Author: José Carlos
Date: 06:30:37 06/01/02
Go up one level in this thread
On June 01, 2002 at 09:19:43, Omid David wrote:
>On June 01, 2002 at 09:13:53, José Carlos wrote:
>
>>On May 31, 2002 at 18:50:18, Omid David wrote:
>>
>>>I forgot to change "negascout" to "AlphaBeta" in 7th line for better
>>>illustration. Here is the correction:
>>
>>1: int AspWin(int estimate, int delta, int depth)
>>2: {
>>3: int alpha = estimate - delta;
>>4: int beta = estimate + delta;
>>5: int best;
>>6:
>>7: best = AlphaBeta(alpha, beta, depth);
>>8:
>>9: /* re-search */
>>10: if(best <= alpha)
>>11: best = AlphaBeta(-10000, beta, depth);
>>12: else if(best >= beta)
>>13: best = AlphaBeta(alpha, 10000, depth);
>>14:
>>15: /* is there any better way to handle this very rare case? */
>>16: if(best >= beta || best <= alpha)
>>17: best = AlphaBeta(-10000, 10000, depth);
>>18:
>>19: return best;
>>20: }
>>
>> I numbered the lines for clarity. I think it should be:
>>
>>11: {alpha = -10000; best = AlphaBeta(-10000, beta, depth);}
>>13: {beta = 10000; best = AlphaBeta(alpha, 10000, depth);}
>>
>> If you don't update alpha and beta when researching, it can happen that you
>>fail low for [alpha,beta], then you research with [-inf,beta] and return best
>>with -inf < best < alpha which is a true score and there's no need to research.
>>Same for fail high.
>> Other than this, you're code looks correct.
>>
>> José C.
>
>But then in line 13 it would be in fact AlphaBeta(-10000, 10000, depth) since
>alpha = -10000 in line 11.
Maybe I'm still sleeping :)
In your code I see:
------------------------------------
Search asp window;
If fail low
search -inf,beta
else if fail high
search alpha,+inf
if still outside asp window
search -inf,+inf
------------------------------------
Now, if you fail low, research [-inf,beta], find a true score below alpha,
there's no need to research [-inf,+inf], which you do unless you change alpha
and beta according to the research.
Please, if forgive me if I'm wrong. I've had a bad night.
>My question is how can a full window search of
>(-10000,10000) be avoided at all costs?
You can try some ideas like research small windows around the score found. For
example:
alpha=-30;
beta=30;
search returns -50;
you try search [-80,30]
search returns +40 (very improbable)
you try search [-80,+80]
You get the idea. But I don't think it's any better than plain full with
research, and it's surely much more complex to code.
José C.
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.