optimal aspiration window for stockfish question

[bob]

testing still underway, but one surprise already. Turns out that narrow aspiration windows might be worthless. I ran a test with several starting windows to see if there was a clear winner. There were clear losers like +/-1 and +/- 2 and such. But +/-16, +/-50, +/-100, +/-200 all seem to be equivalent in real-game testing. More when the test is completely done...

BTW this is the "delta" value. Which means with a value of 200 the first search is old+/- 200, the first research is old+/-400...

hmmm...

[michiguel]

testing still underway, but one surprise already. Turns out that narrow aspiration windows might be worthless. I ran a test with several starting windows to see if there was a clear winner. There were clear losers like +/-1 and +/- 2 and such. But +/-16, +/-50, +/-100, +/-200 all seem to be equivalent in real-game testing. More when the test is completely done...

BTW this is the "delta" value. Which means with a value of 200 the first search is old+/- 200, the first research is old+/-400...

hmmm...

I believe this will depend quite a bit on the specific engine. I observed a clearly distinct peak at about 40 cp (IIRC) with a slow decline toward higher deltas and a sharper decline towards lower ones. However, this peak was not very high. IIRC, when I tested this, it may have been ~3 elo points or so (compared to no aspiration). But the curve was clear. Of course I needed to run 80k games each time to have enough precision.

Anyway, there are two effects here. One of them has not been mentioned. The obvious one is the size of the tree and researches, but the ignored one is that fail lows may trigger longer thinking times to catch problems. Too narrow windows will cause the engine to think longer in positions that are not needed, and too wide windows, will become insensitive to potential problems (they won't fail low so often).

Miguel

[bob]

testing still underway, but one surprise already. Turns out that narrow aspiration windows might be worthless. I ran a test with several starting windows to see if there was a clear winner. There were clear losers like +/-1 and +/- 2 and such. But +/-16, +/-50, +/-100, +/-200 all seem to be equivalent in real-game testing. More when the test is completely done...

BTW this is the "delta" value. Which means with a value of 200 the first search is old+/- 200, the first research is old+/-400...

hmmm...

I believe this will depend quite a bit on the specific engine. I observed a clearly distinct peak at about 40 cp (IIRC) with a slow decline toward higher deltas and a sharper decline towards lower ones. However, this peak was not very high. IIRC, when I tested this, it may have been ~3 elo points or so (compared to no aspiration). But the curve was clear. Of course I needed to run 80k games each time to have enough precision.

Anyway, there are two effects here. One of them has not been mentioned. The obvious one is the size of the tree and researches, but the ignored one is that fail lows may trigger longer thinking times to catch problems. Too narrow windows will cause the engine to think longer in positions that are not needed, and too wide windows, will become insensitive to potential problems (they won't fail low so often).

Miguel

You might test the fail-low hypothesis. Why? In Crafty, if the score drops 0.01 from the last iteration, I will use more time, just as if it had dropped 1.0... and a lot of games proved this to to be best. And not just by 1 or 2 elo either. I was surprised but have been running with this for a good while now (almost a year). I don't just search longer on a fail low, I search longer on any score drop at all, fail-low or not... It tests significantly better.

[bob]

Here is the output from BayesElo first:

2 Crafty-23.5R06-200 2650 4 4 30000 64% 2539 22%
3 Crafty-23.5R06-24 2650 4 4 30000 64% 2539 22%
4 Crafty-23.5R06-100 2649 4 4 30000 63% 2539 22%
5 Crafty-23.5R06-30 2648 4 4 30000 63% 2539 22%
6 Crafty-23.5R06-50 2648 4 4 30000 63% 2539 22%
7 Crafty-23.5R06-300 2648 4 4 30000 63% 2539 22%
8 Crafty-23.5R06-20 2648 4 4 30000 63% 2539 22%
9 Crafty-23.5R06-10 2645 4 4 30000 63% 2539 22%
10 Crafty-23.5-2 2645 4 4 30000 63% 2539 22%
11 Crafty-23.5R06-1 2645 4 4 30000 63% 2539 22%
12 Crafty-23.5R06-8 2644 4 4 30000 63% 2539 22%
13 Crafty-23.5R06-5 2643 4 4 30000 63% 2539 22%
14 Crafty-23.5-1 2641 4 4 30000 63% 2539 22%
15 Crafty-23.5R06-2 2636 4 4 30000 62% 2539 22%


version 23.5-1 and 23.5-2 are simply two consecutive runs with the same version to provide a normal result. The rest of the tests are version 23.5R06 and were tested where the -n is the aspiration window (delta value in the code posted yesterday). 23.5R06-1 means the aspiration window was +/- 1 with delta=1. 1 and 2 are a bit low, and by the time it gets to 10, it is pretty optimal. Bigger doesn't seem to hurt at all up to +/- 3.0 pawns... I was expecting a better result in the 20-40 range, the reason I ran the big numbers was to produce some worse results so that there is a recognizable curve with a clear optimal value, and worse results on either side. Didn't get exactly what I expected, as you can see...

