The seventh release of Altair.
Many changes have been made in Altair including search features, its evaluation, and move generation. Two new neural networks have been created through self-data-generation, Taffreta and Solaris, with the latter being the current and strongest net trained on over 2B FENs.

A few large FRC (Chess 960) bugs have been fixed, resulting in much larger Elo gain compared to standard chess.

Standard LTC:

Elo   | 79.17 +- 12.12 (95%)
Conf  | 40.0+0.40s Threads=1 Hash=128MB
Games | N: 1000 W: 278 L: 54 D: 668
Penta | [0, 32, 243, 194, 31]
http:https://antares2262.pythonanywhere.com/test/183/

FRC LTC:

Elo   | 146.45 +- 15.41 (95%)
Conf  | 40.0+0.40s Threads=1 Hash=128MB
Games | N: 1002 W: 455 L: 56 D: 491
Penta | [1, 17, 162, 224, 97]
http:https://antares2262.pythonanywhere.com/test/184/

As always, huge thanks to everyone of the OpenBench Testing Instance that Altair is on for their support and contribution of hardware towards Altair's testing.
Thank you again to @Ciekce for helping me to train neural networks with his hardware.

The binaries are included below, with x86-64 binaries being based on the microarchitecture levels

It is recommended that most people use the v3 binary, with v4 meant for CPUs that support AVX512, and v2 and v1 for older CPUs (more details in the link above).

Version 6.0.0 marks a major release of Altair introducing NNUE (Efficiently Updatable Neural Networks).
Neural Networks have now replaced the use of a classical evaluation in Altair, resulting in large strength improvements.

As of version 5.0.1, Altair's evaluation weights were completely set to zero, allowing for the generation of completely original data from zero knowledge.
Altair's current net Trappist is the product of repeated data generation, training and testing.
It has a network architecture of (768->768)x2->1 and is trained on a combination of standard and DFRC FENs.

Other improvements in Altair include slight changes in its search and time management, further details are visible in the change-log.
Altair now also supports analysis with MultiPV along with evaluation normalization and WDL output.

Test Results:

Standard LTC:

ELO   | 374.11 +- 28.25 (95%)
CONF  | 40.0+0.40s Threads=1 Hash=128MB
GAMES | N: 1000 W: 815 L: 23 D: 162
https://chess.swehosting.se/test/4904/

Fischer Random LTC:

ELO   | 415.05 +- 35.84 (95%)
CONF  | 40.0+0.40s Threads=1 Hash=128MB
GAMES | N: 1000 W: 873 L: 41 D: 86
https://chess.swehosting.se/test/4908/

Once again, all my thanks to everyone in the OpenBench Testing Instance for their tremendous support and contribution of hardware towards Altair's testing.
Thank you to @Ciekce for helping me to train a few of my neural networks with his hardware, and also helping me implement NNUE.
Also, thanks to @jw1912 for his Bullet trainer which Altair's nets have been trained with.

The binaries are included below, with x86-64 binaries being based on the microarchitecture levels

It is recommended that most people use the v3 binary, with v4 meant for CPUs that support AVX512, and v2 and v1 for older CPUs (more details in the link above).

This is the fifth and final HCE (Hand Crafted Evaluation) version of Altair. The next planned release will introduce Neural Networks.

Altair has experienced a large rewrite of its board representation, move generation, and evaluation in order to incorporate bitboards in this version. Along with the introduction of bitboards, Altair has had many search patches which contribute to its strength improvements.

Standard Chess LTC

ELO   | 121.85 +- 15.29 (95%)
CONF  | 40.0+0.40s Threads=1 Hash=128MB
GAMES | N: 1000 W: 418 L: 81 D: 501
https://chess.swehosting.se/test/3975/

FRC / Chess960 LTC

ELO   | 138.18 +- 17.37 (95%)
CONF  | 40.0+0.40s Threads=1 Hash=128MB
GAMES | N: 1000 W: 491 L: 113 D: 396
https://chess.swehosting.se/test/3980/

Major thanks to @archishou and @Ciekce for helping me with the bitboard rewrite of Altair. And per usual, much thanks to everyone here OB testing instance for contributing resources towards Altair's testing.

