Skip to content

Latest commit

 

History

History
245 lines (189 loc) · 9.56 KB

CHANGELOG.md

File metadata and controls

245 lines (189 loc) · 9.56 KB
Raw

Changelog

Unreleased

  • Evaluate passed pawns (38.30 +- 13.27)
  • Don't consider mobility for squares that are attacked by opponent pawns (14.90 +- 7.93)

Misc

  • Determine the evaluation coefficients in the main eval module, removing the need for a side-by-side impl in the tuner

5.0

  • Tune all evaluation parameters with https://github.com/GediminasMasaitis/texel-tuner (53.17 +- 16.76)
  • Evaluate piece mobility (41.36 +- 13.82)
  • Add a texel tuner in-repo and tune, resolving an issue where mobility scores were not computed correctly (28.53 +- 11.44)
  • Evaluate king safety (24.27 +- 10.43)

Misc

  • Define PSTs directly using phased evaluations to prep for tuning
  • Move all tunable parameters to a single file to prep for tuning
  • Add the 'default' feature to enable functionality for local from-source builds without polluting release builds
  • Add a CLI to the non-release build
  • Hide fields from PersistentState inside SearchContext

4.1

  • Add tablebase probing in search based on 'fathom' (18.97 +- 8.35 (5-man))
  • Use arrayvec for PrincipalVariation (15.52 +- 8.72)
  • Fix an accidental fail hard in quiescence - fail soft when eval >= beta (12.54 +- 7.15)
  • Add futility pruning (10.82 +- 6.41)
  • Pack Move into u16 and store extra information, e.g. if the move is a capture (7.03 +- 4.89)
  • Add tablebase support and follow tablebase lines (6.79 +- 4.59 (5-man vs none))
  • Add an evaluation bonus for having a pair of bishops (4.35 +- 3.46)
  • Add countermove history (2.66 +- 2.03)
  • Use separate functions for scoring tacticals vs. quiets (2.64 +- 4.43)
  • Enable LTO (2.09 +- 3.32)
  • Pre-compute move flags (2.07 +- 4.01)
  • LMR: Reduce in check, but reduce less than normal (1.51 +- 2.70)
  • Use arrayvec for MoveList (1.05 +- 3.35)

Misc

  • Allow sending only 'go wtime' or 'go btime'
  • Avoid constructing Move objects directly where possible, preferring to extract them from MoveList
  • Hide the internal implementation details of Move
  • Add a ByPlayer struct for easily working with values stored for both players

4.0

  • Add late move reductions (100.80 +- 22.10 Elo)
  • Use SEE to order bad captures later in moves to try (6.69 +- 4.80 Elo)
  • Do not allow TT cutoffs in PV nodes (~0 Elo)

Misc

  • Avoid locking up the UCI thread if the 'Hash' option is set during a search
  • Add a new Github Action for building and publishing new releases

3.0

  • Use hard and soft time limits in our time management strategy (~28 Elo STC, ~43 Elo LTC)
  • Store the board state as [PieceOccupancy; Pieces] and [ColorOccupancy; Colors] (~24 Elo)
  • Collect the principal variation during search (~18 Elo)
  • Use 3.3% of the remaining time as base rather than 5% (~9 Elo)
  • Skip losing captures in quiescence (~8 Elo)
  • Add aspiration windows (~5 Elo)
  • Return best_eval in quiescence (~5 Elo)
  • Fail soft on TT cuts (~2 Elo)
  • Prefer TT nodes with a higher depth (~2 Elo)
  • Don't do RFP or NMP in TT nodes (~0 Elo)

Misc

  • Clear PersistentState and Control on ucinewgame (fails SPRT at -9 Elo but is strictly more correct)
    • Allocate TT before the first search (gains +10 undoing the -9 from the time spent allocating after ucinewgame and before the first search)
  • Add prettier search output if being used interactively
  • Rename MoveProvider -> MovePicker for consistency with other engines
  • Improved debug output when MovePicker perft tests fail
  • Move search termination check and 'force stop' of search into TimeStrategy
  • Fix OOM-kills due to briefly allocating two transposition tables during ucinewgame
  • Remove git-version for setting UCI version dynamically during development

2.5

  • Fix not storing moves that caused beta cutoffs in the TT (~50 Elo)
  • Pack midgame and endgame PhasedEval i16s into a single i32 (~19 Elo)
  • Do PVS by searching first move with the full window and the remainder with a zero-window (~4 Elo)
  • Perform all TT updates in the same place in negamax (~1 Elo)

