Re: [Bug-gnubg] Resign bug

[Holger]

At 14:58 18.08.2003 +0000, Joern Thyssen wrote:
> On Mon, Aug 18, 2003 at 10:06:00AM -0400, Ric wrote
> > I don't doubt that there are legitimate concerns from the programming side
> > on this issue. In fact, I'll even argue that the current approach is 
> fine IF
> > the user is notified in some obvious way about what is going on. Even a
> > popup box announcing "Gnu wins a single game" (or whatever) and thus ending
> > the game would be a useful approach.
>
> That's an alternative solution, but a bit dangerous, since terminating
> the game before end by an implicit resignation will be based on gnubg's
> evaluator. People may argue that the evaluator is not perfect, and it's
> perfectly imaginable that gnubg overlooks a winning
> 66-21-66-21-66-21-66-21 sequence if analysed at 0-ply (or perhaps even
> at 2-ply).

( Since no one has yet replied to your suggestions I will. )

A forced termination of the game is certainly no good idea as this will 
likely lead to more confusion. I would think that gnubg should resign at 
this occasion. If for any reason gnubg doesn't recognise this it should 
play on and hope for less than perfect play from the opponent.

> The only reason why I entered the discussion was to describe the current
> implmentation and because I found the extra "scoring" function
> interesting. Also, if implemented we have (another) feature that Snowie
> hasn't -- at least until they read the mailing list :-)

You do are a bit competitive.  :-)

At 13:16 18.08.2003 +0000, Joern Thyssen wrote:
> For cubeful chequerplay evaluations gnubg will sort the moves by cubeful
> equity and secondarily by cubeless equity if the cubeful equities are
> identical. In this case, both the cubeful and cubeless equity are
> identical thus gnubg cannot see the difference: it selects an
> arbitrary one, which in this case happens to be the annoying (and
> provoking on some players) move.
>
> A simple algorithm would be to use the one sided bearoff evaluator to
> find the moves which bears off in fewest rolls. This costs one lookup in
> the onesided bearoff database which would be close to zero.

I was to suggest to look for a move that bears off 2 chequers and take this 
one, but your approach is much more general.

> Another example of provoking behaviour by gnubg is random moves in races
> where the match is lost or money game with Jacoby rule where the
> opponent hasn't doubled. This will only occur if a human opponent or
> buggy bot has rejected a legitimate resign or if gnubg "forgets" to
> resign. Do we want gnubg to make the least provoking move in this case
> too? My simple algorithm above can be extended to cover such positions
> as well by using Joseph one-sided rollout code. However, now the this
> features comes with a non-zero prize tag -- do we want to spend a few
> seconds of CPU time to find the move  that won't provoke the opponent?

The calculation of a one-sided rollout seems to be quite instantaneous so 
it wouldn't be a big waste. I suppose 144 trials should be enough.

Regards,

        Holger