SimpleChess 3.2 Public Release - CCRL 2+1 Rating Estimates

[mysammoore]

SimpleChess 3.2.0 is out at https://github.com/LinkLikesLattes/SimpleChess under GPL-3.0, and since there's no CCRL rating for the engine to point at I measured one against CCRL's own, 464 games against 29 CCRL-rated engines from Stockfish 19 down to c4ke 3.0, each held at its published 1-CPU Blitz rating while 3.2's was fitted with BayesElo. That puts 3.2 at roughly 3529 on the CCRL Blitz 2'+1" scale, somewhere in 3501 to 3557, and it's the figure I'd use as a starting point for pairings. It's a source release with no binaries as always, so anyone willing to compile it and run it under real CCRL conditions would be doing me a real favour, and that's mostly why I'm posting.

As for what changed, the playing network is a fresh selfplay generation. The adaptive search width no longer collapses to nothing in the endgame, the material term that had driven it toward zero as pieces came off is gone. The correction history now weights each update by evaluation tension, the difference between the static eval and the same network queried with the side to move swapped, so at depth six and up a higher-tension position takes a proportionally larger correction step. It plays Chess960, FRC and DFRC under UCI_Chess960, castling is sent and received king-to-rook and the parser accepts both X-FEN and Shredder castling fields. The Move Overhead option now feeds the time budget instead of being parsed and then discarded for a hard-coded reserve. And a movegen bug is fixed, the vendored library was counting two diagonal sliders as a single check and could have generated an illegal interposition.

[Roland Chastain]

Congratulations for that release.

Don't take this the wrong way, but I can't help pointing out that "Simple Chess" is a rather poor choice of name — firstly, because there are already a thousand other programs with that name, and secondly, because your program isn't "simple" at all.

[mysammoore]

Poor choice of name? Point me to a "SimpleChess" anyone has ever used or rated. And of course there's nothing simple about it. The original concept was an HCE that relied solely on "simple chess principles" hence the name, but since then, it has transitioned to NNUE from the original HCE labels. The name is reference to the origin of the evaluator. It's also catchy. I have not found reference to any other engine under that name (please direct me to the thousands of SimpleChess programs). The closest relatives are the long-dead "Simplex" and "SSSimpler" on the CCRL rating list. Stockfish is not literally a fish, nor does "stock" reference anything (the literal word "stockfish" is a very stupid name for an engine). PlentyChess is as stupid a name as SimpleChess. I'm not trying to be combative, but this comment is not really necessary because chess engines famously have extremely nondescript or stupidly simple names.

[mysammoore]

Now I read my own reply thinking I come across as an egomaniac. I can't seem to help myself. I defend relentlessly even when the "attack" in question is a footnote. The truth is I have an obsession with picking fights and I want to stop doing that. I'm sorry, Roland.

[Roland Chastain]

I'm sorry, Roland.

No need to be sorry I think. I must admit that my comment was a bit bold (especially since we don't know each other) and exaggerated.

Your response makes sense.

If I ask my search engine to find "simplechess chess engine", your project appears in third place.

Good luck with the project.