ChangeLog.txt

2024-01-12: Texel 1.11 (ELO: +63)
Neural network improvements:
- Fewer non-incremental updates of the first layer state.
- Implement quantization aware training.
- Better GPU utilization by using separate data loading threads.
- Efficiently handle training data that is too large to fit in RAM.
- Train a new network based on new training data evaluated by Texel 1.10.
- Optionally use AVX-512 instructions.
Other:
- Re-calibrate UCI_Elo for NNUE.
- Fix a null-move verification problem.
- Update GoogleTest code to version 1.14 to fix warnings when using new CMake
  versions.
- Add state to the Square class, i.e. make it contain the square number.

2023-09-25: Texel 1.10 (ELO: +173)
Speed up evaluation:
- More handwritten code using compiler intrinsics.
- Switch from gcc to clang for Windows compiles.
- Cache-line align data for more efficient SIMD load/store.
- Avoid "horizontal adds" in matrix multiplication by re-arranging weight data
  ahead of time.
Better neural network:
- Train a new network based on new training data evaluated by Texel 1.09.
- Use both search scores and game results when generating training data.
- Compute quantized validation error during training.
- Set unused layer 1 weights to 0 during training, to reduce compressed network
  size.
- Print weight/bias statistics during net quantization, to make it possible to
  see if quantization parameters are reasonable.
- Permute the quantized net for more efficient sparse handling of the outputs
  from the first layer.
Compilation fixes:
- Fix Visual Studio debug compilation.
- Fix cygwin/msys2 compilation problems.
- Better SIMD support for 32-bit CPUs.
Other:
- Add source code for "syncengine" program.
- Allow MultiPV mode also when playing games.
- Re-introduce evaluation score reduction when the half-move clock is large.
  No strength improvement but avoids shuffling moves in some situations.

2023-07-30: Texel 1.09 (ELO: +252)
- Implement NNUE evaluation.

2023-06-10: Texel 1.08
New features:
- Allow arbitrary hash table sizes up to 1TiB of memory.
- Add "contempt" UCI parameter.
- Add automatic contempt depending on opponent name.
- Add support for 7 man Syzygy tablebases.
- Add "runcmd" Windows program. Useful for GUIs that don't allow specifying
  command line arguments to the engine.
Bug fixes:
- Fix crash in Windows NUMA/CPU discovery code.
- Fixed rare crash when running in cluster mode.
- Changed default "ponder" UCI option value to false.
Time management:
- More dynamic time management.
- Better time allocation for ponder on games.
Search:
- Use depth-dependent singular extension threshold.
- Add tempo bonus.
Reduced strength:
- Add MaxNPS UCI parameter.
- Make reduced strength work for Threads > 1.
- Add UCI_LimitStrength and UCI_Elo support, range -625 - 2540.
Infrastructure:
- Use CMake instead of a handwritten Makefile.
- Use Google Test framework instead of CUTE.
texelutil:
- Use C++ threads instead of OpenMP.
- Build texelutil with reduced functionality even if Armadillo and GSL libraries
  are not found.
- Add a reverse move generator, "texelutil revmoves".
- Improve "texelutil proofgame" for random positions.
- Add "texelutil rndfen" to generate random positions.
Documentation:
- Document how to estimate number of legal chess positions.
- Document the Texel tablebase probing algorithm.