[Rémi Coulom]

Here is the output from BayesElo first:

2 Crafty-23.5R06-200 2650 4 4 30000 64% 2539 22%
3 Crafty-23.5R06-24 2650 4 4 30000 64% 2539 22%
4 Crafty-23.5R06-100 2649 4 4 30000 63% 2539 22%
5 Crafty-23.5R06-30 2648 4 4 30000 63% 2539 22%
6 Crafty-23.5R06-50 2648 4 4 30000 63% 2539 22%
7 Crafty-23.5R06-300 2648 4 4 30000 63% 2539 22%
8 Crafty-23.5R06-20 2648 4 4 30000 63% 2539 22%
9 Crafty-23.5R06-10 2645 4 4 30000 63% 2539 22%
10 Crafty-23.5-2 2645 4 4 30000 63% 2539 22%
11 Crafty-23.5R06-1 2645 4 4 30000 63% 2539 22%
12 Crafty-23.5R06-8 2644 4 4 30000 63% 2539 22%
13 Crafty-23.5R06-5 2643 4 4 30000 63% 2539 22%
14 Crafty-23.5-1 2641 4 4 30000 63% 2539 22%
15 Crafty-23.5R06-2 2636 4 4 30000 62% 2539 22%


version 23.5-1 and 23.5-2 are simply two consecutive runs with the same version to provide a normal result. The rest of the tests are version 23.5R06 and were tested where the -n is the aspiration window (delta value in the code posted yesterday). 23.5R06-1 means the aspiration window was +/- 1 with delta=1. 1 and 2 are a bit low, and by the time it gets to 10, it is pretty optimal. Bigger doesn't seem to hurt at all up to +/- 3.0 pawns... I was expecting a better result in the 20-40 range, the reason I ran the big numbers was to produce some worse results so that there is a recognizable curve with a clear optimal value, and worse results on either side. Didn't get exactly what I expected, as you can see...

You should use CLOP

Rémi

[Karlo Bala]

Here is the output from BayesElo first:

2 Crafty-23.5R06-200 2650 4 4 30000 64% 2539 22%
3 Crafty-23.5R06-24 2650 4 4 30000 64% 2539 22%
4 Crafty-23.5R06-100 2649 4 4 30000 63% 2539 22%
5 Crafty-23.5R06-30 2648 4 4 30000 63% 2539 22%
6 Crafty-23.5R06-50 2648 4 4 30000 63% 2539 22%
7 Crafty-23.5R06-300 2648 4 4 30000 63% 2539 22%
8 Crafty-23.5R06-20 2648 4 4 30000 63% 2539 22%
9 Crafty-23.5R06-10 2645 4 4 30000 63% 2539 22%
10 Crafty-23.5-2 2645 4 4 30000 63% 2539 22%
11 Crafty-23.5R06-1 2645 4 4 30000 63% 2539 22%
12 Crafty-23.5R06-8 2644 4 4 30000 63% 2539 22%
13 Crafty-23.5R06-5 2643 4 4 30000 63% 2539 22%
14 Crafty-23.5-1 2641 4 4 30000 63% 2539 22%
15 Crafty-23.5R06-2 2636 4 4 30000 62% 2539 22%


version 23.5-1 and 23.5-2 are simply two consecutive runs with the same version to provide a normal result. The rest of the tests are version 23.5R06 and were tested where the -n is the aspiration window (delta value in the code posted yesterday). 23.5R06-1 means the aspiration window was +/- 1 with delta=1. 1 and 2 are a bit low, and by the time it gets to 10, it is pretty optimal. Bigger doesn't seem to hurt at all up to +/- 3.0 pawns... I was expecting a better result in the 20-40 range, the reason I ran the big numbers was to produce some worse results so that there is a recognizable curve with a clear optimal value, and worse results on either side. Didn't get exactly what I expected, as you can see...

What is the average depth?

