seer-v2.8.0

This release adds an estimated 50 Elo in self play under UHO testing conditions. This improvement in strength can be attributed to an improved network fine-tuned exclusively on a self play game result (WDL) target, improved repetition detection, SPSA search parameter tuning and a number of miscellaneous search refinements. As was the case for the prior release, the embedded network was trained solely on data originating from Seer produced by way of many self play iterations starting with the initial "retrograde learning" network bundled with v2.0.0-v2.1.0 as a base.

Elo   | 52.22 +- 3.42 (95%)
Conf  | 40.0+0.40s Threads=1 Hash=64MB
Games | N: 20002 W: 6531 L: 3547 D: 9924
Penta | [33, 1217, 4716, 3803, 232]
http://chess.grantnet.us/test/34901/

Thanks to everyone involved with the OpenBench project for making Seer's continued development possible!

seer-v2.7.0

This release adds an estimated 50 Elo in self play under UHO testing conditions. This improvement in strength can be attributed to further optimizations to the network inference implementation, an improved network trained using a new regularization scheme, and a number of miscellaneous search improvements in roughly equal measure. As was the case for the prior release, the embedded network was trained solely on data originating from Seer produced by way of many self play iterations starting with the initial "retrograde learning" network bundled with v2.0.0-v2.1.0 as a base.

Elo   | 52.37 +- 3.38 (95%)
Conf  | 40.0+0.40s Threads=1 Hash=64MB
Games | N: 20000 W: 6442 L: 3450 D: 10108
Penta | [45, 1230, 4639, 3860, 226]
http://chess.grantnet.us/test/34137/

Thanks to everyone involved with the OpenBench project for making Seer's continued development possible!

seer-v2.6.0

This release is estimated to add 100 Elo in self play under UHO testing conditions. This improvement in strength can be attributed to both a number of further optimizations to the network inference implementation and an improved network trained by way of several additional self play training iterations in roughly equal measure. As with the prior release, the embedded network was trained solely on data originating from Seer produced by way of many self play iterations starting with the initial "retrograde learning" network bundled with v2.0.0-v2.1.0 as a base.

ELO   | 108.05 +- 3.46 (95%)
CONF  | 40.0+0.40s Threads=1 Hash=64MB
GAMES | N: 20000 W: 8175 L: 2148 D: 9677

Thanks to everyone involved with the OpenBench project for making Seer's continued development possible!

seer-v2.5.0

This release adds roughly 100 Elo in self play. The majority of this strength improvement can be attributed to a variety of miscellaneous optimizations and an improved network trained by way of several further self play training iterations. Additionally, time management was revisited with this release which should notably improve performance at shorter time controls. As with the prior release, the embedded network was trained solely on data originating from Seer produced by way of several self play iterations starting with the initial "retrograde learning" network bundled with v2.0.0-v2.1.0 as a base.

ELO   | 106.30 +- 2.93 (95%)
CONF  | 40.0+0.40s Threads=1 Hash=64MB
GAMES | N: 20000 W: 6913 L: 978 D: 12109

Thanks to everyone involved with the OpenBench project for making Seer's continued development feasible. Discussing various search ideas with fellow OpenBench engine authors proved invaluable in further improving Seer!

seer-v2.4.0

This release adds approximately 100 elo in self play. This increase in strength can be attributed to both further refinements to the search function and an improved network in roughly equal measure. As with the previous release, the network was trained solely on data originating from Seer produced by way of several self play iterations starting with the initial "retrograde learning" network bundled with v2.0.0-v2.1.0 as a base. Associated training code can be found at the "seer-training" repository: https://github.com/connormcmonigle/seer-training/tree/shared-affine.

Notably, this release adds support for both ponder and Syzygy EGTB probing.

