Lambergar - UCI chess engine Rating Chess Engines Diary CEDR=3364
Lambergar is a chess engine developed in the Zig programming language.
The name "Lambergar" is a nod to the Slovenian folk romance, Pegam and Lambergar, which recounts the epic struggle between Jan Vitovec and Krištof Lambergar (Lamberg). This narrative of fortitude and rivalry provided a fitting namesake for this chess engine.
Lambergar is licensed under the MIT License. Check out LICENSE.txt for the full text. Feel free to use this program, but please credit this repository in your project if you use it. github:https://github.com/jabolcni/Lambergar/releases/
Lambergar 1.5 - what's new?
Lambergar 1.5 is a major release over 1.3. It adds Syzygy tablebases, Chess960/FRC and DFRC support, a full data-generation and HCE tuning pipeline, a larger hand-crafted evaluation, more UCI-exposed search tuning, many stability fixes, and search and robustness improvements.
Builds
Currently, there are four release builds:
x86-64 Windows AVX2: recommended Windows build for most modern CPUs.
x86-64 Windows AVX-512: optional high-end Windows build for CPUs with AVX-512 support.
x86-64 Linux AVX2: recommended Linux build for most modern CPUs.
x86-64 macOS AVX2: recommended macOS build for Intel Macs with AVX2 support.
Zig 0.15.1 is recommended for release builds. Zig 0.16.0 can compile the engine,
but local testing showed a large NPS regression in both NNUE and HCE modes.
Older x86-64 Vintage/POPCNT builds and aarch64 builds are not included in this
release because the current NNUE path uses x86-64 AVX2-oriented code.
Release Notes
Syzygy tablebases: Added experimental Syzygy endgame tablebase support through the Fathom library, including SyzygyPath, SyzygyProbeDepth, probe, probebest, WDL/root probing, and tablebase probing inside search.
Chess960/FRC: Added UCI_Chess960, Chess960-aware UCI castling notation, Shredder-FEN castling rights, Chess960 castling make/unmake, Chess960 castling legality checks, and position startposfrc <index>.
DFRC: Added Double Fischer Random support with independent white/black start indices through position startposdfrc <white_index> <black_index>.
Position handling: Reworked FEN parsing/export around fullmove numbers, Chess960 castling metadata, castling-right clearing by square, FEN diagnostics, and internal pawn/non-pawn/material hash keys.
Move generation: Added a shared MoveGenContext carrying occupancy, danger, checker, pin, and king data for reuse across legal, quiet, noisy, evasion, and search move generation.
Pinned-piece legality: Reworked pinned pawn and pinned slider generation, including quiet pinned moves, noisy pinned moves, en-passant legality with discovered-check handling, and single-check/double-check evasion paths.
Attack tables: Added line and in-between square tables (LINE, SQUARES_BETWEEN_BB) and centralized initialization for rook, bishop, line, between-square, and pseudo-legal attack databases.
Search parameters: Added UCI-exposed search parameters for null move pruning, futility pruning, razoring, LMR, SEE pruning, history scaling, and qsearch margins.
Search framework: Reworked iterative deepening, aspiration windows, root score tracking, seldepth reporting, node accounting, early-stop logic, and soft/hard time limits.
Lazy SMP: Added per-thread IDs, per-thread move-ordering seed diversification, helper-thread aspiration deltas, and worker node aggregation in search output.
Tablebase search integration: Added main-thread tablebase gating, TB win/loss score handling, and tablebase hit reporting in search info.
Move picker: Added a staged main move picker with TT move, captures, killers, countermove, deferred quiet scoring, and quiet-batch skipping.
Move ordering: Added capture history, threat-aware quiet history, continuation-history scoring, deferred quiet scoring, randomized tie-breaking salt, and improved promotion/capture scoring.
SEE: Reworked SEE helpers with see_ge, faster pruning checks, en-passant-aware capture handling, promotion-aware exchange values, and cheap lower-bound tests.
Null move pruning: Fixed null move pruning to return the actual null-window score on successful prune instead of stale best_score, and added tunable reduction, beta margin, and verification controls.
Pruning: Added or expanded history quiet pruning, deep continuation-history pruning, late move pruning, futility pruning, razoring, main-search SEE pruning, and qsearch delta pruning.
Futility pruning: Changed futility pruning from setting skip_quiets after searching the triggering move to a hard skip when the collected fail-low data justified it.
LMR and extensions: Added tuned LMR scaling, PV/cutnode/improving/context adjustments, singular extensions, double extensions, multi-cut correction updates, and removed a dead killer-move LMR adjustment.
Quiescence search: Added TT reuse/storage in qsearch, qsearch TT move ordering, noisy move ordering, stand-pat handling, delta pruning, and SEE pruning.
Correction history: Expanded correction history from the older pawn/non-pawn form into pawn, white non-pawn, black non-pawn, major, minor, continuation, and previous-move correction components.