The binaries are included below, with windows binaries being based on the microarchitecture levels

It is recommended that most people use the v3 binary, with v4 meant for CPUs that support AVX512, and v2 and v1 for older CPUs.

This is the fourth version of Altair. Altair now supports Fischer Random Chess (Chess 960) by toggling the UCI option.

Compared to previous releases, this release is not as major, and this is due to the release primarily marking the final version of Altair using an internal mailbox-based board representation. Due to the mailbox-based board representation, Altair is comparatively slow and has much more potential utilizing bitboards.

Altair is now at an estimated 3080 Elo, and almost all the patches since v3.0.0 have been search or time management related.

Self-play LTC test:

ELO   | 85.41 +- 13.46 (95%)
CONF  | 40.0+0.40s Threads=1 Hash=128MB
GAMES | N: 1000 W: 320 L: 79 D: 601

Thanks to @archishou for a speed-up patch, and thanks to everyone in this OB testing instance for contributing invaluable resources that have enabled me to improve Altair.

This is the third version of Altair. The engine is at an estimated 3000+ Elo and has gained around 160 Elo over the previous version in self-play. Most of the improvements in this version have been from development in Altair's search, with a few evaluation improvements sprinkled in. A huge thanks and acknowledgement to everyone in this Testing Instance for contributing resources and invaluable support to the development of Altair.

Altair now supports multi-threading and the usage of multiple CPUs to make it stronger.

Score of Altair3.0.0 vs Topple0.8.1: 144 - 144 - 177 [0.500]
...      Altair3.0.0 playing White: 73 - 68 - 92  [0.511] 233
...      Altair3.0.0 playing Black: 71 - 76 - 85  [0.489] 232
...      White vs Black: 149 - 139 - 177  [0.511] 465
Elo difference: 0.0 +/- 24.9, LOS: 50.0 %, DrawRatio: 38.1 %
473 of 5000 games finished.

Score of Altair3.0.0 vs Altair2.0.0: 59 - 7 - 53 [0.718]
...      Altair3.0.0 playing White: 30 - 4 - 25  [0.720] 59
...      Altair3.0.0 playing Black: 29 - 3 - 28  [0.717] 60
...      White vs Black: 33 - 33 - 53  [0.500] 119
Elo difference: 162.8 +/- 46.9, LOS: 100.0 %, DrawRatio: 44.5 %
119 of 5000 games finished.

This is the second version of Altair. There have been around 150-180 ELO rating gains from changes in its search, and the rest 250+ have been from tuning the evaluation and new evaluation features.

A huge thanks to Gedas, as many of the rating gains are the result of tuning using his texel tuner.

8 + 0.08 Versus Altair1.0.0
Score of Altair1.5.6 vs Altair1.0.0: 89 - 1 - 10 [0.940]
...      Altair1.5.6 playing White: 46 - 1 - 4  [0.941] 51
...      Altair1.5.6 playing Black: 43 - 0 - 6  [0.939] 49
...      White vs Black: 46 - 44 - 10  [0.510] 100
Elo difference: 478.0 +/- 121.0, LOS: 100.0 %, DrawRatio: 10.0 %```

60 + 0.6 Versus Glaurung2.2 1CPU
Score of Altair1.5.6 vs glaurung-w64: 155 - 150 - 82 [0.506]
...      Altair1.5.6 playing White: 75 - 77 - 42  [0.495] 194
...      Altair1.5.6 playing Black: 80 - 73 - 40  [0.518] 193
...      White vs Black: 148 - 157 - 82  [0.488] 387
Elo difference: 4.5 +/- 30.8, LOS: 61.3 %, DrawRatio: 21.2 %

The engine should be around 2800-2850 in CCRL ratings.

Release Version of Altair. This version is a port of Antares to c++. It has additional features that make it stronger too.

1 minute games against SofCheck 0.9

...      Altair0.47 playing White: 20 - 21 - 19  [0.492] 60
...      Altair0.47 playing Black: 23 - 25 - 12  [0.483] 60
...      White vs Black: 45 - 44 - 31  [0.504] 120
Elo difference: -8.7 +/- 53.9, LOS: 37.5 %, DrawRatio: 25.8 %
120 of 120 games finished.