Elephant Gambit 1.3.2 "Matriarch"

Discussion of anything and everything relating to chess playing software and machines.

Moderator: Ras

looa
Posts: 8
Joined: Sat Apr 15, 2023 9:17 pm
Full name: Alexander Ek

Elephant Gambit 1.3.2 "Matriarch"

Post by looa »

Hi,

A few weeks ago I pushed the release button on my engine with 0.15.2, and soon after that 1.0.0. Today I present 1.3.2 "Matriarch". My versioning is all over the place: it makes sense to me, but it follows how I feel about the engine more than any rule about what makes a release.

1.0 was about making Elephant correct and multi-threaded. Since then I've spent three cycles rebuilding its static evaluation, and Matriarch is where that work lands. PeSTO's tables are still the base, on top of them are mobility, pawn structure, passed pawns, mobility and king safety, each one fitted with Elephant's own tuner and kept only if it passed an SPRT.

Download: github.com/looak/elephant/releases/tag/1.3.2 (C++20, GPL-3.0)

Strength
My most recent gauntlet of 1.2.0 against 13 engines from the CCRL Blitz list (rated 2663–3095, 1300 games at 2'+1") put it at about 2894 ± 16 CCRL Blitz (my own test, statistical error only). 1.0.0 was about 2740 on a comparable field. 1.3.2 should be slightly stronger.

What's new since 1.0
  • Tapered, tuned evaluation: every term is a midgame/endgame pair, fitted to millions of quiet positions with their game results (Texel tuning, Adam). Using public position databases. The plan is to replace this with Elephants' own games.
  • Mobility (1.1): safe squares for knights, bishops, rooks and queens.
  • Pawns (1.2): passed, isolated, doubled, backward and connected pawns, and a pawn hash table.
  • King safety (1.3): shelter and pawn storm, plus king attack scored from a fitted danger table.
  • UCI: buildinfo, an info nodes line counting every node, bestmove 0000 with no legal move, stricter FEN checks.
  • 1.3.2: perft counts the last ply in bulk, about 530M moves/s on one thread on my Ryzen 9 9900x. Same search, node for node.
SPRTs (each change against the one before it, 1 thread, OpenBench)

Code: Select all

Change                     Short TC                        Long TC
Mobility (1.1)             -                               +99.2 ± 20.6 (40+0, 608)
Pawn structure (1.2)       +52.3 ± 15.5 (8+0, 1224)        +70.2 ± 18.3 (40+0, 848)
Passed pawn extras (1.2)   +71.4 ± 18.1 (8+0, 824)         +53.2 ± 15.5 (40+0, 1020)
Pawn table rebalance (1.2) +18.0 ±  8.9 (8+0, 3318)        +27.0 ± 10.7 (40+0, 2090)
King shelter & storm (1.3) +20.0 ±  9.4 (8+0, 2910)        +30.5 ± 11.1 (40+0, 1668)
King attack (1.3)          +3.9 ± 5.8 (8+0), undecided     +11.1 ±  6.6 (40+0, 5464)
Builds
Windows and Linux, at three levels: x86-64 (any 64-bit CPU), sse42-popcnt, and avx2-bmi2 (most CPUs from about 2015). Pick the newest your CPU supports. Bench: 520775 nodes.

What's next
Tuning the search parameters with SPSA, and later a network trained on Elephant's own games.

Thanks to Ronald for PeSTO, Andrew for OpenBench, the fastchess and cutechess authors, Peter for the Texel tuning method, the author of the E12.52 dataset, the Chess Programming Wiki and the Engine Programming Discord, the CCRL for the rating lists I test against, the authors of the engines in my gauntlets, Graham for his endless testing, and last but not least, talkchess.com. Feedback and games are very welcome!

Alex