Chess Engines for Android

Hans Obermeier wrote:

How to download?

Archimedes wrote:

As Outskirts Chess Forum is closed, i will temporary post some newer Android builds here. For that purpose, i choose a more common thread title.  

ChessFan1 wrote:

Archimedes wrote:

Hello Archimedes, you came here too!  Did you read my PM (on talkchess)?

Yes i did.

Archimedes wrote:

Hello Archimedes, you came here too!  Did you read my PM (on talkchess)?

I found a serious bug in Blue Marlin 14.4. Sometimes it could resign in winning position. It turned out that the problem was already in version 14.3. It is fixed in build 14.4a. Swordfish 14.4 does not need any corrections.

- Make exploration constant configurable
- Add openbench compatability
- Use previous playouts to set root policy
- Learn 50 move rule
- Add pawn shield feature
- Use average reward, not visits, to choose pv
- Add movelist command
- Output info each second
- Add Bishop pair to evaluation

Added history heuristics for a very small Elo gain and fixed a hash move bug for a much larger Elo gain (I think). I'd like to figure out why HH isn't doing much for me. Also discovered and fixed a scenario that can stall the engine.

Another release of Berserk with an improved NNUE evaluation and small search patches.

The new network was trained using a custom trainer on 950 million FENs from Berserk 6 self play games. Along with this, the input arch has changed such that the pieces are now indexed relative to the side of the board the king is on. This should hopefully improve the network's ability to detect king attacks.

Minor search patches were added as well, most of which were many late move reduction tweaks.

Almost "exactly" the same as v5.9. But a few cleanups here and there. No warnings/errors.

Fixed a variable having the wrong type. This caused the "extra" module failing to compile.

NB: For normal chess engine usage nothing has changed. A default compile does not include the "extra" module. Therefore this release doesn't include any new binaries. You can just keep using the binaries for Alpha 3.0.0.

Implemented SEE, hidden attackers work now but it seems to only have given an improvement of 50-100 Elo. With the improvements in move ordering and LMR I expected it to make a bigger difference. SEE is not used for quiet moves and does not account for promotions or en passant, so those might make it more effective. Timeouts also should not be occurring any more as another condition was added to time management.

Changes in Arasan 23.1:
1) Disable book when running the "bench" command. Ensure tablebases and NNUE are initialized before it runs.
2) Support SIMD on older Intel/AMD chips. Default builds now assume at least SSE2 is available.
3) Fix #34 - specify C++17 for clang.
4) Various fixes to selfplay utility
5) New network, based on reinforcement learning from 280m positions
6) Singular extension changes. Allow some pre-search pruning during a singular search.
7) Some code cleanup and refactoring.
8) Fixes and cleanup to Makefiles, VC++ projects and CMakeLists.txt.
9) Output error messages related to intialization to stdout not stderr, and use the appropriate prefix (e.g. "info" for UCI).

This release is not a notable strength improvement over 7.1.0. However, 7.2.0 does introduce a polyglot opening book loader as the primary new feature. Several UCI options are provided to make usage of the loader. Additionally, various places in the codebase have been refactored and cleaned-up, and the evaluation for Blunder has been retuned and restructured to allow for more granularity. Lastly, a makefile is now included so Blunder can be more easily compiled across platforms.