First release featuring a true NNUE evaluation
NNUE
Replaced the previous neural evaluation with a true NNUE implementation
King-centric HalfKP feature set
Efficient incremental evaluation updates
5-million-parameter network (≈20 MB)
Trained using my own custom LibTorch/C++ trainer
Training data generated using Ceibo 1.0 self-play games
Completely redesigned training pipeline
Time Management
Fixed several issues affecting time management
Improved behavior across a wide range of time controls
Better support for increment and sudden-death time controls
Search
Search tuning to complement the new NNUE evaluation
Various search and performance improvements
Misc
Bug fixes
Code cleanup and general optimizations
First release featuring a Neural Network evaluation
Neural Network
Replace hand crafted evaluation with a neural network implementation
Trained using my own custom LibTorch/C++ trainer
Training data generated using Ceibo 0.9 self-play games
Search
Tuned Futility Pruning
Tuned Null Move Pruning
Tuned History Heuristics
Added Aspiration Window
Added Syzygy Tablebase support
Added Dynamic Time Management
Reworked multithreading
Move generator
Modified into staged move generator
Various Minor speed improvements on the generator and make/undo
Improved validations to prevent illegal moves
Eval terms
Scaling based on rule 50
Reworked mobility eval
Added tempo eval term
Knight/Bishop outpost
Knight/Bishop behind pawn
Blocked Pawns on bishop color
Rook/Queen infiltration
Misc
Added upper/lowerbound info strings
Various other smaller changes.
Move Ordering
Static Exchange Evaluation (SEE): applied to captures only for now.
Other minor changes
Search
Qsearch: SEE Pruning
Internal Iterative Reductions
Fixed Futility Pruning issue
Transposition Tables
Fixed Major move ordering issue
Fixed Minor en passant issue
4 buckets per entry
Evaluation:
semi and open file rook
isolated pawn
enhancements to mobility
Board Representation: Bitboards! (old school; no rotated/no magic)
Movegen: completely reworked from scratch
Evaluation:
added king safety (my engine actually tries to protect the king ! how about that? )
added safe mobility
added more pawn structure terms
texel tuned
Null Move: tuned parameters, removed useless conditions and added new criteria
LMR: tuned for longer Time Controls
fixed issue where moves to mate where incorrectly reported
removed bugs and added new ones
+ many minor changes, enhancements, code clean up, etc
Ceibo 0.7 64-bit in CCRL 40/15 : ELO 2611 +/- 70 (preliminary)
Ceibo 0.7 64-bit in CCRL Blitz : ELO 2601 +/- 25
Late Move Reduction
SMP *
minor enhancements and code clean up
* Up to 4 threads. Unfortunately I only have a 4 core CPU and therefore could not not test beyond that. However, I have not enforced a max thread count and should work with higher concurrency just as well.
Ceibo 0.6 64-bit in CCRL 40/15 : ELO 2378
Ceibo 0.6 64-bit in CCRL Blitz : ELO 2390
Ceibo 0.5 64-bit in CCRL 40/15: ELO 2351
Ceibo 0.5 64-bit in CCRL Blitz : ELO 2325
Ceibo 0.4 64-bit in CCRL 40/15 : ELO 2021
Ceibo 0.4 64-bit in CCRL Blitz : ELO 1994