[Bug-gnubg] Re: race-database and possible bot weakness

[Robert-Jan Veldhuizen]

At 21:32 10/1/2002 -0300, Albert wrote:

[from a discussion with Neil:]

> NON-contact races where bots are generally very
> weak.

Yes. I really think Neil was thinking SNOWIE 3.2 here, and maybe Jellyfish. 
GNUBG isn't weak and definitely not very weak in these positions, in fact 
I'd say it's very strong in these positions, since GNUBG has a good RACE 
neural net which Snowie apparently lacks.

I think Neil hasn't tried GNUBG enough to tell and perhaps just assumed 
GNUBG would be similar to Snowie, but it's not. I've never seen GNUBG make 
the silly racing plays Snowie sometimes makes.

Of course GNUBG's RACE NN isn't perfect, so it might still be worth trying 
to improve something here, but I really hope most of the efforts are spent 
trying to improve other sorts of position evaluations. I feel very 
confident that there's a LOT more to gain in many other areas than racing 
with GNUBG, like bearing in/off versus contact as you showed with your 
examples (will look at those more closer now!).

BTW, note that 3-ply is sometimes worse than 2-ply (in general, not sure if 
it applies to these positions as well). So I think I'll do some rollouts to 
see what's going on here.


--
Robert-Jan (Zorba on FIBS)