1.0

Release 1.0 (22 Jul 2024)

• NNUE evaluation (768 - 1024)x2 -> 5 buckets
• Introduced bucketed L2 based on queen-count/phase, increased HL to 1024
• Increased training dataset to 700m positions
• Relabeled positions using Odonata's NNUE rather than HCE
• Fixed some bugs in the tuner
• Expected Elo +150 vs Odonata 0.9.0

0.9.0

Odonata 0.9.0 (6 Jun 2024)

• NNUE evaluation (768 - 512x2 - 1 network)
• Trained using simple self-written CPU-device NN trainer (mini-batch, AdamW), and own self-play and evals.
• Expected Elo +250 to +300 vs Odonata 0.8.0

v0.8.0

Odonata 0.8.0 Changelog

• Connected rooks and doubled rooks evals fixed
• Training migrated from LBFGS (argmin) to Adam (custom coded) in anticipation of Odonata NNUE
• Increase in training dataset to 4m+ positions
• Migration to rustc 1.75
• Elo +30 (maybe?) vs Odonata 0.7.0

Use the "modern" binary for best performance. Use the "generic" binary if the modern binary does not run on your hardware (pre-Haswell CPU for example)

v0.7.0

Odonata 0.7.0 changelog

• Release to kickstart development again after a long pause
• Updated rust library dependencies
• Improved zugzwang detection on null move pruning
• Futility pruning uses classical piece evals
• Eval changes - candidate passed pawns and distant pawn bishops bonus
• Rust dependencies updated
• Node counts and NPS include q-search nodes

v0.6.2

Odonata 0.6.2 changelog

• Pawn structure examined during eval
• Eval caching
• Promos accounted for in q-search
• Internal search metrics

(There was a compilation issue with windows-0.6.0 - please upgrade to 0.6.2)
(22/06/22 windows binaries updated again just to correct reported version number - no change in play/performance)

v0.5.1

Release 0.5.1 replaces release 0.5.0 which had a fatal issue involving Hash resizing on Windows. Please upgrade to 0.5.1.

Release 0.5.1

• Elo improved by 200+ in self-play vs release 0.4.0 (expected Elo 2500+)
• Late move reductions, razoring, check extensions, time control tuning
• Greatly simplified UCI options
• Rook behind passer eval adjustment
• Phase adjusted logistic k-steepness incorporated into texel-style tuning
• Some items built but not tuned, finished or adding Elo yet. Notably
  • Aspiration searches
  • (Alternate/direct addressed) bucket hashing and aligned hash
  • MultiPV (experimental)
  • TOML configuration/profile files
  • Tri-valued evaluation weights

v0.4.0

Odonata 0.4.0

Changelog

• Evaluation of king tropism, passed pawn on 7th
• Bishop and knight outposts
• Rook batteries
• History table
• Xray attacks considered in SEE
• Texel tuning uses Rayon library for parallel evals (138 M/sec)
• Check for insufficient material in quiesensce
• CLOP tuning of non-eval params
• Support for "searchmoves" and "info hashfull" in UCI interface
• ELO in the region of 2300

v0.3.37

Highlights

• Texel style tuning for piece values
• Pre-compiled binaries should run on more CPUs (not fully tested)
• Tuning tooling
• Recursive null move pruning
• Material balance, move to compiler version nightly-2021-08-04
• Fix repetition detection bug
• Check extensions and late move reduction
• Piece square table for rook end game, pawn pst corrected
• Futility margin adjustment, PVS re-enabled

Building from source and testing

See Compilation

v0.3.26

Highlights

• 180 Elo improvement from 0.3.17
• Futility pruning
• Null move pruning
• Multi-threading (still experimental)
• Command-line switches for benchmarking and perft (use --help)
• Bishop pair bonus
• Json-rpc for python interface (still experimental)
• Lockless transposition table

v0.3.17

• pondering
• fixed bug with pv lenght on uci info
• Python: added move_attributes and attacks_from
• added mate_in_4 tests
• make_moves taking a variation
• removed pseudo_legal_move logic
• replace MoveList vec with array
• added magic bitboards for benchmarking (and completeness). Not my own code. Just a rust "port" of C++.
• cargo features to enable release-like configs
• disable piece mobility for low plys