[Bug-gnubg] Multi-threaded rollout

[Jim Segrave]

I was looking at Bob Koca's report of oddities when doing some
rollouts of initial positions. They probably were the result of using
a player setting with noise.

However, running what should be identical rollouts with threads = 2 on
a CoreDuo machine gave differing results, indicating there are still
problems with threading and rollouts. I'll fix the code so that
logging of rolled out games works with threads (it's currently broken,
as the file handle and file names are global) to try to identify why
the win/loss ratio differs.

I was using today's CVS (31 July, 11:00 GMT) with the following 2
command files:

===== cubeless ========
set player 0 human
set rollout player 0 cubedecision plies 0
set rollout player 0 cubedecision prune off
set rollout player 0 chequer plies 0
set rollout player 0 chequer prune off
set rollout truncation cube plies 0
set rollout truncation cube prune off
set rollout truncation chequer plies 0
set rollout truncation chequer prune off
set rollout players-are-same off
set rollout player 1 cubedecision plies 0
set rollout player 1 cubedecision prune off
set rollout player 1 cubedecision noise 0.0
set rollout player 1 chequer plies 0
set rollout player 1 chequer prune off
set rollout player 1 chequer noise 0.0
new match 5
set board 4HPwATDgc/ABMA
set matchid cAkgAAAAAAAA
set rollout initial on
set rollout trials 50
set rollout seed 1001
set matchequitytable "/usr/local/share/gnubg/met/g11.xml"
set rollout jsd stop off
set rollout jsd move off
set rollout cube-equal-chequer on
set threads 2
set rollout player 0 cubedecision cubeful off
set rollout player 0 chequer cubeful off
set rollout late player 0 cube cubeful off
set rollout late player 0 chequer cubeful off
set rollout player 1 cubedecision cubeful off
set rollout player 1 chequer cubeful off
set rollout late player 1 cube cubeful off
set rollout late player 1 chequer cubeful off
set rollout truncation cube cubeful off
set rollout truncation chequer cubeful off
rollout

====== cubeful =======

set player 0 human
set rollout player 0 cubedecision plies 0
set rollout player 0 cubedecision prune off
set rollout player 0 chequer plies 0
set rollout player 0 chequer prune off
set rollout truncation cube plies 0
set rollout truncation cube prune off
set rollout truncation chequer plies 0
set rollout truncation chequer prune off
set rollout players-are-same off
set rollout player 1 cubedecision plies 0
set rollout player 1 cubedecision prune off
set rollout player 1 cubedecision noise 0.0
set rollout player 1 chequer plies 0
set rollout player 1 chequer prune off
set rollout player 1 chequer noise 0.0
new match 5
set board 4HPwATDgc/ABMA
set matchid cAkgAAAAAAAA
set rollout initial on
set rollout trials 50
set rollout seed 1001
set matchequitytable "/usr/local/share/gnubg/met/g11.xml"
set rollout jsd stop off
set rollout jsd move off
set rollout cube-equal-chequer on
set threads 2
set rollout player 0 cubedecision cubeful on
set rollout player 0 chequer cubeful on
set rollout late player 0 cube cubeful on
set rollout late player 0 chequer cubeful on
set rollout player 1 cubedecision cubeful on
set rollout player 1 chequer cubeful on
set rollout late player 1 cube cubeful on
set rollout late player 1 chequer cubeful on
set rollout truncation cube cubeful on
set rollout truncation chequer cubeful on
rollout

===============

Changing the threads line to 'set threads 1' produces identical
results for cubeful and cubeless (and if you log the games, ,they are
identical) 

-- 
Jim Segrave           address@hidden