Cadie 1.5

[janus]

Cadie 1.5
@mwyngaarden mwyngaarden released this 2 days ago

This release includes many bug fixes, improvements, and search refinements. In self-play, Cadie 1.5 is ~70 Elo stronger than Cadie 1.4 at the 120"+1" time control. Hopefully, this will translate to at least 45 Elo in the rating lists.

Most of the code changes are refactoring. Elo gains come mainly from improved move ordering, the addition of counter moves, a new transposition table replacement scheme, and improved time management. There are also several refinements in the main search, and quiescence search, which make the style of play more tactical.

This is the first release to include binaries for Windows. These have only been tested on Windows 10.

Cadie_1.5D.exe: Dynamic linkage
Cadie_1.5T.exe: Static linkage

Both should be equally performant.

Binary >> https://github.com/mwyngaarden/Cadie/re ... e_1.5D.exe

Binary >> https://github.com/mwyngaarden/Cadie/re ... e_1.5T.exe

Source >> https://github.com/mwyngaarden/Cadie/ar ... s/v1.5.zip

Source >> https://github.com/mwyngaarden/Cadie/ar ... 1.5.tar.gz

[janus]

Caissa v1.10
@Witek902 Witek902 released this 2 days ago

Summary
This release introduces new neural network architecture. The accumulator is split into whites and blacks perspective. Trained on 1.9B positions. The biggest gains are visible in DFRC. Accumulator size is now 1024 neurons, so the net file is even smaller than in previous release (1536).
Also includes a small fix in search (not updating best move if searched move doesn't beat alpha), worth around 15 Elo.

More here >> https://github.com/Witek902/Caissa/releases/tag/1.10

Binary >> https://github.com/Witek902/Caissa/rele ... a_1.10.zip

Source >> https://github.com/Witek902/Caissa/arch ... s/1.10.zip

Source >> https://github.com/Witek902/Caissa/arch ... .10.tar.gz