Dragon v4.7

[Krzysztof Grzelak]

Thank you for the new version of the program.

[Sidiki]

Hi Michel,

What weights to use please?
I launched et dxp match and so "missing weights, poor performance"

Friendly, Sidiki

[MichelG]

Hi Michel,

What weights to use please?
I launched et dxp match and so "missing weights, poor performance"

Friendly, Sidiki

Hi Sidiki,

You can ignore the warnings. They are about the old evaluation function, but dragon does not use them any more. I will remove the messages in a new version later.

Michel

[Sidiki]

Hi Michel,

What weights to use please?
I launched et dxp match and so "missing weights, poor performance"

Friendly, Sidiki

Hi Sidiki,

You can ignore the warnings. They are about the old evaluation function, but dragon does not use them any more. I will remove the messages in a new version later.

Michel

Hi Michel,

Thanks very much for the reply.

Friendly ,Sidiki .

[Krzysztof Grzelak]

Hi Michel,
I've been waiting for 3 days and nothing is happening. Can you fix this or do something about it. Applies to program version 5.0. Screenshot from version 4.8.0 but the same thing happens.



Krzysztof

[MichelG]

Version 5.0.1 is now available

It fixes the "missing weights, poor performance" message. In retrospect it may have affected strength in some cases after all.

5.0.1 is much stronger than 5.0. In my internal benchmark the score moved from 42.5% win to 47.2% wins against scan 3.1. A yet to be released version does about 49.7% and i can see a clear path for further improvement.

5.0.1 introduces a new neural net for endgames with 8 pieces or less, which is used for the endgames that are not installed. Also the neural net is now split into positions with and without kings. Training is based on 607 milion positions that have been fully rolled out.

@Krzysztof if the database generator problem still persist, can you post a new image, and indicate which of the 3 buttons you used to start it?

[Sidiki]

Version 5.0.1 is now available

It fixes the "missing weights, poor performance" message. In retrospect it may have affected strength in some cases after all.

5.0.1 is much stronger than 5.0. In my internal benchmark the score moved from 42.5% win to 47.2% wins against scan 3.1. A yet to be released version does about 49.7% and i can see a clear path for further improvement.

5.0.1 introduces a new neural net for endgames with 8 pieces or less, which is used for the endgames that are not installed. Also the neural net is now split into positions with and without kings. Training is based on 607 milion positions that have been fully rolled out.

@Krzysztof if the database generator problem still persist, can you post a new image, and indicate which of the 3 buttons you used to start it?

Hi Michel,

Thanks again for this update.
It seem that strenght reached a good level.

Friendly, Sidiki

[Krzysztof Grzelak]

Hi Michel,
I present a problem with generating databases. Program version 5.0. After starting the program, I select options, then Endgame and Create endgame databases.I have this window - I select the options marked in red. Applies to 5v2 and 4v3 bases. And nothing happens for several days.

[Joost Buijs]

I noticed that on the Dragon Database Statistics website the counts for the sub-database 1221 are missing, are these counts (capture/non capture) by any means known (including the depth-counts for DTW)?

[Ed Gilbert]

Joost, I have this list of WLD counts for 2 through 7 pieces in my db. IIRC the counts exclude direct captures.