2017-09-30: Texel 1.07 (ELO: +35)
Search:
- More aggressive LMR in expected CUT nodes.
- Don't use history and killer heuristics for captures.
- Prevent q-search explosion in pathological cases.
- Allow LMR for checking moves with negative SEE value.
- Don't do internal iterative deepening when in check.
- Don't do razoring when in check.
- Don't check extend SEE<0 moves at depth<=3.
- More LMR/LMP reduction if the evaluation score is worse than it was two plies
  ago. Idea from stockfish.
- More aggressive reverse futility parameters.
- In hash replacement, prefer a deep bound over a shallow exact entry if the
  depth difference is larger than 3.
- Use history score to control LMR reductions.
Parallel search:
- Parallel search algorithm changed to hybrid lazy SMP / ABDADA.
- Removed code that on NUMA avoids probing the TT at depth 0 for nodes > 0.
Evaluation:
- Removed the piece trade bonus.
- Implemented fortress detection for KQvsKRM+pawns endgames.
- Added endgame correction for KQKNNNN which is generally won by the knights.
Speed:
- Support for large and huge page allocations for the transposition table.
- Added alpha/beta pruning to the SEE function.
- Added support for sliding move generation using the PEXT BMI2 machine code
  instruction.
Other:
- Include "hashfull" information in UCI output.
- Added UCI option to control whether the transposition table is aged when
  starting a new search in analysis mode.

2016-07-10: Texel 1.06 (ELO: +56)
Search:
- Don't use ply-2 killers for move ordering.
- Be more restrictive when extending losing (SEE<0) checks.
- Don't do check extensions for moves that are reduced by LMR.
- Don't do singular extensions if the unextended score is a mate score.
- Don't do LMR for the first 2 pseudo-legal moves.
- More aggressive LMR.
- More aggressive null move searches.
- Don't resolve a known loss score bound if there are other moves that are
  potentially not losing.
Parallel search:
- Allow a helper thread that fails high to interrupt the master thread.
- Changed MIN_SPLIT_DEPTH from 10 to 7 to make better use of multiple threads in
  short time control games.
- Changed maximum number of search threads to 512.
- Added support for windows NUMA systems having more than 64 cores.
Evaluation:
- Evaluation term to encourage "pawn breaks".
- Bonus for "latent" pawn attacks on enemy minor and major pieces.
- Bonus for safe queen contact checks.
- Only treat pawns on rank 2 and 3 as backwards pawns.
- King safety bonus for pieces (except bishops) on the same side of the board as
  the king.
- Parameter tuning.
Endgame:
- More patterns for fortress detection in bishop endings.
- On-demand tablebase generation for 4-men pawn-less positions.
- Knowledge about KRKBNN endgame, which is generally a draw.
- Better evaluation of KRBNKRB endgame.
- Larger evaluation score in KRRMKRR endgames
- Improved 50 move draw handling for TB probes.
Speed:
- Reduced TB probe overhead.
- Implemented a small (512KB) per-thread evaluation hash table.
- Removed support for fractional depth.
- Made the transposition table 4-way set associative.
Other:
- Removed "64-bit" from the engine name. 64 bit is now considered standard.
- UCI option to disable null moves.
- Made "BufferTime" a UCI parameter to control time handling safety margin.
- Don't start a new search iteration if it is unlikely that it can finish before
  the allocated thinking time is up.
- In analysis mode, don't age transposition table entries when starting a new
  search.

2015-01-24: Texel 1.05 (ELO: +57)
- Added support for Gaviota and Syzygy tablebases.
- Improved the built in opening book.
- Implemented UCI option "Clear Hash".
- Optimization for NUMA hardware.
- Earlier transposition table prefetch.
- Added evaluation of "semi-trapped" rooks.
- Removed the guarded passed pawn evaluation.
- Don't rely on the transposition table to communicate search results from helper to master thread.
- Use more time early in sudden death and increment games.
- Smaller delta in aspiration window loop.
- Optimized history heuristic update formula.
- Improved fortress detection for bishop and pawns vs pawns.
- Detection of "16 pawn fortress positions".
- Added bonus for pieces protected by pawns.
- Fixed incorrect draw scores in multipv mode.
- Made piece square tables have left/right asymmetry.
- Detection of KQKRP fortress positions.
- Disabled LMP in PV nodes.
- Only do check extension for negative SEE moves if the remaining depth is small.
- Implemented singular extensions.
- More aggressive LMR.
- Fixed makefile to make multi-threaded search work for Android.

