2025-01-05: Texel 1.12 (ELO: +85)
Neural network improvements:
- Speed up matrix multiplication for AVX-512.
- New network based on new training data.
- Use 4 "heads" (all NN layers except the first) for evaluation. The active head depends on the number of pieces on the board.
- Increase NN layer 1 output size from 256 to 384.
Tablebases:
- Measure thinking time more accurately. Useful when TB files are on mechanical disks.
- Better handling of DTZ scores in search.
- Avoid expensive (wrong side to move) DTZ probes in search.
- Add UCI parameter to control required search depth for DTZ probes.
- Measure time required to perform a TB probe instead of guessing. This avoids time losses in fast games when TB files are on mechanical disks.
- Allow RTB WDL probes in some cases even if hmc > 0.
- Change TB swindle depth from 16 to "15+minProbeDepth", to avoid slowdown in late middlegame/early endgame when TB files are on mechanical disks.
Search:
- Implement the counter move heuristic.
- Implement multi-cut pruning based on singular search score.
- Adjust singular extension search parameters.
- Allow both extension and reduction in the same search node.
- Remove no longer useful "recapture" and "going into pawn endgame" extensions.
- Reduce aspiration window size.
Other:
- Use multiple threads to initialize the transposition table.
- Better handling of mate scores in the transposition table.
- Fix handling of empty strings in UCI string options.