Notation: bBwW
0101b: wins 16, draws 2024, losses 2
0101w: wins 16, draws 2024, losses 2
0110b: wins 1516, draws 281, losses 0
0110w: wins 1, draws 574, losses 1510
1010b: wins 616, draws 799, losses 450
1010w: wins 616, draws 799, losses 450
0201b: wins 2330, draws 39402, losses 0
0201w: wins 0, draws 40920, losses 236
0210b: wins 33973, draws 1879, losses 0
0210w: wins 0, draws 7807, losses 39602
1101b: wins 2969, draws 81465, losses 2
1101w: wins 15, draws 72355, losses 150
1110b: wins 64099, draws 9776, losses 1
1110w: wins 0, draws 21465, losses 63377
2001b: wins 337, draws 30939, losses 11312
2001w: wins 14028, draws 17809, losses 7
2010b: wins 23532, draws 13754, losses 572
2010w: wins 1858, draws 17204, losses 18801
0202b: wins 10519, draws 697448, losses 11
0202w: wins 10519, draws 697448, losses 11
0211b: wins 81329, draws 1138856, losses 3
0211w: wins 11250, draws 1581692, losses 5833
0220b: wins 433166, draws 90324, losses 0
0220w: wins 1308, draws 337495, losses 561713
0301b: wins 134804, draws 434756, losses 0
0301w: wins 0, draws 528990, losses 17210
0310b: wins 469575, draws 7865, losses 0
0310w: wins 0, draws 78063, losses 627030
1111b: wins 66781, draws 2734001, losses 1537
1111w: wins 66781, draws 2734001, losses 1537
1120b: wins 805687, draws 418002, losses 65
1120w: wins 6734, draws 831965, losses 763459
1201b: wins 257206, draws 1452834, losses 0
1201w: wins 0, draws 1426390, losses 19115
1210b: wins 1405115, draws 56765, losses 0
1210w: wins 0, draws 275813, losses 1617548
2020b: wins 170641, draws 451336, losses 88695
2020w: wins 170641, draws 451336, losses 88695
2101b: wins 141334, draws 1567532, losses 2
2101w: wins 13, draws 1265493, losses 5672
2110b: wins 1333265, draws 153291, losses 2
2110w: wins 0, draws 374087, losses 1316430
3001b: wins 17373, draws 537278, losses 13271
3001w: wins 30562, draws 340477, losses 364
3010b: wins 366308, draws 135152, losses 242
3010w: wins 2739, draws 191952, losses 307138
0302b: wins 477082, draws 7835942, losses 0
0302w: wins 33564, draws 7952722, losses 1310
0311b: wins 3472867, draws 10540561, losses 0
0311w: wins 33918, draws 19541907, losses 396123
0320b: wins 5568878, draws 304486, losses 0
0320w: wins 4569, draws 3348598, losses 9121448
0401b: wins 5817506, draws 20558, losses 0
0401w: wins 0, draws 1826, losses 5372442
0410b: wins 4778684, draws 0, losses 0
0410w: wins 0, draws 2831, losses 7710421
1202b: wins 999766, draws 26558244, losses 4
1202w: wins 392679, draws 20323616, losses 520
1211b: wins 6668510, draws 40619881, losses 2
1211w: wins 400091, draws 51843444, losses 359006
1220b: wins 17867630, draws 2327578, losses 1
1220w: wins 47385, draws 10878582, losses 22389102
1301b: wins 22874114, draws 225028, losses 0
1301w: wins 0, draws 151242, losses 18838998
1310b: wins 19294022, draws 0, losses 0
1310w: wins 0, draws 33214, losses 27592088
2102b: wins 513014, draws 29883099, losses 42974
2102w: wins 1140669, draws 16697115, losses 68
2111b: wins 3370895, draws 49697252, losses 5953
2111w: wins 925412, draws 45028366, losses 74540
2120b: wins 17206795, draws 5843384, losses 214
2120w: wins 86449, draws 13163908, losses 16326844
2201b: wins 33333880, draws 928409, losses 3
2201w: wins 19, draws 1120407, losses 23966664
2210b: wins 29148467, draws 693, losses 2
2210w: wins 0, draws 119525, losses 36891780
3002b: wins 43721, draws 6345691, losses 4800884
3002w: wins 3483768, draws 1614055, losses 0
3011b: wins 368406, draws 15078074, losses 4347100
3011w: wins 5285093, draws 8089131, losses 4181
3020b: wins 4117646, draws 4422451, losses 186597
3020w: wins 630419, draws 5370435, losses 2727283
3101b: wins 20716598, draws 1844239, losses 1
3101w: wins 6, draws 2513106, losses 12169448
3110b: wins 19510908, draws 3809, losses 1
3110w: wins 0, draws 221654, losses 21759842
4001b: wins 3337999, draws 2215868, losses 6219
4001w: wins 10612, draws 1906697, losses 1294363
