This is the first release with executables made using PyInstaller. Running the engine is now easier than ever - no need to set up a full python dev environment.
Last time I tried this, I bundled heavy things like torch, cuda, etc and the package ended up being 6GB+. I never released those versions. This time, I've built relatively lean executable (~30-40 MB) by using onnx instead of torch (cpu-only onnx included, no cuda). You'll also need to download the new small neural network (~10 MB) - the smallest ever released for the Dinora project.
Strength estimation
In previous releases, I tested rating with elofish, which limits Stockfish via UCI_Elo. That turned out to be (unexpectedly) overly optimistic. Why?
Only one opponent (Stockfish) is tested. It uses constant time per move (ignores time management). No opening book is used.
This time, I aimed for more reliable estimates, I downloaded 17 FOSS engines from the CCRL Blitz list, set up a gauntlet tournament using the "Perfect 2021" opening book,
and ran games with a time control of 100 seconds + 0.9 second increment (close to CCRL 2 minutes + 1 second)
After 68 games, the results placed Dinora v0.3.0 at 2216 Elo—a more reasonable estimate. By contrast, elofish overestimated by just +500 Elo.
Dinora 0.3.0 chess engine - position test
The discussed version of the Dinora 0.3.0 engine was given a tactical position to solve, as shown in the diagram. The move 1.Rg4 wins!!
However, the Dinora 0.3.0 engine failed to find a winning move for 10 minutes. Instead, it proposed the draw move 1.Bd5.
How strong is the new version of the Dinora 0.3.0 chess engine?
The author states that Dinora 0.3.0 in his tests reached 2216 Elo. We have tested this engine yet, our team will start tournaments with this version from tomorrow. We invite you to follow the results on our website.
Comments
Post a Comment