Cheng 4.48

• 9-bit fixed point (16-bit weights and biases), faster inference, new bigger NN, 4 billion selfplay, new topo: 736-576*2-1
• more tactical search at the cost of some elo loss:
  • dropped recapture extensions
  • reenabled fractional plies
  • relaxed singular margin
  • more check extensions in scout nodes
  • SEE pruning
• bench 10044316

Cheng 4.47

• new bigger NN, 3 billion selfplay positions, 10k nodes, new topo: 736-448*2-1
• bench 8501031

Cheng 4.46

• new NN, 800M selfplay positions, 10k nodes, new topo: 736-224*2-1
• augmented net eval with old HCE EG scalers/recognizers
• bench 9213631

Cheng 4.45

• new NN, 400M selfplay positions, 10k nodes
• syzygy cutoff at scout nodes (finds mates slower, but gives more accurate eval in won/lost positions)
• bench 10988267

Cheng 4.44

• NN switched from floats to fixed point - now eval/bench is deterministic again
• NN topo changed to 736-224-1 (performance)
• switched from leaky relu to plain relu (performance)
• NN cache incremental updates (performance)
• NN outputs scaled pawn scores instead of win prob (fixed oscillations at high evals) - thanks Jay
• avx2 binaries provided for x64 (the preferred ones to be tested, ~20% faster)
• minor bugfixes
• bench 10429956

Cheng 4.43

• NN based evaluation (topo 736-192-4-1), goodbye HCE
• no incremental inference updates, so a bit slower
• custom inference with layer 1 zero skip trick
• inverse sigmoid to convert win probability to cp, oscillates a lot with big evals
• trained on 1.7M cheng self-play fixed node games, yielding ~125M filtered positions
• labeled on 6-ply cheng search and game outcome
• includes custom hacky libtorch-based tuner
• bench 11152861
• dropped support for Windows XP

Cheng 4.42

a very modest improvement over 4.41 (12 elo maybe)
notable new feature: syzygy tablebase support

Cheng 4.41

Cheng 4.41 release
only about half the improvement of 4.40 over 4.39, estimating about +15 Elo against other engines

Cheng 4.40

official release, please see readme.txt for further details