4010b: wins 4793612, draws 87717, losses 2
4010w: wins 1210, draws 308320, losses 4572916
0303b: wins 1102432, draws 80327686, losses 32
0303w: wins 1102432, draws 80327686, losses 32
0312b: wins 13696756, draws 193362214, losses 7
0312w: wins 2614961, draws 293115197, losses 9004
0321b: wins 41201559, draws 133415523, losses 0
0321w: wins 1420237, draws 353500438, losses 3388782
0330b: wins 43719979, draws 5106126, losses 0
0330w: wins 128698, draws 61866696, losses 82692145
0402b: wins 13231266, draws 61577205, losses 0
0402w: wins 69348, draws 63745250, losses 3720055
0411b: wins 115152007, draws 8292207, losses 0
0411w: wins 74687, draws 49985110, losses 135244212
0420b: wins 50529263, draws 52280, losses 0
0420w: wins 12235, draws 5859346, losses 121442757
0501b: wins 47821032, draws 41552, losses 0
0501w: wins 0, draws 5724, losses 41848548
0510b: wins 38328164, draws 0, losses 0
0510w: wins 0, draws 2286, losses 66169926
1212b: wins 30877933, draws 733184209, losses 1495
1212w: wins 30877933, draws 733184209, losses 1495
1221b: wins 86768707, draws 568453887, losses 260
1221w: wins 16683337, draws 919726026, losses 2830156
1230b: wins 143850456, draws 42613174, losses 49
1230w: wins 1452604, draws 194469458, losses 188395854
1302b: wins 36530982, draws 286860420, losses 0
1302w: wins 1220006, draws 227382860, losses 5739557
1311b: wins 482825538, draws 60605020, losses 0
1311w: wins 1283501, draws 216665045, losses 433999804
1320b: wins 226119869, draws 891737, losses 0
1320w: wins 168777, draws 23490823, losses 430002382
1401b: wins 233584933, draws 332217, losses 0
1401w: wins 0, draws 78993, losses 185068200
1410b: wins 191109550, draws 0, losses 0
1410w: wins 0, draws 19517, losses 296324931
2121b: wins 44320225, draws 773081052, losses 380608
2121w: wins 44320225, draws 773081052, losses 380608
2130b: wins 133245978, draws 103206511, losses 11884
2130w: wins 3832201, draws 219776697, losses 115668391
2202b: wins 33877337, draws 490966511, losses 134
2202w: wins 5294580, draws 295905166, losses 2566311
2211b: wins 725666884, draws 171063983, losses 213
2211w: wins 5534146, draws 362577315, losses 489269775
2220b: wins 377009568, draws 4223499, losses 81
2220w: wins 631884, draws 40275061, losses 563676841
2301b: wins 456377028, draws 1154417, losses 1
2301w: wins 8, draws 459923, losses 326181325
2310b: wins 380834496, draws 0, losses 0
2310w: wins 0, draws 78156, losses 529476315
3030b: wins 20789593, draws 69794834, losses 8947685
3030w: wins 20789593, draws 69794834, losses 8947685
3102b: wins 11146803, draws 367330453, losses 116698
3102w: wins 8499662, draws 165526467, losses 273616
3111b: wins 344925367, draws 311780541, losses 13329
3111w: wins 7905328, draws 348982472, losses 142555504
3120b: wins 270761658, draws 12916370, losses 705
3120w: wins 750271, draws 51243509, losses 305099868
3201b: wins 445098954, draws 2254223, losses 5
3201w: wins 28, draws 1355821, losses 285888587
3210b: wins 378815431, draws 47, losses 4
3210w: wins 0, draws 191611, losses 471739949
4002b: wins 780249, draws 76478438, losses 25050582
4002w: wins 17757371, draws 19587067, losses 2195
4011b: wins 27368387, draws 143152535, losses 9404233
4011w: wins 17354654, draws 86770497, losses 4590483
4020b: wins 61199403, draws 17530177, losses 120406
4020w: wins 1272655, draws 27974496, losses 49614455
4101b: wins 216002317, draws 2481538, losses 1
4101w: wins 6, draws 1927758, losses 123963350
4110b: wins 187939689, draws 191, losses 1
4110w: wins 0, draws 256423, losses 209468705
5001b: wins 40275014, draws 2326739, losses 6442
5001w: wins 6110, draws 2003040, losses 19985818
5010b: wins 37166724, draws 11431, losses 0
5010w: wins 517, draws 283980, losses 36891266
0403b: wins 22947628, draws 624882484, losses 0
0403w: wins 1712978, draws 599580394, losses 2251058
0412b: wins 283913660, draws 1332445036, losses 0