2014-05-29: Texel 1.04 (ELO: +72)
- Added separate queen piece square tables for middle game and endgame.
- Added recognition of some drawn endgames involving bishops and pawns.
- Added evaluation corrections for QvsRR, QvsRM, RvsM, RvsMM.
- Made bishop pair bonus depend on number of opponent minor pieces.
- Added more evaluation terms for pawns (own and opponent) in front of the king.
- Added evaluation terms for king assisting passed pawns.
- Made king endgame piece square table contain 32 parameters.
- Made the isolated and doubled pawn penalties file dependent.
- Added file dependent passed pawn bonus.
- Only use the protected passed pawn bonus when the non-pawn material is balanced.
- Improved knight mobility evaluation.
- Added bonus for a piece blocking an opponent passed pawn.
- Added non-linear king safety and tuned other king safety parameters.
- Improved KRKP evaluation.
- Added evaluation of connected passed pawns.
- Added drawishness factor to rook endgame evaluation, based on white/black pawn asymmetry.
- Added endgame corrections for KRPKM.
- Added square-dependent bonus for pawns protecting/attacking other pawns.
- Added bonus for minor pieces protecting the king.
- Added penalty for semi-backward pawns.
- Reduce evaluation score when the halfmove clock is large.
- Added bonus for rook behind passed pawn.
- Optimized razor margins and delta pruning margin using CLOP.
- Improved the built in opening book.

2014-01-19: Texel 1.03 (ELO: +112)
- Multi-threaded (SMP) search.
- Remove moves before the last zeroing move from the repetition history list before starting search.
- Removed bogus transposition table store in Search::iterativeDeepening().
- Disable aspiration window logic when searching for a faster mate.
- Added null move verification search.
- Prefetch entries from the main transposition table.
- Bonus for candidate passed pawns.
- Skip move ordering after enough moves to trigger LMP have been searched.
- Implemented pawn storm bonus for opposite castling positions.
- Removed the Botvinnik-Markoff extension.
- More aggressive LMR.
- Added knowledge for KBpKB, KNpKB, KBpKN and KNpK endgames.
- Store endgame flag in material hash table.
- Optimized lots of evaluation function parameters.
- More consistent evaluation of different classes of pawnless endgames. KQK, KRK, KQKM, KBBK, KBNK, KQKR, KQKMM, KMMKM (M = B or N).
- MultiPV-support.
- Mate distance pruning.
- Evaluation correction for 7 knights vs 3 queens.
- Display engine evaluation using "info string" when starting search in analysis mode.
- Changed internal parameter handling to make it easier to switch between UCI parameters and compile-time constants.

2013-06-08: Texel 1.02 (ELO: m=99 s=4.7)
- Fixed KNNpK evaluation bug.
- Fixed infinite loop when engine received EOF on standard input.
- Only do one check in q-search.
- Speeded up and improved the search tree log file analyzer.
- Remember history information between searches.
- Remember best move (for LMR threat detection) also when there is a
  transposition table hit or when going into q-search on the next ply.
- Cache line aligned the main transposition table and the pawn hash table.
- Use cache-optimized magic bitboard constants.
- Handle stalemate in kqkp endings with pawn on a2/h2 and king on a1/h1.
- Detect illegal moves before playing them. Makes code look nicer.
- Reduce aspiration window from 20cp to 15cp.
- Gradually widen aspiration window after fail high/low at root.
- Store stale-mate scores in transposition table.
- Don't do LMP if best move so far is negative mate score.
- Smaller razor margin at depth <= 1.
- Support for SSE4.2 and POPCOUNT.
- Added knight mobility evaluation.
- Added KRPKR end game evaluation based on table base statistics.
- Use different bishop piece square table in the end game.
- Some speed optimizations.
- Fixed castle bonus asymmetry.
- Use an incrementally updated material identifier to simplify end game
  classification.
- When ordering root moves, use accumulated node counts instead of just last ply
  counts.
- Added evaluation term for bad bishops.
- Rescale history counters before starting a search, so that they adapt faster
  to the new situation, but still remember information from the previous search.
- Use a material hash table with material score and interpolation factors for
  other evaluation terms.
- Added evaluation term for redundancy of major pieces.
- Made history aging symmetric with respect to success/failure.
- Implemented knight outpost evaluation.
- Optimized pawn structure, major piece redundancy and knight outpost weights
  using CLOP.
- Disabled late move pruning in late end games.
- Fixed crash if EOF is received on standard input while engine is thinking.
- Better time management in "ponder on" games with increment.
- When only one legal move is available, think slightly longer than 2 ply.

