Skip to content

1.3.2

Latest

Choose a tag to compare

@looak looak released this 25 Oct 23:38
· 87 commits to main since this release

v1.3.2

The first release since 1.0.0, three evaluation cycles later. 1.0 made Elephant correct and multi-threaded; 1.1 to 1.3
rebuild its static evaluation. PeSTO's material and piece-square tables are joined by mobility, pawn structure, passed
pawns and king safety, each term fitted with Elephant's own tuner and kept only after it passed an SPRT.

A gauntlet of 1.2.0 against 13 engines rated 2663–3095 put it at 2894 ± 16 CCRL Blitz (1300 games at 120+1, a
fixed-anchor fit, statistical error only). 1.0.0 was about 2740 on a comparable field. 1.3 adds king safety on top.

Evaluation

  • Tapered evaluation, tuned. Every term is a midgame & endgame pair, summed and blended once by game phase.
    ElephantTuner (in tuner/) fits them to millions of quiet positions with their game results (Texel tuning,
    gradient descent with Adam). New terms are fitted with everything else held fixed.
  • Mobility (1.1): safe squares for knights, bishops, rooks and queens; squares attacked by enemy pawns don't count.
  • Pawns (1.2): passed, isolated, doubled, backward and connected pawns; passed pawn extras for king distances, a
    blocked or free path and a rook behind; the pawn piece-square table retuned with them. Pawn structure is cached
    per thread in a pawn hash table.
  • King safety (1.3): shelter and pawn storm on the king's file and the two beside it; king attack, the enemy
    pieces bearing on the king zone and files open to its rooks and queens, scored from a fitted danger table.
  • Tempo: 15 cp to the side to move.

UCI

  • buildinfo reports the commit, branch, dirty state and build settings.
  • go ends with an info nodes line counting every node searched.
  • bestmove 0000 when there's no legal move.
  • A FEN without exactly one king per side is rejected.

Under the hood

  • 1.3.1: dead code removed (the old Move class and unused evaluation constants), logging only where something
    logs, compile time roughly halved. Same search, node for node.
  • 1.3.2: perft counts the last ply in bulk, about 530M moves/s on one thread (was about 140M). Same search, node for
    node.

OpenBench

Each change against the one before it, 1 thread. Short time control first, then a long one to confirm.

Change Short TC Long TC
Mobility (1.1) +99.2 ± 20.6 (40+0, 608 games)
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)
Tempo (1.3) +2.3 ± 6.3 (8+0, 5998), stopped undecided, merged
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, 7742), stopped undecided +11.1 ± 6.6 (40+0, 5464)

Bench 520775 nodes.

Downloads

elephant-gambit-1.3.2-<os>-<build>: x86-64 runs on any 64-bit CPU, x86-64-sse42-popcnt needs SSE4.2 & POPCNT,
x86-64-avx2-bmi2 needs AVX2 & BMI2 (most CPUs from about 2015 on). Pick the newest your CPU supports.