seer-v1.2.1

This release should resolve an issue reported with v1.2 where the engine would fail to play the shortest path to mate. It also resolves a minor race condition related to the printing of the best move.

seer-v1.2

This release should resolve some of the previous release's UCI compatibility issues. This release, in self play, adds around +250 elo, but early results seem to indicate significantly less improvement against other engines. About 15-30 elo originates from improvements to the search. The remaining elo gain is accounted for by more careful training and testing of the network.

With this release, the engine will refuse to run if not supplied with a network and will print an error such as the following prior to exiting when the user issues the go command:

error: weight signature mismatch
got f21c3cd7, expected c2355852

As with the previous release, the network architecture and training code are all original. However, the network is still trained on Stockfish self play data. With any luck, this will be the final release for which this is the case as my tabula rasa self play training seems to be in a tentatively working state.

seer-v1.1

This release should perform more than 100 elo stronger than v1.0. Most of the gains originate from some optimizations to the network architecture. As a consequence of these optimizations, the new networks are both much faster to evaluate and more accurate. A nps increase of 50% to 100% should be expected. Move generation now relies on plain magics as opposed to BMI2. This change lead to a minor performance improvement on systems lacking BMI2. Additionally, this release corrects an issue in v1.0 where an additional thread would be used.

To run the engine, the "save.bin" weights file found below must be downloaded and the engine pointed to the downloaded weights file by setting the "Weights" UCI parameter to the absolute path to the downloaded weights file.

seer-v1.0

Intitial release. Estimated playing strength is around 2900+ elo. However, this estimate is highly dependent upon the availability of vector instructions. For the engine to function, you must download the attached save.bin file. The engine must then be pointed to the downloaded save.bin file, containing the model's weights, by setting the "Weights" UCI parameter to the absolute path to the downloaded save.bin file. Seer only supports 64-bit processors. At this point in time, only Windows binaries are supplied. However, Linux users should experience minimal difficulty in following the compilation instructions found in the readme (it is highly recommended to compile with g++ over clang).

seer-v0.1-alpha

• This initial release should be around 2800 elo on the CCRL scale. I'm only providing Linux binaries for this pre-release, though compiling on Windows should be straight forward. Time management is still the super rudimentary "budget 1/7 remaining time" strategy which results in difficulties with certain time controls. At STC, it is expected to underperform due to the heavy evaluation function and, at LTC, overperform.
• Provided are binaries supporting both systems with AVX2+PEXT (seer-modern_x86_64) as well as systems lacking AVX and PEXT (seer-old_x86_64). Additionally, a weights file (nn.bin) is supplied which the engine must be provided an absolute path to.
• On old systems, 200-250 knps should be expected. On modern systems, 350-400 knps should be expected