg could
have done 323 trials and then several threads (lets say 3) all rollout one
trial, now the jsd is calculated and found to be too high. All we know is that
somewhere between 324-326 trials the move is no longer valid.
One fudge I could add is to say that if a move ends within the number of threads
being used and the jsd trial limit, set the number of trials to the jsd limit
(as it's likely the move actually hit the jsd limit before the min number of
trials). It just depends if this is worth adding.
Get Windows Live Messenger on your Mobile. Click Here!
|[Prev in Thread]||Current Thread||[Next in Thread]|