Re: Joseki libraries (was: Re: computer-go: About Viking)

[magnus.persson@xxxxxxxxxx]

On 17 Jan 01, at 3:15, Mark Boon wrote:
> 
> Hmmmm, so you are a purist. I must say that in recent years I tend
> believe in a more purist approach as well, but in a completely
> different way. In my opinion, the correct way is to have the program
> to generate the moves from first principles and then use patterns to
> generate the same moves for speedup (generate them faster, use extra
> knowledge for better sorting etc.). To use patterns for moves
> otherwise not generated doesn't seem purist at all.

I agree. For some of my patterndatabases this exactly what I try to 
do (but I skip the part where the program generates the move first - 
I do all the thinking). But it depends on how you use the patterns. 
The important point to note about my full board move generation 
patterns are that the meaning of them only is "this is a possible 
move in some positions". What I am really is doing is trying to 
exclude those moves that cannot possible be a good move such as 
playing into safe territory of either player. The only thing these 
patterns do is to filter away a large portion of legal moves, that 
would be a waste of time to evaluate.

But I am compromising this idea slightly in my current 
implementation. Because my program grossly overestimate the 
value of some tactical blunders that are sente but causes a loss if 
the opponent answers, I do not always have a pattern for moves 
that might be a tesuji. Instead tactical search will generate such 
moves if they save or capture a block that has an unstable status.

If my program in the future will be able to make selective local 
lookahead, then I will probably include *any* move that might be a 
good move in some position, but that is not possible yet.

Best wishes 
Magnus