Re: [Bug-gnubg] Predefined move filters

[Gary Wong]

On Tue, Dec 31, 2002 at 11:03:23AM +1300, Joseph Heled wrote:
> Gary Wong wrote:
> >I agree.  In fact, I think 1-ply comes out worst only when the judge
> >is an even-ply evaluation; head-to-head, 1-ply beats 0-ply by a tiny
> >margin.
> 
> I disagree. 0ply beats or equals 1ply when the judge is 0ply rollouts, 
> as my benchmark shows for both race and crashed positions.

Oh dear.  "set player 1 chequerplay evaluation plies 1" and "set
player 0 chequerplay evaluation plies 0" eventually has player 1 ahead
by a statistically significant margin, tens of thousands of games
later.  I would have expected this result to match 0-ply rollouts, so
I am puzzled about what to conclude.  One possibility is that the
behaviour of the most recent nets differs from when I last ran 0-ply
vs. 1-ply; another explanation might be that 0-ply rollouts are
sufficiently biased in favour of 0-ply evaluations to distort the
result.

Considering everything I've heard, then, I still don't have enough
information to derive a confident conclusion either way.

How about we try:

    set rollout cubeful off
    set rollout player 0 chequerplay plies 0
    set rollout player 1 chequerplay plies 1
    set rollout varredn on
    set rollout trials $as_many_as_we_can_stand
    set rollout initial on

and see what happens?  (Until we implement "real" initial position
rollouts, we'll have to run equal numbers with the players each
way and add the results...)

Cheers,
Gary.
-- 
   Gary Wong           address@hidden           http://www.cs.arizona.edu/~gary/