Re: [Bug-gnubg] HELP program hangs and does not complete roll outs!

[Jim Segrave]

On Thu 13 Feb 2003 (06:11 -0600), Neotropical Bats wrote:
> Hello again,
> 
> I can now set up the position so I can do a roll out but the program now 
> has been running for 9 hours and only 18% complete???
> 
> I did complete several roll outs before and all worked fine.  But this 
> position just will not work.  I have rebooted my machine etc.  I did click 
> on the roll out but it then stopped.
> 
> Thanks that worked fine now that I know how to do it.
> How can I speed up roll outs?
> 
> The first few completed in an hour or so, but this one has been going for 4 
> hours and still only 3% complete.
> 
> Are there some settings I can tweak?  Maybe I set some option that are 
> causing this very long roll out time?
> 
> I have a number of computers in the office, but the one I use for most 
> "number crunching" is a Dell with an Intel 2.4 GHz pentium with 1024 MB 
> RAM.  Hard drive space on this machine for data (not the OS drives) > 
> 270GB.  So I could set aside a fair amount of space for results.

The only thing I know of which makes rollouts faster but doesn't have
any effect on the accuracy or not is to ensure that you have set the
cache size large enough - Settings->Options->Other->Cache size. I use
a setting of 200000 on my laptop (128M), which uses about 30M of
Ram. You might try a value of maybe 1000000 or so.

After that, your only choices are to change the analysis. If you are
doing 2 ply, 1 ply will be faster, but less accurate. If you use
truncation, you don't ask gnubg to play the game out to the end,
instead you let it play some number of moves and then evaluate the
resulting position assuming that the evaluation will be sufficiently
accurate. You can tell it to play the first few moves at two ply, then
the following moves at 1 ply. You can adjust the move filters to
examine fewer moves.

The particular position you are rolling out is only 2 oves into a
game. I tried it on a 2.4Ghz PC, 1296 games with 2 ply evaluation and
very generours filters (accept 0 moves at 0 ply, add up to 10 more
moves below 0.400, no 1 ply filtering). This took about 4 to 6 hours
to roll out. The results were not useful - the side on roll had an
tiny equity advantage, but the standard deviation was several (5 or 10
times) the equity, so you'd have to roll out some huge number of games
to get any meaningful result.


-- 
Jim Segrave           address@hidden