2012-03-24: Texel 1.01
- Ported from Java to C++. About twice as fast as a result.
- Made node counters 64-bit to avoid overflow in long searches.
- Fixed handling of UCI "searchmoves" command.
- Added evaluation term to avoid walking into wrong corner in KRKB
  endings.
- Implemented reverse futility pruning.
- Implemented late move pruning.
- Made strength setting between 0% and 10% even weaker. 0% now plays
  random legal moves and 10% corresponds to the old 0% setting.
- Replaced cuckoo hashing with a more cache-friendly alternative.
- Try harder not to lose on time in really short time control games.
- Implemented passed pawn race evaluation.
- Increased passed pawn bonus for pawns on rank 6 and 7.

2011-10-24: Version 1.12 (ELO: m=61 s=1.6)
- Don't use aspiration window when mate score has been found.
- Optimized the legal move generator. This has minimal effect on
  playing strength, but improves perft speed a lot.
- Added bishop back rank penalty.
- Don't LMR reduce captures more than one ply.
- Avoid array index out of bounds exception if the engine is asked to
  search a position where there are no legal moves.
- Optimized piece values using CLOP.
- Store razoring results in the transposition table.
- Use SEE sign and MVV/LVA to order moves in non-q-search nodes.
- Try equal captures before non-captures in non-q-search nodes.
- Update killer moves when a transposition table hit gives a beta
  cutoff.
- Implemented signSEE function that can be used when only the sign of
  SEE is needed.
- Ignore squares attacked by enemy pawns when computing mobility.
- Added evaluation of KRKP end games, based on statistics gathered
  from EGTB information.
- Skip null-move search if the static score is below beta.
- Update hash entry "generation" value after a transposition table
  hit.
- Optimized bishop pair bonus using CLOP.
- Optimized futility margins using CLOP.

2011-06-12: Version 1.11 (ELO: m=47 s=4.8)
- Do extended futility pruning also on depth 4.
- Implemented razoring.
- Added support for fractional ply extensions.
- Use MVV/LVA ordering (but SEE pruning) in the q-search.
- Some speed optimizations.
- Use IID also at non-PV nodes.
- Made aspiration window a bit smaller.
- More aggressive futility pruning.
- Increase beta in two steps when a root move fails high.
- More accurate evaluation of end games with insufficient mating
  material.

2011-04-23: Version 1.10 (ELO: m=13 s=2.9)
- More aggressive futility pruning.
- Increased passed pawn bonus a little.
- Added small penalty for non-developed bishops.
- Implemented adjustable playing strength.

2011-03-16: Version 1.09
- Speeded up the SEE function by using bitboard techniques.
- Use a lookup table to speed up castle bonus calculations.
- Implemented specialized move generator function for check evasions.
- Some speed optimizations.
- Implemented specialized make/unmake move functions used by SEE.
- Allow more time to resolve fail lows at the root.
- Bug fix. Reset en passant square when making a null move.
- Don't allow too much time to resolve a fail low if too close to the
  next time control.
- Stop searching immediately after a ponder hit if only one legal
  move.
- Store successful null-move prunings in the transposition table.
- Give extra bonus to passed pawns if enemy king is outside "the
  square".
- Only try two hash slots when inserting in the transposition table.
- Try harder to keep PV nodes in the transposition table.
- Use a hash table to speed up king safety evaluation.
- Adjusted the knight piece square tables.
- Added code to dump the search tree to a binary file, and code to
  browse and search the dump file.