Misc

  • Use UciMove instead of Move in uci
  • Encapsulate the history table in HistoryTable
  • Encapsulate the killers table in KillersTable
  • Remove color-eyre dependency

2.4

  • Expand move scoring range from 200000 to 1000000000 (~10 Elo)
  • Refactor duplicate code in MoveProvider (~9 Elo)
  • Score quiets after yielding killers and avoid scoring captures with killer scores (~7 Elo)
  • Persist and decay history heuristic data (~6 Elo)
  • Do null move pruning when static eval == beta (~5 Elo)
  • Avoid yielding the prior best move from the transposition table twice (~2 Elo)
  • Killer move fixes: bound history scores and don't allow dupes (~0 Elo)

Misc

  • Added a 'Threads' UCI option (which isn't used)
  • Changed various Cargo options
    • Disable incremental compilation in release mode
    • Switch to panic=abort
    • Stop generating debug symbols
    • Set codegen-units=1
  • Report 'uci name' as 'name version' instead of 'name (version)'

2.3

  • Implement lazy (staged) move generation (~25 Elo)
  • Index the history heuristic array by player (~9 Elo)
  • Fix a bug which overwrote killer moves with moves from another ply (~7 Elo)

Misc

  • Use atomics instead of a mutex for the shared 'stop' flag
  • Allow making multiple moves with d move [moves]

2.2

  • Restore zobrist hash, incremental evaluation fields and castle rights from history when undoing move (~17 Elo)
  • Fix throwing away old en passant target during null moves (~19 Elo)
  • Use a dedicated MoveList struct instead of Vec<Move>
  • Store castle rights as an array indexed by player
  • Correct stored mate values in TT
  • Make a panic move if there wasn't enough time to find a PV move during search

Misc

  • Split eval tapering into its own module
  • Bundle midgame and endgame evals into a PhasedEval struct
  • Fix taking up more memory than needed when the transposition table is resized repeatedly
  • Remove the default 50ms move overhead and add a UCI option to configure it
  • Always log crashes to a .crash.log file
  • Check for time termination in the root
  • Don't try reporting PV beyond actual depth searched

2.1

  • Use transposition table entries from the same depth (~101 Elo)
  • Always extend when in check (~29 Elo)

Misc

  • Avoid double-counting 'root' quiescence nodes

2.0

  • Add null move pruning
  • Add killer move ordering
  • Sort moves via individual move scoring
  • Sort moves incrementally
  • Add reverse futility pruning
  • Add history heuristic move ordering

Misc

  • Remove the ability to specify alternate strategies

1.1

  • Disable logging by default
  • Use Rust 1.75
  • Switch movegen to use orthogonal/diagonal pin approach from this article
  • Use .get_unchecked() for all static array accesses (-1.18% perft(8) time)
  • Store Square as a u8 instead of a Bitboard internally
  • Use an array for PlayerPieces (-6.37% search(9) time)
  • Generate attackers for single squares instead of all attacks in movegen (+8.0 Elo)
  • Optimise castle move generation
  • Don't generate non-capture underpromotions in quiescence search
  • Remove the Ctx struct from movegen (-3.13% perft(7) time)
  • Reorganise everything into a single crate
  • Re-enable incremental compilation for an unexplained performance boost in sort_unstable_by
  • Consider bishops more valuable than knights for MVV-LVA

Misc

  • Add SAN parsing and formatting
  • Add the 'Win At Chess' test suite
  • Add Justfile commands for STC and LTC tests
  • Use u64 for node counts to prevent overflows with large perft results
  • Remove support for go searchmoves and go mate
  • Collapse castle detection for kingside/queenside into a single code path
  • Various refactoring and simplification around Bitboard and Square abstractions
  • Remove EngineGame
  • Add a wait extension to allow piping go commands to the engine for benchmarking
  • Make the halfmove clock and fullmove number optional in FEN parsing
  • Add a way to easily jump to useful debugging positions (e.g. d position kiwipete)
  • Add the ability to pass UCI commands to run as command line arguments

1.0

Initial release with the following major features:

  • Board

    • Bitboard board representation
    • Redundant mailbox representation for square lookups
    • Zobrist hashing

  • Move generation

    • Fully legal move generation (~200 million NPS)
    • Fancy Magic bitboards

  • Search

    • Negamax
    • Iterative deepening
    • Quiescence search
    • Principal Variation Search (PVS)
    • Check extensions
    • Transposition table

  • Move ordering

    • Previous best move
    • Most Valuable Victim - Least Valuable Aggressor (MVV-LVA)

  • Evaluation

    • Material difference
    • Midgame and endgame piece square tables
    • Tapered midgame vs. endgame evaluation
    • Incremental updates