Download

• DanaSah v.9.1 for Windows 32-64-bit and Linux 64 bit (processors avx2, moderm-sse41, sse2 and basic) 19-10-2023 danasah91.zip

Changes:

1. 1. 1. Updated NNUE to dananet3.nnue. The network was created with about 600 million positions generated with the previous network and also evaluated in depth 9. It looks slightly stronger, but only about 15 Elo points more.

      2. Fixed a bug that when nnue was not used in the evaluation it showed a wrong evaluation.

      3. There is a Linux version for the first time for these NNUE versions.

      4. There is a limited strength version. The limited-strength version does not use NNUE in the evaluation, so although there are some changes from the last released limited-strength versions, these changes are not very big.

• DanaSah v.9.0 danasah_90.zip and v.9.0 LimitStrength danasah_limitstrength_90.zip for Android 32 y 64 19-10-2023 

Archimedeschess has compiled DanaSah v.9.0 for Android, you can find arm7 (32bit) and arm8 (64bit) versions for both normal and limited strength version.

Archimedeschess website --> https://sourceforge.net/projects/chess-engines-for-android/

Since Android version 10, the app and in this case the Chess GUI has problems to install third party chess engines with xboard or uci protocol, the new compatible GUI's only support oex compatible engines.

DanaSah is uci compatible and at the moment it is not oex compatible, so to use it on android 10 or higher you need to follow a series of steps. You need to install a version of Droidfish from outside the store, for example from F-droid, another possibility is to use an older version of Chess for Android, recommended version 6.2.1. To use NNUE you have to indicate the full path to the dananet2.nnue file, for example /storage/emulated/0/danasah9/dananet2.nnue. I have also tried to make gaviota tablebases work, you also have to specify the full path, e.g. /storage/emulated/0/gtb4/. Use internal memory, not sdcard.

DanaSah v.9.0 has a bug in evaluation when not using NNUE, so it shouldn't be used like that (this is fixed in version 9.1). I am not sure if this also affects the strength-limited version. When analysing for example the start position it should give a value close to 0 and not something like 1.5.

• DanaSah v.8.8 and v.9.0 for Windows 64-bit (proccesors avx2, sse41, sse2)  31-12-2021 danasah88-90.zip

Changed:

1. The engine can read NNUE-type neural networks. This gives a lot of strength.

2. The search has many changed parameters, but the strength with the classic evaluation is the same as the previous version.

3. There are changes to the limited strength version (although it is not included here).

4. Other minor changes.

dananet1.nnue --> This is the default network used by 8.8 version, it has been trained with 340 million positions with ccrl 40/2 games (learn by result) and then with 340 millon positions in depth 5 generated with the first net.

dananet2.nnue --> This is the default network used by 9.0 version, it has been trained with 340 millon positions in depth 9 generated with dananet1.

DanaSah 9.0 64-bit Elo CCRL 40/15 --> 3190

DanaSah 8.8 64-bit Elo CCRL 40/15 --> 3143