I've forked Ethereal to Etherlito, which will not be seeing any support or further commits. This allows you to place a weight file in the /src/ directory named "weights.nn", and compile. To make people happy, I've pasted a Stockfish net into the repo. One can compile out of the box, probably. You maybe have to play with the Makefile if you don't have a modern CPU.
Using various Networks, I am able to beat Stockfish 11 in a 500 game match by over +50 elo. Chess is dead. I'm not providing binaries for this, as Etherlito is the only code I've written in the last 6 years that I am ashamed of. I want everyone to know just how easy this was. 2 hours, 25 minutes. That is the time it took me to start by opening a tab to look at CFish, and to end by pushing a working version. Congrats to Stockfish for the incredible achievement. My middle finger to everyone else using it.
Binaries for Windows & Android (ARMv7, ARMv8) can be found here. As always, it is highly preferable for you to compile your own binary on your machine. A simple "make", or "make pext" (for those with Intel machines supporting the BMI2 instruction set), should produce a binary as fast and likely faster than the ones I provide. Of course, test it to be sure.
The following are the patches since V12.50.
Code: Select all
Evaluate Stacked & Isolated Pawns by File
Tune all NORMAL terms using mini-batches
Reduce the base Aspiration Window size to 10cp
Fix MoveOverhead to help with Sudden Death games
Add a Capture History Table (Stockfish) to move sorting
Tune all NORMAL positions using a new set of 42.5m games
Increase the importance of Move Stability in the Time Management
Add a small, [224, 32, 1] NN to augment the eval via Pawns and Kings
Fix an off-by-one error causing crashes with the NN weights
Introduce a miniature Shelter/Storm for use in King Safety
Update the Network, now trained using a mix of FRC data
Tune all of the NORMAL evaluation terms following the NN addition
Use Capture History (Stockfish) to apply LMR to some tactical moves
Increase reductions for tactical moves with very poor history
Add mirroring to the training, tune the eval around the updated NN
Remove the "UE" portion of the NN, as we can compute it faster from scratch
Remove a useless line when reverting a NULL move
Remove a useless operation when generating Quiet Evasions
Augment the existing eval with a RPK Endgame Network
Kick to version 12.70
Kick to version 12.71
Kick to version 12.72
Kick to version 12.73
Kick to version 12.74
Kick to version 12.75
Code: Select all
ELO | 42.90 +- 2.90 (95%)
CONF | 60.0+0.6s Threads=1 Hash=64MB
GAMES | N: 18640 W: 4334 L: 2044 D: 12262
Thanks to all of the old guard on these forums. I learned almost everything I know about chess, computer chess, and even programming in the pursuit of Ethereal. I got damn near the top. If you toss Houdini and Fire, I held the #3 spot in the world for a year. A testament to my obsession, and to the support of this community.
I'm not sure what to say as a closer. I can't find words to thank the many members of CCC for their contributions over the decades. I've been apart of the community for 6 years, but I don't see a future in this hobby. By this time next year, I expect to see Stockfish, Leela, and a bunch of forks or clones or near copy-paste jobs of them. Good luck. A republic, if you can keep it.
A wise man, one of the few true friends I've ever had, told me this: "fais ce que dois, advienne que pourra"