2011-01-30: Version 1.08
- Modified the bishop piece square table.
- Added penalty for bishop trapped behind pawn on a2/a7/h2/h7.
- Added isolated pawn penalty.
- Fixed race condition in UCI mode with quick ponder misses.
- Added backward pawn penalty.
- Modified the history heuristic calculations.

2011-01-18: Version 1.07
- Speed optimizations.
- Only apply the "drawish factor" for opposite colored bishops if the
  non-pawn material is balanced.
- Added evaluation of KQKP end games.
- Added evaluation term for two rooks on the 7th rank.
- Implemented delta pruning in the q-search.
- Improved reduction logic. Don't allow reduction of hash moves and
  killer moves, but allow larger reductions for other moves.
- Fixed bug related to stalemate traps and the 50-move rule.
- Allow null moves and LMR also in PV nodes.
- Don't do futility pruning and LMR for passed pawn pushes to the 6th
  and 7th ranks.
- Added bonus in end games for having king close to opponents passed
  pawn promotions squares.
- Don't allow null moves when side to move doesn't have pawns. This
  fixes problems in KQKR end games.
- Extend the search one ply when going into a pawn end game.
- Reduced queen mobility bonus.
- Increased the passed pawn bonus and made it non-linear.
- King safety adjustment.
- Added a threat bonus evaluation term.
- Penalize unmoved d/e pawns.
- Undo LMR if the threat move slides through the LMR move "from
  square". Idea from Stockfish.
- Ponder even if there is only one valid move.
- Allow LMR also at the root.

2010-12-31: Version 1.06
- Speed optimizations.
- Implemented non-linear mobility scores.
- Fixed bad interaction of futility pruning/hashing/fail soft.
- Tweaked rook piece square table a little.
- Order root moves according to search tree size on previous ply.
- Don't trust mate scores from null move searches.
- Increased the bishop pair bonus.
- Tweaked king safety a bit.
- Removed mate threat extension.
- Only extend recaptures if the recapture is reasonable according to
  SEE.
- Use a built-in bitbase for KPK evaluation.
- Reduced the aspiration window to 25 centipawns.
- Prune checking moves with negative SEE score in q-search.
- Allow reductions and pruning of losing (according to SEE) captures.
- Added end game evaluation of rook pawn and wrong-colored bishop.
- Added penalty for rook trapped behind non-castled king.
- Don't allow double null moves.
- Allow null move search to go directly into q-search.
- More depth reduction in IID at large depths.
- Implemented null move threat detection, to be able to undo unsafe
  LMR reductions. Idea from Stockfish.
- Don't overwrite a transposition table hash entry with inferior
  information for the same position. Could happen in IID.
- Added evaluation of unstoppable passed pawns in pawn end games.
- Include attacked squares close to the king in king safety
  evaluation.
- Implemented magic bitboard move generator.
- Use bitboard techniques for evaluation.
- Created specialized move generator functions to generate checks.
- Correctly handle pseudo-legal but not legal en passant squares.

2010-07-04: Version 1.05
- Speed optimizations.
- Avoid dynamic memory allocations during search.
- Don't do mate threat extensions too deep down in the search tree.
  Fixes search tree explosion in WAC 21.
- Incremental update of material scores.
- Created specialized move generator functions to generate captures.
- Created a function givesCheck that decides if making a move gives
  check to the opponent.
- Implemented a pawn hash table.
- Store static evaluation scores in the transposition table.
- If there is only one legal move at the root, make that move
  immediately.

2010-06-12: Version 1.04