History learning: Added capture-history updates, threat-indexed quiet-history updates, continuation-history updates at multiple offsets, countermove updates from node state, and depth-1 history learning.
Transposition table: Reworked TT storage into two-entry buckets with age tracking, replacement scoring, exact-bound preference, hashfull sampling, hit-rate counters, prefetch helpers, and lock-free published-entry fetches.
Thread safety: Added TT publish flags with acquire/release semantics, spin locks for age/hashfull/clear paths, and safer TT resize/deinit behavior.
HCE parameters: Moved tuned HCE parameters into hce_params.zig, including material, PSQT, pawn structure, mobility, threats, king safety, pawn storm, outposts, rook files, bishop pair, and tempo.
Pawn hash: Added a pawn hash table for HCE pawn structure, pawn attacks, passed pawns, blocked passers, candidate pawns, supported pawns, phalanxes, backward pawns, isolated pawns, doubled pawns, and pawn attack data.
HCE features: Added new HCE features for knight/bishop outposts, bishop and rook traps, rook behind passer, connected rooks, blockade of passed pawns, king pawn shield, king virtual mobility, pawn storm, candidate pawns, and piece threats.
Endgame evaluation: Added extra endgame handling for bare-king material, bishop+knight mate guidance, king centralization, and king-distance terms.
Tuner integration: Reworked the tuner feature counters to match the expanded HCE and to support the binhce pipeline instead of only the old text/pickle-style flow.
Datagen: Added engine selfplay datagen with configurable games, depth, plies, random opening moves, strict mode, noisy-position skipping, minimum nodes, phase sampling, start-position mix, evaluation mode, and output format.
Datagen formats: Added bin40 writing, binhce writing with HCE feature vectors, SFEN32 packing/unpacking helpers, move16 encoding, and bin_to_binhce conversion tooling.
Fast binhce loading: Added a C ABI loader for binhce files so Python tuning tools can load scores, moves, plies, results, SFEN32 records, and feature tensors efficiently.
NNUE loading: Added embedded NNUE initialization, external NNUE file initialization, integrity checks, feature/output layer loading, and shared initialization helpers.
NNUE accumulator: Reworked NNUE accumulator refresh and incremental update around king buckets, oriented squares, piece-index tables, delta pieces, null-move copying, and full-refresh fallback.
NNUE diagnostics: Added debug-only NNUE stats, search-context tracking, king-refresh cause tracking, accumulator validation support, and NNUE microbenchmark tooling.
SIMD NNUE path: Kept AVX2-oriented SIMD affine evaluation and vectorized accumulator refresh paths for current x86-64 release builds.
UCI protocol: Added safer parsing, exit alias, immediate isready behavior, better malformed-input handling, option clamping, Chess960 synchronization, and UCI output fixes.
UCI debug commands: Added or expanded validate, perftdiv, perftstats, perftchild, hcefeatures, hcefeatures_sfen32, probe, and probebest.
Bench commands: Added bench for OpenBench-style validation and benchhce for HCE-only validation.
Perft: Reworked perft around the new move generator, added richer statistics, divide output, child-node inspection, timing, and more regression test coverage.
Tests: Added Zig test coverage for perft, SEE, FEN round-tripping, repetition, insufficient material, in-check detection, evaluation symmetry/bounds, HCE feature determinism, and NNUE incremental-vs-full-refresh consistency.
Python validation tools: Added helpers for UCI compliance, FRC/DFRC start positions, tablebase testing, PGN validation, datagen validation, binhce inspection, and HCE visualization.
Automation: Added Python automation for gauntlets, datagen runs, large-scale datagen orchestration, HCE tuning, bench statistics, go-depth statistics, and Zig-version comparison.
Documentation: Added dedicated docs for datagen, binhce conversion, data validation, HCE tuning, and search-feature impact tracking.
Build system: Added Fathom C sources, optional tablebase build wiring, Zig 0.15.1/0.16.0 compatibility shims, main_015.zig, bin_to_binhce_015.zig, and release tooling with selectable Zig executable.
Bug fixes: Fixed a latent LMP table out-of-bounds read, Threads values above MAX_THREADS, go movetime 1 unsigned underflow, transposition table double-free during Hash resizing, malformed debug/UCI input panics, and several Chess960/FEN/move-parsing edge cases.
Cleanup: Removed dead search constants and dead branches, consolidated NNUE initialization, replaced repeated TT hash-size literals with HASH_SIZE_DEFAULT + 1, and moved datagen NNUE initialization after common engine initialization for consistency.
