Re: [gnugo-devel] status of twin code?

[Alain Baeckeroot]

Le lundi 31 décembre 2007, Joseph Piche a écrit :
> When browsing through some of the enhancements in trac, I ran across the
> Twin code. I was wondering what the status of it is. Knowing where your
> opponent wants to move is a vital aspect in becoming a better player, so I
> imagine we would eventually want something that looks for these types of
> moves. If it is currently abandoned, is it because of lack of time/interest
> (if this is the case, I might be able to help a little) or is there some
> other factor, i.e. integrating it with other parts of the code?
> 
> Joseph Piche
> 

There are various reasons:
- i had time when i tried this and now i have full time job, so much less
time :)
- it was my first C hack, so it is a bit ugly.
- I had a bot running on kgs to try to measure the average improvement. It was
 less than half a kyu, for 50% slow down, and due to this slowness, the
 twinbot sometimes was weaker in tournaments due to time constraints: it falls
 to level 0 or so then play very bad moves.

I fully agreed with Gunnar summary : it is not worth the additionnal complexity
and slowness.

Anyway, it was interesting:
- the "twin" is in fact the "miai value" (in go terms) or the "temperature" in
game theory, with the limitation of the evalutation function.
- it would need refinements to prevent weird side effects: using only numerical
value is too crude, an attempt to use move_reasons when there is an urgent
semeai move showed very little improvement.
- The main improvement could come from better ko handling: it is possible 
to count all threats for both players, and then (assuming the count is correct)
give correct value to moves according to this (instead of multiplying by 0.5).
Maybe this can be done without the twin, but it seems possible to do with
the twin.

Alain