[bob]

Here is the output from BayesElo first:

2 Crafty-23.5R06-200 2650 4 4 30000 64% 2539 22%
3 Crafty-23.5R06-24 2650 4 4 30000 64% 2539 22%
4 Crafty-23.5R06-100 2649 4 4 30000 63% 2539 22%
5 Crafty-23.5R06-30 2648 4 4 30000 63% 2539 22%
6 Crafty-23.5R06-50 2648 4 4 30000 63% 2539 22%
7 Crafty-23.5R06-300 2648 4 4 30000 63% 2539 22%
8 Crafty-23.5R06-20 2648 4 4 30000 63% 2539 22%
9 Crafty-23.5R06-10 2645 4 4 30000 63% 2539 22%
10 Crafty-23.5-2 2645 4 4 30000 63% 2539 22%
11 Crafty-23.5R06-1 2645 4 4 30000 63% 2539 22%
12 Crafty-23.5R06-8 2644 4 4 30000 63% 2539 22%
13 Crafty-23.5R06-5 2643 4 4 30000 63% 2539 22%
14 Crafty-23.5-1 2641 4 4 30000 63% 2539 22%
15 Crafty-23.5R06-2 2636 4 4 30000 62% 2539 22%


version 23.5-1 and 23.5-2 are simply two consecutive runs with the same version to provide a normal result. The rest of the tests are version 23.5R06 and were tested where the -n is the aspiration window (delta value in the code posted yesterday). 23.5R06-1 means the aspiration window was +/- 1 with delta=1. 1 and 2 are a bit low, and by the time it gets to 10, it is pretty optimal. Bigger doesn't seem to hurt at all up to +/- 3.0 pawns... I was expecting a better result in the 20-40 range, the reason I ran the big numbers was to produce some worse results so that there is a recognizable curve with a clear optimal value, and worse results on either side. Didn't get exactly what I expected, as you can see...

You should use CLOP

Rémi

I intend on looking at it. But I can test this so simply at present, I just say "runtest" and it runs the test with each parameter change as needed. Of course it is not optimally tuning a parameter, just using the choices I give...

[bob]

Here is the output from BayesElo first:

2 Crafty-23.5R06-200 2650 4 4 30000 64% 2539 22%
3 Crafty-23.5R06-24 2650 4 4 30000 64% 2539 22%
4 Crafty-23.5R06-100 2649 4 4 30000 63% 2539 22%
5 Crafty-23.5R06-30 2648 4 4 30000 63% 2539 22%
6 Crafty-23.5R06-50 2648 4 4 30000 63% 2539 22%
7 Crafty-23.5R06-300 2648 4 4 30000 63% 2539 22%
8 Crafty-23.5R06-20 2648 4 4 30000 63% 2539 22%
9 Crafty-23.5R06-10 2645 4 4 30000 63% 2539 22%
10 Crafty-23.5-2 2645 4 4 30000 63% 2539 22%
11 Crafty-23.5R06-1 2645 4 4 30000 63% 2539 22%
12 Crafty-23.5R06-8 2644 4 4 30000 63% 2539 22%
13 Crafty-23.5R06-5 2643 4 4 30000 63% 2539 22%
14 Crafty-23.5-1 2641 4 4 30000 63% 2539 22%
15 Crafty-23.5R06-2 2636 4 4 30000 62% 2539 22%


version 23.5-1 and 23.5-2 are simply two consecutive runs with the same version to provide a normal result. The rest of the tests are version 23.5R06 and were tested where the -n is the aspiration window (delta value in the code posted yesterday). 23.5R06-1 means the aspiration window was +/- 1 with delta=1. 1 and 2 are a bit low, and by the time it gets to 10, it is pretty optimal. Bigger doesn't seem to hurt at all up to +/- 3.0 pawns... I was expecting a better result in the 20-40 range, the reason I ran the big numbers was to produce some worse results so that there is a recognizable curve with a clear optimal value, and worse results on either side. Didn't get exactly what I expected, as you can see...

What is the average depth?

A quick looks says that the average is in the 13-14-15 range. I looked at a few logs and there are plenty of re-searches going on, so it is having a chance to exert influence.