0412w: wins 4648489, draws 2283871173, losses 92442071
0421b: wins 1138292886, draws 197838832, losses 0
0421w: wins 2808661, draws 1573165088, losses 1561645302
0430b: wins 363662706, draws 2156778, losses 0
0430w: wins 296975, draws 168572305, losses 1211671344
0502b: wins 489226234, draws 59194628, losses 0
0502w: wins 175224, draws 80452496, losses 376519010
0511b: wins 885686145, draws 1562537, losses 0
0511w: wins 210906, draws 125860616, losses 1236942539
0520b: wins 355950583, draws 277, losses 0
0520w: wins 41269, draws 21121224, losses 999519077
1303b: wins 75041376, draws 2964492283, losses 26
1303w: wins 31976341, draws 2023408571, losses 2553027
1312b: wins 893300142, draws 6813961789, losses 25
1312w: wins 80008423, draws 8013813737, losses 139511210
1321b: wins 4981650641, draws 1499012585, losses 10
1321w: wins 46295542, draws 6370190391, losses 4574983194
1330b: wins 1775068306, draws 31491058, losses 1
1330w: wins 4275925, draws 705003928, losses 4185372540
1402b: wins 2420945170, draws 478103942, losses 0
1402w: wins 2919505, draws 505610056, losses 1481333376
1411b: wins 4758418717, draws 17191554, losses 0
1411w: wins 3455934, draws 650775239, losses 5351633668
1420b: wins 1953253420, draws 4689, losses 0
1420w: wins 556644, draws 96725305, losses 4452095825
2203b: wins 78200925, draws 5281029059, losses 45804
2203w: wins 167951930, draws 2450549147, losses 703986
2212b: wins 925342523, draws 12867400378, losses 13442
2212w: wins 401236248, draws 10176739540, losses 56094484
2221b: wins 7019056270, draws 4761276537, losses 7709
2221w: wins 231885784, draws 10071366000, losses 4086903333
2230b: wins 3167431065, draws 170903135, losses 1861
2230w: wins 20578346, draws 1269512002, losses 5198855329
2302b: wins 4198230472, draws 1945895748, losses 47
2302w: wins 16862729, draws 1637041099, losses 1798168572
2311b: wins 10182489629, draws 110029711, losses 46
2311w: wins 19578663, draws 1420700913, losses 9112328689
2320b: wins 4285400830, draws 67685, losses 4
2320w: wins 2580705, draws 173577085, losses 7913400965
3103b: wins 27509423, draws 4159173566, losses 17997625
3103w: wins 308558016, draws 1165822996, losses 6169
3112b: wins 331139648, draws 10626204330, losses 12840692
3112w: wins 699473807, draws 5376500875, losses 2944008
3121b: wins 2207921855, draws 7295616081, losses 1359296
3121w: wins 390368689, draws 7449981971, losses 503075535
3130b: wins 2222661901, draws 511643243, losses 36832
3130w: wins 32063859, draws 1438257550, losses 2341507574
3202b: wins 2766598999, draws 3752544678, losses 1448
3202w: wins 39277759, draws 2265684610, losses 678102698
3211b: wins 10629556269, draws 462037384, losses 2014
3211w: wins 44712860, draws 1739337215, losses 7456542312
3220b: wins 4691761177, draws 2864648, losses 617
3220w: wins 4975720, draws 174292186, losses 6993155833
4003b: wins 1944214, draws 708582014, losses 526780449
4003w: wins 251924262, draws 57727265, losses 0
4012b: wins 23901292, draws 2171713568, losses 1073416435
4012w: wins 819983024, draws 477436143, losses 3838
4021b: wins 136394440, draws 2249042761, losses 484459502
4021w: wins 674883270, draws 1125942653, losses 6475631
4030b: wins 422782725, draws 397218117, losses 17051943
4030w: wins 69994641, draws 495944034, losses 271109700
4102b: wins 729618536, draws 2729677088, losses 162285
4102w: wins 39952296, draws 1169006275, losses 74833267
4111b: wins 5193794066, draws 776854157, losses 22047
4111w: wins 43479839, draws 1167655012, losses 2820935999
4120b: wins 2548903337, draws 16608246, losses 1178
4120w: wins 4088648, draws 131122372, losses 3035223931
5002b: wins 53579246, draws 601212602, losses 78987756
5002w: wins 58375294, draws 159922205, losses 1784475
5011b: wins 669863526, draws 605186934, losses 8116476
5011w: wins 30414574, draws 444409694, losses 226409920
5020b: wins 526663112, draws 32319102, losses 31098