ELO   | 117.96 +- 3.52 (95%)
CONF  | 8.0+0.08s Threads=1 Hash=32MB
GAMES | N: 20000 W: 8576 L: 2035 D: 9389

ELO   | 96.85 +- 3.04 (95%)
CONF  | 40.0+0.40s Threads=1 Hash=64MB
GAMES | N: 20000 W: 6859 L: 1424 D: 11717

As with the previous release, thanks to everyone involved with the OpenBench project for making Seer's continued development feasible. Discussing various search ideas with fellow OpenBench engine authors proved invaluable in further improving Seer's search function.

seer-v2.3.0

This release adds an estimated 100 elo in self play. As with the previous release, the majority of this increase in strength originates from an improved network produced by further self play iterations starting with the initial "retrograde learning" bundled with v2.0.0-v2.1.0. Notably, the network was trained on a mixture of self play game result and search score which yielded far superior results in my testing.

ELO   | 107.29 +- 3.52 (95%)
CONF  | 40.0+0.40s Threads=1 Hash=64MB
GAMES | N: 20000 W: 8299 L: 2312 D: 9389

Thanks to all OpenBench contributors for making Seer's continued development possible!

seer-v2.2.0

This release adds roughly 70 to 100 Elo in self play depending on the testing conditions. The majority of this increase in strength can be attributed to an improved network generated via self play training using the previous "retrograde learning" network used in versions 2.0.0 and 2.1.0 as a base. The specific self play code used can be found under the selfplay branch of the "seer-training" repository. The remaining 20 or so Elo originates from search patches.

ELO   | 101.47 +- 3.54 (95%)
CONF  | 40.0+0.40s Threads=1 Hash=64MB
GAMES | N: 20008 W: 8210 L: 2527 D: 9271

As with the previous release, thanks to everyone involved with the OpenBench project for making Seer's continued development feasible.

seer-v2.1.0

This release adds +70 to +100 elo in self play over v2.0.1 depending on the configuration. The entirety of this strength increase can be attributed to search improvements as this release features the same network bundled with v2.0.1 trained via retrograde learning. Many thanks to everyone involved with the OpenBench project for making Seer's continued development feasible. In particular, thanks to Andrew Grant for developing and maintaining OpenBench and Bojun Guo (noobpwnftw), the primary OpenBench contributor.

ELO   | 104.23 +- 3.58 (95%)
CONF  | 8.0+0.08s Threads=1 Hash=32MB
GAMES | N: 20000 W: 8373 L: 2547 D: 9080

Note that as of v2.1.0, I'm only officially supporting AVX/AVX2 processors and requesting that testers please refrain from testing binaries for older processors found in the "unsupported.zip" attached below. For casual use as well as unofficial rating lists, feel free to use the unsupported binaries if none of the supported binaries prove compatible.

seer-v2.0.1

This release resolves time management issues when playing using cyclical TCs associated with v2.0.0. The cyclical time management has also been adjusted to be less aggressive which, in limited testing, gained a modest amount of elo given cyclical TC conditions.

As always, Seer performs optimally on modern x86 processors with AVX extensions, but it should still be reasonably strong on other hardware. Only binaries for Windows are supplied at this time. Linux users should not experience much difficulty in and can expect superior results by compiling locally (see README for instrutions).

For Skylake generation and newer Intel processors, the seer_skylake.exe binary is recommended. For Ivybridge generation through Skylake generation exclusive Intel processors, the seer_ivybridge.exe binary should yield best performance. For older Intel systems, a seer_core2.exe binary is supplied. For AMD Ryzen 1000 and 2000 series processors, the seer_znver1.exe binary should be optimal and for 3000 series and newer AMD processors, the seer_znver2.exe binary should yield best results.

seer-v2.0.0

This release adds around +120 elo in self play over v1.2.1. Limited testing indicates that slightly more elo can be expected against other opponents. Additionally, the reported NPS should be slightly increased relative to v1.2.1.

The embedded network is trained entirely on data generated by Seer's search starting with a randomly initialized network. The network is no longer directly or indirectly trained on Stockfish derived training data. The training technique employed is a variant of semisupervised learning and involves starting with a large number of <=6 man positions labeled using Syzygy EGTBs. A lengthier description of the unique training process can be found in the README.