Author: Daniel Shawul
Date: 03:21:44 05/13/04
Go up one level in this thread
On May 13, 2004 at 05:37:35, Vasik Rajlich wrote:
>On May 13, 2004 at 05:22:41, Daniel Shawul wrote:
>
>>On May 13, 2004 at 05:13:10, José Carlos wrote:
>>
>>>On May 13, 2004 at 04:56:23, Daniel Shawul wrote:
>>>
>>>>On May 13, 2004 at 04:20:02, José Carlos wrote:
>>>>
>>>>>On May 13, 2004 at 03:35:00, Daniel Shawul wrote:
>>>>>
>>>>>>Hello
>>>>>>
>>>>>>My search is fail soft(i return the actual score)
>>>>>>when i fail high at the root i widen the window by 3 pawns (300).
>>>>>>
>>>>>> if(score<=r_alpha)
>>>>>> {
>>>>>> r_beta=r_alpha;
>>>>>> r_alpha=score-300; //r_alpha = -MATESCORE;
>>>>>> }
>>>>>> else if(score>=r_beta)
>>>>>> {
>>>>>> r_alpha=r_beta;
>>>>>> r_beta=score + 300; //r_beta = MATESCORE;
>>>>>> }
>>>>>>If the search fail's high at 1.75 score but the real score was 10
>>>>>>i get a score of 4.75 (1.75 + 3) in the next iteration. Why? my search is fail
>>>>>>soft and the score returned should be independent of beta. If i change the 200
>>>>>>to 300,score returned is 4.75??
>>>>
>>>> a correction (1.75 + 2) = 3.75
>>>>
>>>>>
>>>>> Are you sure you're doing fail soft? I mean, you need some extra logic, like
>>>>>starting off with -INFINITE in alpha nodes and increase the best score even in
>>>>>fail lows. You also need to make sure of returning true bounds in qsearch. Also,
>>>>>if you do some kind of forward prunning depeding on alpha and beta you won't be
>>>>>able to return true scores.
>>>>
>>>> i do futility pruning and other purnings but i always return
>>>> (score + margin) in all of the cases. Not alpha.
>>>>
>>>>for example in the following test position
>>>> rnb2rk1/ppq2p1p/4p1p1/3pP1B1/3P1Q2/2b2N2/P1P2PPP/2RK1B1R w - - 8 14
>>>>
>>>> 1& 34 -0.31 0.05 1. Qxf7+?
>>>> 1 36 -0.10 0.05 1. Bd3
>>>> 2 138 -0.44 0.09 1. Bb5
>>>> 2 172 -0.24 0.09 1. Bf6
>>>> 3 447 0.10 0.14 1. Bf6 Qa5
>>>> 3 679 0.16 0.16 1. Bh6!
>>>> 3 781 0.18 0.16 1. Bh6 Rd8
>>>> 4 1899 0.15 0.20 1. Bh6 Rd8 2. Qf6 Nc6
>>>> 5 3778 0.47 0.27 1. Bh6 Re8 2. Qf6
>>>> 6 10330 0.71 0.33 1. Bh6 Rd8 2. Qf6
>>>> 7 26798 1.06 0.44 1. Bh6!
>>>> 7 86808 1.48 0.63 1. Bh6 Qe7 2. Bxf8 Qxf8 3. Qf6 Nd7
>>>> 8 208184 1.20 0.95 1. Bh6 Qe7 2. Bxf8 Qxf8 3. Bd3 Nd7
>>>> 4. Ng5 Qg7
>>>> 9 684592 1.40 2.27 1. Bh6 Nd7 2. Ng5 a6 3. f3
>>>>10 3253930 1.42 9.00 1. Bh6 Qe7 2. Qe3 Qb4 3. Bxf8 Kxf8
>>>> 4. Qh6+ Kg8 5. Bd3 f6
>>>>10 5270072 1.75 15.05 1. Bf6!
>>>>10 11120355 3.75 33.33 1. Bf6! //here score is 1.75+2=3.75
>>>> //full window opened
>>>>10 32259429 12.82 87.59 1. Bf6 Qxe5 2. Nxe5 Nd7 3. Nxd7 Bd2
>>>> 4. Kxd2 Bxd7 5. c4
>>>>11& 32784106 12.47 89.17 1. Bf6? Qxe5
>>>>
>>>> if i change margin to 3,the result will be something like this
>>>>10 11120355 4.75 33.33 1. Bf6!
>>>>10 32259429 12.82 87.59 1. Bf6 Qxe5 2. Nxe5 Nd7 3. Nxd7 Bd2
>>>> 4. Kxd2 Bxd7 5. c4
>>>>11& 32784106 12.47 89.17 1. Bf6? Qxe5
>>>>
>>>> I am very very sure i don't return alpha/beta anywhere in my search.
>>>>I also tried turning off hashtable ,nullmove,iid etc but no success.
>>>>
>>>>best
>>>>daniel
>>>
>>> I don't find the effect you present here any strange. I can be caused by many
>>>things:
>>> - Null move: you can cutoff with a much lower score than true score cause you
>>>allow the opponent two moves in a row.
>>> - Hashing: cutting off by a stored bound from a different depth
>>> - Move ordering: if you search moves a, b and c with scores +1, +2, +10 you'll
>>>cutoff with beta <= +1 in the first move, with beta <= +2 in the second.
>>> - Forward pruning: if score > beta + margin and (some conditions) return(beta
>>>+ margin) or return(beta)-not true score-.
>>> - Lazy cutoffs
>>> Etc...
>>
>> I do all of the things you said above but there is no return (alpha) or
>>return (beta) in my search. The search somehow converges to beta. What i store
>>in hashtables is score. Infact i tried turning off hashtable, null move , and
>>pruning techniques but it all comes back to beta.
>>
>
>Yes, this is normal, and it's the reason why MTD (f) is a reasonable alternative
>to PVS. The search is really good at doing the minimum amount of work.
>
>Just glancing at the logs from my last run (I use MTD (f)), I see that 100% of
>the fail-lows at the maximum depth returned the exact bounds value, and 73.3% of
>the fail-highs did so.
>
>These numbers are normal. It's also normal for soft fail-highs to be softer than
>soft fail-lows, even at big remaining depths. (It's obvious why it should be so
>at remaining depth == 0, but it also holds true throughout the search.)
I accept your logic. This thing has been troubling me for a long time. I
thought it was a bug. But it may be the similarity of PVS to MTD(f) that is
causing this.
Thank you all for your help
daniel
>
>Vas
>
>>
>>>
>>> If all of that is disabled, then make sure you have something like (simplified
>>>search, and also in qsearch):
>>>
>>>search(alpha,beta,depth)
>>>{
>>> int best = -INFINITY; // Make sure you start with -INF, not with alpha
>>> int score;
>>>
>>> while(all moves)
>>> {
>>> score = -search();
>>> if (score >= beta)
>>> return(score);
>>> if (score >= best)
>>> best = score; // update this even when score <= alpha
>>> }
>>>
>>> return(best)
>>>}
>>
>> That is exactly what i do in my search.
>>
>>>
>>> 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.