Re: New Ideas for bots (Was: BGBlitz 1.7 won against GNU-BG 0.14 at the 8 Computer Olympiad.....)

Achim Mueller wrote:

> Hmm, there is still some big difference in evaluation of some positions
> between the plies and sometimes even bigger surprise after a deep rollout.
> I still think there is a lot of improvement in checkerplay. When you have
> the time, check 0,2 and 4ply in volatile or (at least for us human)
> puzzling positions. I guess, too often we tend to believe th 2ply truth of
> the bots because it's more comfortable to say "Ah, yes, this move is
> appropiate" rather than to doubt and doing an exhausting analysis.
>  
>> As you said, the real challenge is in handling the doubling cube. In
>> that sense I still believe there is room for improvement.
> 
> ACK, and I'm still waiting for the perfect MET.

You always get the best ideas having a morning coffee and a cig ;-). What
about a "learning" neural net while playing. In Graz - after three
doubtfull passes by BGBlitz - I once thought it would be a good thing to
apdapt cube decisions based on former reactions of your opponent (like
worldclass bg players do). Frank changed cube strategy from "normal" to
aggressive" in the second match (though this only had an effect in money
play he later figured out).

But it should basically be possible to adapt doubles to the opponents
assumed point of first pass or point of last take based on a 2ply analyis
during a match/money session. You may also try this for checker play to
produce uncomfortable positions for your opponent, though I think this is
much more difficult. 

You see, there is still a lot of room for improvements.

Ciao

Achim