[computer-go] Re: improving random players

[alain Baeckeroot]

Le jeudi 31 août 2006 00:40, Sylvain Gelly a écrit :
> Don Dailey wrote:
> >>>This was based on the observation that AnchorMan at level 8 (8 random
> >>>games) plays MUCH better than random even with such a small number of
> >>>games. 
> >>>
> >>>      
> >>>
> >>Sorry, I think I miss a point here. How AnchorMan at level 8 chooses the 
> >>8 moves to try? You draw randomly 8 moves, run a random simulation for 
> >>each, and take the move with the best score (so you count the territories)?
> >>
> >>Sylvain
> >>    

I hope i don't misunderstand your remark ;)

I tried some times ago with brown2 a very dumb montecarlo player:
it was trying 16 random moves, and for each did 100 simulations
(finish the games and count), as a result it was ranked near
 400 Elo on CGOS.

Then i think a lot ;) and try to use all the available information:
instead of trying 16 random moves i kept the 8 top moves from previous
simulation, considering the whole board has not change a lot (something
like quasi-static approximation in physics), and try 8 more random moves.

with this it improved a lot and jumped to 800 Elo, and beat all single
strategy bots (liberty count, influence, capture....)

Alain