Since you’re here...

We hope you will consider supporting us today. We need your support to continue to exist, because good entries are more and more work time. Every reader contribution, however big or small, is so valuable. Support "Chess Engines Diary" even a small amount– and it only takes a minute. Thank you.
============================== My email: jotes@go2.pl



Triumviratus 5.1 - new version chess engine

on

Triumviratus is a strong, UCI-compliant chess engine written in C++. 
Author: Francesco (fraTorsello)

Triumviratus 5.1 what's new?

vs 5.0

+64.7 ± 7.6 Elo vs 5.0 (2000 games, 20+0.2, AVX2 build, same network, no adjudication, LOS 100%).
This is the official version-bump gate — see README → Results for the full table.

What's new since 5.0

  • Second-audit patch (+26 Elo cumulative, SPRT-confirmed): threat-indexed quiet history,
    refined TT-cutoff, aspiration/alpha-raise/fail-high tweaks, an SPSA-tuned singular/extension
    vector.
  • NPS-optimization patch (+26 to +42 Elo depending on time control, SPRT-confirmed): a lazy
    NNUE-mirror apply (board mirror + threat computation for a move deferred until an evaluation
    actually needs it) plus a -mtune=native PGO build flag — a real, measurable strength gain,
    not just raw NPS.
  • Recalibrated eval scale, two-level TT, hindsight extensions, faster SEE/AVX-512 accumulators,
    re-tuned time management.
  • A display-only eval normalization ("+1.00" ≈ 50% win probability) — search, TT, and time
    management are untouched; only the printed score cp is rescaled.
  • Quieter startup: the ASCII banner on launch is now suppressed (engine stays silent until it
    answers uci, matching strict UCI etiquette and making log/GUI parsing cleaner).

Screened and left off (no net gain)

This cycle, ~12 additional search toggles were implemented, bench-swept for a sensible parameter
range, and SPRT-tested — all resulted in a neutral or negative signal and remain off by default:
GoodCapHistDiv, QSMoveCap, RFPHistThresh, NMPStaticMargin/NMPTTNoisy/EasyCapGate,
TTCutMalus (retested), PostLMRHist, KillerReset, HistTrivGuard, MalusPct,
SingularPlyGuard, AspAvg, FHBoostMargin, QSDrawCheck.

A methodological note for anyone reading the SPRT logs: three of those (AspAvg,
FHBoostMargin, QSDrawCheck) initially looked like ~9-11 Elo gains and were briefly baked in,
but a clean re-test (isolated, no early stopping) showed the effect was noise — a combination of
winner's-curse (picking the best of many near-zero-signal tests) and stopping the SPRT early on a
favorable-looking LOS. They were reverted before this release; the lesson is now written up in
source/PIPELINE_5.1.md for future tuning sessions.

A first Time-Management toggle (TMStability, soft-time reduction on a stable root best-move)
was implemented and is ready, but flat at short time controls — it needs re-testing at long TC
(40+/60+) before a verdict, since it's a reduction feature that needs long searches to matter.

vs. external engines

  • Pawnocchio 1.9.1: essentially even at long TC (60+0.6, 2000 games: +2.6 ± 7.5, within
    noise), a small edge at faster TC (20+0.2, 800 games: +9.6 ± 12.9).
  • Berserk 14: −25.9 ± 18.1 (25+0.25, 322 games).

Build

  • Network: nn-rubicon-alea-v1.nnue (SFNNv13, own-lineage), unchanged from 5.0.
  • Shipping binaries: clang-cl + ThinLTO + PGO, both AVX2 (CCRL-compatible) and AVX-512 targets.
    See source/BUILD_NOTES.md.

Downloads — two variants per architecture

  • Standard (Triumviratus_5.1_avx2.zip / ..._avx512.zip): the release build
    (-DTRIUMV_RELEASE) — only the standard UCI options are exposed (Hash, Threads,
    Move Overhead, EvalFile, SyzygyPath). Use this for play/tournaments.
  • ..._DEVELOPMENT_VERSION.zip: same search and network, built without -DTRIUMV_RELEASE,
    so every internal tuning parameter (SPSA spins, A/B toggles) is exposed as a UCI option. Use
    this only if you want to experiment with the search constants yourself — the extra options do
    not change default behavior (all default to the same values as the standard build).

Tooling (not shipped, internal)

A standalone self-play datagen suite for a possible future rubicon-alea refinement pass
(T80 + a self-generated data fraction) now targets 5.1 (previously built for the experimental
6.0 bullet-net line). Not part of this release; mentioned here for provenance.



Comments

Post a Comment