[computer-go] A problem with selectivity and AMAF bias

[Magnus Persson]

Quoting Michael Williams <michaelwilliams75 at gmail.com>:

> Magnus Persson wrote:
>> Yesterday I noticed an odd phenomena in Valkyria wich was caused by  
>>  high selectivity and AMAF.

>> Then since there is nothing in the AMAF scores that indicate that   
>> move 2 is any good it is never searched, since the search is so   
>> selective that the bounds will not grow large to compensate for the  
>>  bias in a reasonable time.

> Isn't this fixed by never straying from a move in the tree until it
> loses and then trying an untried move?
> Or something like that.  It wasn't my idea and I don't remember the
> details, but it seems like it fixes what you describe.

No, it does not because the AMAF score for move 2 is strictly lower  
than the evaluation for move 1 and all other moves for some reason. It  
will try other moves deeper in the tree instead and the position is  
sufficently complex to generate a very large tree until it gives up on  
the move at the root level. The thing is that if it searches move 2  
for at least 100 simulations it will discover it is a good move. But  
because of the AMAF score is so low and all other moves are indeed  
losing moves it sticks to move 1 because it at least makes it into a  
fight although at bad odds.

Otherwise I am quite happy with the current implementation since it is  
strong in testing, this only happens when there are two hot candidates  
and the first one is searched first because of a limitation in move  
ordering, and a particularly strong bias works against the second best.

The annoying thing is that is can suddenly lose a game it was winning.

But I found a better fix. I also tried to enter my pattern priorities  
as the priors of the AMAF scores. And I now strongly believe that this  
is also better in general and not for this particular situation. In  
the position I wrote about yesterday this version get the right move  
almost immediately. I tested as Valkyria 3.2.0 overnight on CGOS and  
it seems to be just as strong as the previous version, and I can still  
tune the parameters of it.

This means that the search of position will be guided in order by

1) My pattern priorities disguised as priors of AMAF scores.
2) Then AMAF score will take over
3) If a move is searched, then the true winrates for those moves will  
be used and there is no bias from the pattern priorities except very  
weakly from the AMAF scores.

-Magnus



-- 
Magnus Persson
Berlin, Germany