Rusty-rival 1.0.77 - new version chess engine
Rusty Rival - UCI (Universal Chess Interface) chess engine written in Rust.
Author:Chris Moreton Rating Chess Engines Diary CEDR=3474
The engine can be about 50% faster if compiled on the machine on which it will be run in order to take advantage of cpu-specific instructions.
Rusty-rival what's new?
Main-search strength improvement
Transposition-table scores can terminate or narrow a search only at scout nodes, whose search window is one point wide. Principal-variation (PV) nodes now search their wider window without those table-score cutoffs or adjustments. Hash-move ordering, cached static evaluation and singular-extension metadata remain available.
This release contains the independently confirmed PV-TT candidate only. The neural network, evaluation scale and other search parameters are unchanged. The separately tested null-move gate is not included.
Strength evidence
A preregistered, independent confirmation against v1.0.76 completed 12,000 games (6,000 colour-swapped opening pairs) at an accelerated engine-test clock of 10 seconds + 0.1 second per move:
Estimated improvement: +15.76 Elo, with a pair-based 95% confidence interval of +11.81 to +19.72 Elo.
Score: 52.2667%. The preregistered requirement was a nondegenerate 95% lower score bound above 50%; this run passed.
Fresh matched PGO builds, one thread per engine, Hash 128 MB, eight concurrent games and the same Syzygy files on one host.
All games ended by normal chess rules: no resign, draw, maximum-move or tablebase adjudication. Independent replay verified legal moves, terminal results, colour-swapped pairs and no clock flags.
Confirmation openings were reserved before the pilot and disjoint from pilot and historical exclusion roots using board, side-to-move and castling rights.
The preceding 4,000-game PV pilot at 3+0.03 selected this candidate with +13.99 Elo [6.49, 21.50]. Pilot results were not pooled with confirmation. No retries, extensions or alternate-candidate confirmation were used.
These measurements support strength at the tested clock and conditions. They are not a CEDR rating, a prediction for human bullet/blitz/rapid clocks, or measured gains on other hardware or thread counts.
Correctness and release qualification
The confirmed source passed 303 release tests, with zero failures and three existing ignored tests, plus its 10 targeted debug tests and mutation checks. A root-fallback test's node budget was recalibrated after trace diagnostics proved that this search change moved the fixture's fail-high interval; its assertions were preserved.
The confirmed PGO binary matched its own non-PGO source on 80 fixed-node and 60 held-out fixed-depth positions. This is compiler/build identity for the changed search, not node-count identity with v1.0.76.
Versioned packaging and independent public-artifact verification passed. The public Linux AVX2 binary matched the qualified versioned PGO build on 80 fixed-node positions, 60 held-out positions and all 16 depth-20 bench signatures. The portable x86-64-v2 build also passed 23 library tests and 80+60 cross-kernel identity checks on the test host. Other platform assets were verified by digest and embedded-network identity; this does not claim execution testing on those platforms.
Comments
Post a Comment