Skip to content

Renegade 1.3.0

Latest
Latest

Choose a tag to compare

View all tags
@pkrisz99 pkrisz99 released this 17 Jun 20:43
v1.3.0
b792453
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Renegade 1.3.0 (June 17, 2026)

I'm thrilled to announce version 1.3.0 of Renegade! 🚀

This release includes dozens of patches, including staged move generation, history pruning, fractional reductions, and a large number of refinements made to existing heuristics. The engine now also manages time better, and spends more time on difficult moves.

The neural network architecture is unchanged, but its training dataset had been entirely regenerated, significantly improving the data quality. The current network was trained on 8.4 billion positions from 91 million games.

As a result of these improvements, this version is up to 90 elo stronger in standard chess, and should scale better to longer time controls as it was tuned across 130 thousand games.

Once again, I'd like to thank everyone who helped me out or shown interest in this project!

Changelog

Detailed changelog
  • Additions to search
    • History pruning
    • Low-depth singular extensions
    • Triple extensions
  • Refinements to existing search heuristics
    • Reducing noisy moves
    • Updating histories on alpha raises
    • Overhauled futility pruning
    • Restrict null-move pruning and internal iterative reductions to cutnodes
    • Multicut returns beta
    • History reductions done based on up-to-date values
    • Limit depth reduction on repeated aspiration fail-highs
    • Late-move pruning move counts are adjusted based on history
    • More singular extensions
    • Always replace transposition entries from a previous search
    • Recency matters more in transposition entry replacement policy
    • Fixed aspiration window behavior
    • Less extensions very deep in the search tree
    • No longer save scores as exact in quiescence search
    • Increase SEE pruning threshold for quiets with good history
    • Score queen promotions in move ordering like regular captures
    • Reset killer moves for the next ply instead of ply+2
    • Extensively tuned search constants
  • More granularity in search
    • Added fractional reductions in LMR calculations
    • Decoupled history bonuses and penalties, both for main and for continuation history
    • The three types of correction histories contribute different amounts
  • Switching over to staged move generation
    • Avoids calculating unused information
    • Added move pseudolegality check for robustness
    • Modernized and faster movegen code
  • New neural net trained on a fully regenerated dataset
    • 91 million games played with a soft node limit of 10k and 20k
    • Using viriformat binpacks instead of raw bulletformat
    • Higher WDL proportion
    • Improved data generation ergonomics
  • Neural network evaluations are scaled according to the half-move clock
  • Improved time management
    • Increased default soft time limit budget
    • Root node count fractions (node time management) matter more
    • Considers best move stability
  • Improved static exchange evaluation
    • Accounts for pinned pieces
    • Short-circuits for castling moves
  • Speedups
    • Faster move generation
    • Fuse update operations for cached accumulators
    • Requesting huge pages on potentially compatible machines
  • Various simplifications
    • Removed check extensions
    • Aspiration window does not revert to full-width search if the window is too large
    • Removed depth condition from main search SEE
  • Other minor changes
    • Fixed uninitialized value in the LMR table
    • Support for non-power-of-2 transposition table sizes
    • More robust UCI parsing
    • Error message if transposition table initialization fails
    • Prettier formatting of non-standard commands
    • Updated WDL model, now based on games with longer time controls
    • PV lines are complete and include moves from quiescence search
    • Cleanups, dead code removal, typo fixes, etc.

Progression testing

Standard chess (UHO_Lichess_4852_v1.epd):

Elo   | 78.32 +- 1.84 (95%)
Conf  | 10.0+0.10s Threads=1 Hash=16MB
Games | N: 40000 W: 14886 L: 6019 D: 19095
Penta | [46, 1812, 8220, 9073, 849]

Elo   | 91.56 +- 3.40 (95%)
Conf  | 50.0+0.50s Threads=1 Hash=128MB
Games | N: 10000 W: 3775 L: 1199 D: 5026
Penta | [0, 319, 1923, 2621, 137]

DFRC (DFRC.epd):

Elo   | 86.41 +- 4.11 (95%)
Conf  | 10.0+0.10s Threads=1 Hash=16MB
Games | N: 10024 W: 3644 L: 1201 D: 5179
Penta | [28, 446, 2042, 2047, 449]

You know the drill, the gain will be more modest for balanced books, and against other engines.

Assets 4
Loading