[computer-go] Caching of local search in NeuroGo?

[Rémi Coulom]

Mark Boon wrote:
> 
> On 9-aug-2006, at 19:42, Markus Enzenberger wrote:
> 
>> I only try the move for the attacker with more empty adjacent points. 
>> Only if 
>>
>> there is a tie, I try both moves.
>>
> 
> Hmmm, I'm sure that's a heuristic that will work most of the time. But 
> 'most of the time' is not nearly good enough for my purposes.
> 
> Obviously if you limit yourself to one candidate for the attacker the 
> branch-factor will remain one. But that's not the same as solving a 
> general 2-liberty problem. Solving them properly also doesn't take that 
> much time, we're talking some microseconds per case, so I don't 
> understand why one would want to skimp on a few candidates and run the 
> risk of getting it wrong.

If you consider that a 2-liberty problem is a problem that can be solved 
without the defender ever having more than two liberties during the 
search, then, with the method of Markus, there is no risk of getting it 
wrong. If one move has less adjacent point than the other, then either 
the other is an immediate killer if it has two liberties, or it is 
mandatory to play there to avoid stopping the search if it has three or 
more.

Also, I would like to add that in case of tie, it is easy to find 
efficient move-ordering heuristics. For instance, by testing whether the 
defender would connect to strings with liberties.

> 
> I also put a limit on the depth, but it's large enough to get some 
> really crazy ladders right. But by limiting the depth I don't need to 
> spend time watching out for triple-ko.
> 
> Mark
> 
> 
> ------------------------------------------------------------------------
> 
> _______________________________________________
> computer-go mailing list
> computer-go at computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/