[Rebel]

Here is the output from BayesElo first:

2 Crafty-23.5R06-200 2650 4 4 30000 64% 2539 22%
3 Crafty-23.5R06-24 2650 4 4 30000 64% 2539 22%
4 Crafty-23.5R06-100 2649 4 4 30000 63% 2539 22%
5 Crafty-23.5R06-30 2648 4 4 30000 63% 2539 22%
6 Crafty-23.5R06-50 2648 4 4 30000 63% 2539 22%
7 Crafty-23.5R06-300 2648 4 4 30000 63% 2539 22%
8 Crafty-23.5R06-20 2648 4 4 30000 63% 2539 22%
9 Crafty-23.5R06-10 2645 4 4 30000 63% 2539 22%
10 Crafty-23.5-2 2645 4 4 30000 63% 2539 22%
11 Crafty-23.5R06-1 2645 4 4 30000 63% 2539 22%
12 Crafty-23.5R06-8 2644 4 4 30000 63% 2539 22%
13 Crafty-23.5R06-5 2643 4 4 30000 63% 2539 22%
14 Crafty-23.5-1 2641 4 4 30000 63% 2539 22%
15 Crafty-23.5R06-2 2636 4 4 30000 62% 2539 22%


version 23.5-1 and 23.5-2 are simply two consecutive runs with the same version to provide a normal result. The rest of the tests are version 23.5R06 and were tested where the -n is the aspiration window (delta value in the code posted yesterday). 23.5R06-1 means the aspiration window was +/- 1 with delta=1. 1 and 2 are a bit low, and by the time it gets to 10, it is pretty optimal. Bigger doesn't seem to hurt at all up to +/- 3.0 pawns... I was expecting a better result in the 20-40 range, the reason I ran the big numbers was to produce some worse results so that there is a recognizable curve with a clear optimal value, and worse results on either side. Didn't get exactly what I expected, as you can see...

What is the average depth?

A quick looks says that the average is in the 13-14-15 range. I looked at a few logs and there are plenty of re-searches going on, so it is having a chance to exert influence.

What should not be forgotten is the saturation of the hash table. With an almost full hash table researches will become very expensive and it would make sense to widen the window.

[bob]

Here is the output from BayesElo first:

2 Crafty-23.5R06-200 2650 4 4 30000 64% 2539 22%
3 Crafty-23.5R06-24 2650 4 4 30000 64% 2539 22%
4 Crafty-23.5R06-100 2649 4 4 30000 63% 2539 22%
5 Crafty-23.5R06-30 2648 4 4 30000 63% 2539 22%
6 Crafty-23.5R06-50 2648 4 4 30000 63% 2539 22%
7 Crafty-23.5R06-300 2648 4 4 30000 63% 2539 22%
8 Crafty-23.5R06-20 2648 4 4 30000 63% 2539 22%
9 Crafty-23.5R06-10 2645 4 4 30000 63% 2539 22%
10 Crafty-23.5-2 2645 4 4 30000 63% 2539 22%
11 Crafty-23.5R06-1 2645 4 4 30000 63% 2539 22%
12 Crafty-23.5R06-8 2644 4 4 30000 63% 2539 22%
13 Crafty-23.5R06-5 2643 4 4 30000 63% 2539 22%
14 Crafty-23.5-1 2641 4 4 30000 63% 2539 22%
15 Crafty-23.5R06-2 2636 4 4 30000 62% 2539 22%


version 23.5-1 and 23.5-2 are simply two consecutive runs with the same version to provide a normal result. The rest of the tests are version 23.5R06 and were tested where the -n is the aspiration window (delta value in the code posted yesterday). 23.5R06-1 means the aspiration window was +/- 1 with delta=1. 1 and 2 are a bit low, and by the time it gets to 10, it is pretty optimal. Bigger doesn't seem to hurt at all up to +/- 3.0 pawns... I was expecting a better result in the 20-40 range, the reason I ran the big numbers was to produce some worse results so that there is a recognizable curve with a clear optimal value, and worse results on either side. Didn't get exactly what I expected, as you can see...

What is the average depth?

A quick looks says that the average is in the 13-14-15 range. I looked at a few logs and there are plenty of re-searches going on, so it is having a chance to exert influence.

What should not be forgotten is the saturation of the hash table. With an almost full hash table researches will become very expensive and it would make sense to widen the window.

That's one problem I really don't have to deal with. I don't hash the q-search, so I really don't see much in terms of saturation. When I was testing this stuff, I found that hashing the q-search reduced the total tree size by about 10%, but it slowed the program down by almost exactly the same amount. A wash. But by not hashing the q-search, the stress on the ttable is greatly reduced, which tends to make this pay off (not hashing qsearch) in real long games without sufficient memory for the ttable. 8 gigs gives 512 million entries, which is a lot even for a long 40/2hr type game...

I run fast games with a modest hash on the cluster, just to try to keep things within perspective...