[Joost Buijs]

Joost, I have this list of WLD counts for 2 through 7 pieces in my db. IIRC the counts exclude direct captures.

Ed, thanks!

This helps me to complete the comparison.

When compressing I exclude direct captures too, because of this the size of the 7P database has shrunk to 30 GB.

Joost

[Ed Gilbert]

Joost, as you may know, there are 2 versions of the kingsrow egdb. In version 1, I only exclude direct captures. In version 2, in addition to direct captures, for positions of 7 or more pieces that have more than 1 king present, I also exclude non-side captures (would-be captures if the side-to-move was reversed), and I also for this same subset I removed all the positions for one of the sides-to-move (which ever side took up more space on disk). These changes for version 2, plus some improvement in the compression algorithm, made a huge difference in the size of the db. You might think that excluding all these positions would have some adverse effect on engine strength, or on its ability of the engine search to return actual db win, loss, or draw values instead of heuristic values. But I found through testing that it does not. There was no measurable difference in either based on my tests, which were engine matches and also a list of test positions that I used to measure the time for the search to return a specific egdb score. Between all these excluded positions, and also from not loading any uncached positions from disk at or near a leaf node, it means that a large majority of positions visited during an engine search with few enough pieces to be in the egdb do not get a db result to use.

The only disadvantage that I have found from not having all those positions is when I am using the db in some other application, not in the kingsrow engine, where I want to get the WLD value of every position in a slice. In those cases, when I need the value of an excluded position, I have to do a recursive tree search and go deep enough to be able to propagate a guaranteed WLD value back to the root. This is a lot slower than just probing the db to get the value. The recursive tree search often has to do hundreds and sometimes thousands of egdb probes. In a very small percentage of cases, I think less than 0.1 %, the tree search can take longer than a second. But typically it is microseconds or a few milliseconds. A single probe from a position in cache is nanoseconds.

Comparing the db sizes:
2 through 8 pieces
version 1: 389 GB
version 2: 55.6 GB

Of course some of that difference is due to compression improvements, but it's relatively minor. Comparing only 2 through 6 pieces, which have the exact same exclusions
version 1: 1.07 GB
version 2: 0.94 GB

-- Ed

[Joost Buijs]

Ed, I heard from Bert that he doesn’t include positions with a 'non-side' capture in his database either. However, I’ve decided to include them because Ares frequently encounters positions where a 'non-side' capture needs to be evaluated. On a modern PC, it matters very little whether a database footprint is 27 GB or 30 GB.

I’m currently using 4K blocks and a cache of 256 blocks per thread. For each block, I maintain a CRC32 checksum which can optionally be used during probing. With 7P, a cache hit takes about 40 ns, whereas a cache miss can take 5 µs. I'm currently working on pushing the cache hit-rate a bit higher.

Unfortunately, I still haven't found the WDL-counts for the 1221 database (1 white man, 2 white kings, 2 black men, 1 black king). Since the rest of the databases match up, this one is likely correct too.

Joost

[Ed Gilbert]

Ed, I heard from Bert that he doesn’t include positions with a 'non-side' capture in his database either. However, I’ve decided to include them because Ares frequently encounters positions where a 'non-side' capture needs to be evaluated. On a modern PC, it matters very little whether a database footprint is 27 GB or 30 GB.

The difference is much greater than 30 to 27. I ran a test for Bert, compressing 2 through 6 pieces and excluding both captures and non-side captures. That one change to additionally exclude the non-side captures decreased the size from 0.94 GB to 0.451 GB.

The reason I don't exclude non-side captures for 2 to 6 pieces or for positions with 0 or 1 kings is that those slices are already relatively small and also some of the most useful.

In the kingsrow Qsearch, I also extend when there are non-side captures, although there are some constraints to prevent it from ballooning to excessive depths. But in general I found that if I cannot get an egdb result at any of the kingsrow search nodes, I don't do anything extraordinary. I just treat it the same way I treat any other node where I cannot probe the egdb. If I need an eval score, I just call the heuristic eval function. At interior nodes I don't normally need eval scores. I don't do anything like futility pruning that might require a static eval score. I only probe the egdb at interiors to try to get a quick fail-high cutoff or a cutoff from a db draw.

The per-thread egdb cache is an interesting idea. I didn't realize you could read a small block from an SSD in 5 us. The last time I benchmarked this stuff was quite a while ago. I have a single egdb cache and use locks to protect critical regions.

-- Ed

[BertTuyt]

Joost,

not 100% sure, can you verify:

Format: configuration all_win all_draw all_lose, noncapture_win noncapture_draw noncapture_lose

1221 678795283 1395118040 7606227 86768707 568453887 260

Bert