--------------------------------------------------
Results of Lambergar_1_5 vs Lambergar_1_3 (10+0.1, 1t, 128MB, UHO_2024_8mvs_big_+115_+134.pgn):
Elo: 106.77 +/- 8.43, nElo: 206.35 +/- 15.30
LOS: 100.00 %, DrawRatio: 36.87 %, PairsRatio: 10.36
Games: 1980, Wins: 852, Losses: 262, Draws: 866, Points: 1285.0 (64.90 %)
Ptnml(0-2): [0, 55, 365, 495, 75], WL/DD Ratio: 1.31
--------------------------------------------------
--------------------------------------------------
Results of Lambergar_1_5 vs Lambergar_1_3 (60+0.5, 1t, 128MB, UHO_2024_8mvs_big_+115_+134.pgn):
Elo: 104.99 +/- 7.32, nElo: 219.32 +/- 14.39
LOS: 100.00 %, DrawRatio: 35.80 %, PairsRatio: 11.84
Games: 2240, Wins: 965, Losses: 308, Draws: 967, Points: 1448.5 (64.67 %)
Ptnml(0-2): [0, 56, 401, 613, 50], WL/DD Ratio: 1.69
--------------------------------------------------
Lambergar 1.2 vs other engines:
| Critter 1.6a | 10.5/18 | +3 | 18 Games |
| Bitbit 1.4 | 9/18 | +0 | 18 Games |
| Velvet 9.0.0 dev7 JA | 2.5/14 | -9 | 14 Games |
| Integral 7 dev.322b JA | 2/14 | -10 | 14 Games |
| Tucano 12.17 JA | 2/14 | -10 | 14 Games |
| Illumina 3.0 JA | 6.5/13 | +0 | 13 Games |
| Knightx 4.5 | 8/8 | +8 | 8 Games |
| Jinx 1.0 JA | 8/8 | +8 | 8 Games |
| Adeon 1.2.1 | 8/8 | +8 | 8 Games |
| Cheese 3.3.0 JA | 7.5/8 | +7 | 8 Games |
| Maelstrom 3.1.1 | 6.5/8 | +5 | 8 Games |
| Tomcat 1.0 JA | 6.5/8 | +5 | 8 Games |
| Maelstrom 3.2.0 | 5.5/8 | +3 | 8 Games |
| Zahak 10.0 JA | 5/8 | +2 | 8 Games |
| Lynx 1.10.0 | 3.5/8 | -1 | 8 Games |
| Oxide 1.4.0 | 3.5/8 | -1 | 8 Games |
| Oxide 1.3.0 | 3/8 | -2 | 8 Games |
| Iris 2.0 | 2.5/8 | -3 | 8 Games |
| Uralochka 3.42 dev14 | 2.5/8 | -3 | 8 Games |
| Astra 6.0 | 1.5/8 | -5 | 8 Games |
| Halogen 13.10.0 | 1.5/8 | -5 | 8 Games |
Lambergar engine's best results in our test tournaments:
Sample game of this engine:
[Event "2024.07.18.NewEnginesTest"]
[Site "Chess Engines Diary"]
[Date "2024.07.20"]
[Round "1.63"]
[White "Lambergar 0.6.0"]
[Black "Knightx 4.1"]
[Result "1-0"]
[ECO "B30"]
[PlyCount "92"]
[EventDate "2024.??.??"]
{[%evp
0,90,19,38,40,41,51,32,66,38,46,19,6,27,34,13,17,12,19,8,15,18,22,-7,13,-3,3,28,72,-21,49,81,92,62,79,84,76,76,75,69,109,129,119,100,113,124,101,122,106,91,127,106,114,120,129,129,110,110,124,159,170,196,194,196,200,196,224,224,245,253,253,253,258,253,250,259,269,277,271,295,310,325,323,333,336,352,355,374,413,419,425,435,446]}
1. e4 c5 2. Nf3 Nc6 3. Bb5 e6 4. O-O Nge7 5. c3 d5 6. exd5 Qxd5 7. Re1
a6 8. Ba4 b5 9. Bc2 Bb7 10. d4 Rd8 11. Be4 Qh5 12. Nbd2 cxd4 13. cxd4
h6 14. h3 Ng6 15. Nf1 Nh4 16. Be3 Na5 17. Nxh4 Qxh4 18. Bxb7 Nxb7 19.
Qf3 Rd7 20. Qc6 Be7 21. Qxa6 O-O 22. Qxb5 Rfd8 23. Ng3 Nd6 24. Qc6 Ne8
25. Qe4 Nd6 26. Qxh4 Bxh4 27. a4 Rb7 28. a5 Rxb2 29. a6 Ra8 30. d5 exd5
31. a7 Bxg3 32. fxg3 Nb5 33. Rab1 Rxb1 34. Rxb1 Nxa7 35. Ra1 f6 36.
Bxa7 Kf7 37. Bd4 Rxa1+ 38. Bxa1 Kg6 39. Kf2 Kf5 40. Ke3 h5 41. Kd4 Ke6
42. Bc3 Kd6 43. Bb4+ Ke6 44. Bc5 g6 45. Ba3 g5 46. Bb4 h4 1-0
Comments
Post a Comment