Improvements

• Better repetition detection. Compared to the previous version, the engine will convert more winning positions and less often run into threefold repetitions.
• Evaluate passed pawns based on their squares (instead of just counting them)
• Evaluate pieces and passed pawns based on their rank and file distances to the kings (and replaced the old evaluation terms which only considered the total distance to the kings)
• The tuner is now faster, uses less memory and it's less cumbersome to add new evaluation terms.

Full Changelog: v0.9.0...v0.10.0

Result against previous version

8s + 0.08s

Score of Fatalii 0.10.0 vs Fatalii 0.9.0: 518 - 153 - 329  [0.682] 1000
...      Fatalii 0.10.0 playing White: 276 - 69 - 155  [0.707] 500
...      Fatalii 0.10.0 playing Black: 242 - 84 - 174  [0.658] 500
...      White vs Black: 360 - 311 - 329  [0.524] 1000
Elo difference: 132.9 +/- 18.2, LOS: 100.0 %, DrawRatio: 32.9 %

Improvements

• Faster static exchange evaluation
• Add SEE pruning
• Transposition table improvements
• Speed up code that checks if the side to move is in check and if a move gives check
• Update conditions and depth reduction formula for late move reductions and null move pruning

Full Changelog: v0.8.0...v0.9.0

Result against previous version

8s + 0.08s

Score of Fatalii 0.9.0 vs Fatalii 0.8.0: 495 - 128 - 377  [0.683] 1000
...      Fatalii 0.9.0 playing White: 266 - 55 - 179  [0.711] 500
...      Fatalii 0.9.0 playing Black: 229 - 73 - 198  [0.656] 500
...      White vs Black: 339 - 284 - 377  [0.527] 1000
Elo difference: 133.7 +/- 17.4, LOS: 100.0 %, DrawRatio: 37.7 %

Improvements

• Improved and simplified transposition table
• Improved history heuristic
• Added late move pruning

Match against previous version

8s + 0.08s

Score of Fatalii 0.8.0 vs Fatalii 0.7.0: 457 - 167 - 376  [0.645] 1000
...      Fatalii 0.8.0 playing White: 245 - 76 - 179  [0.669] 500
...      Fatalii 0.8.0 playing Black: 212 - 91 - 197  [0.621] 500
...      White vs Black: 336 - 288 - 376  [0.524] 1000
Elo difference: 103.7 +/- 17.3, LOS: 100.0 %, DrawRatio: 37.6 %

Improvements

• Add king tropism as a simple king safety evaluation
• Use kindergarten bitboards for faster sliding piece move generation
• Update transposition table entry age on access
• Store move list on the stack instead of the heap if it is small (up to 32 moves)

Match against previous version

8s + 0.08s

Score of Fatalii 0.7.0 vs Fatalii 0.6.1: 485 - 281 - 234  [0.602] 1000
...      Fatalii 0.7.0 playing White: 262 - 130 - 108  [0.632] 500
...      Fatalii 0.7.0 playing Black: 223 - 151 - 126  [0.572] 500
...      White vs Black: 413 - 353 - 234  [0.530] 1000
Elo difference: 71.9 +/- 19.1, LOS: 100.0 %, DrawRatio: 23.4 %

Fixes an issue that prevented the engine from searching if one of the UCI go options was set to a negative value (e.g. go wtime 109264 btime -76 winc 8000 binc 8000). It now prints a warning, replaces the negative value with 0 and starts searching. (Previously, it printed an error message without starting the search.)

Thanks @GrahamCCRL and @Mathmagician8191 for finding the issue and helping with it!

Improvements

Search

• Switch from fail-hard to fail-soft alpha-beta search
• Add static exchange evaluation (SEE)
  • Use SEE for move ordering
  • Skip losing captures in quiescence search
• Use subtree size for root move ordering
• Add check extensions
• Add delta pruning in quiescence search
• Use futility and reverse futility pruning at depths > 1
• Add late move reductions
• Improve null move pruning
  • Only do null move pruning if the static evaluation is greater than or equal to beta
  • Do not play 2 null moves in a row
• Reset killers for the next ply before playing a move
• Persist history table between searches
• Play forced moves immediately, without searching depth 1 (searching depth 1 has sometimes led to weird eval jumps)
• Tune search parameters

Time management

• Make better use of the given time in games with increment

UCI

• Include seldepth in UCI info
• No longer print the last uci info twice

Bugfixes

• Fix replacement scheme in transposition table (could cause duplicate entries)
• Disable reverse futility pruning in PV nodes (could lead to missing checkmates)

Match against previous version

Chess960, 8s + 0.08s

Score of Fatalii 0.6.0 vs Fatalii 0.5.0: 265 - 113 - 122  [0.652] 500
...      Fatalii 0.6.0 playing White: 140 - 52 - 58  [0.676] 250
...      Fatalii 0.6.0 playing Black: 125 - 61 - 64  [0.628] 250
...      White vs Black: 201 - 177 - 122  [0.524] 500
Elo difference: 109.1 +/- 27.4, LOS: 100.0 %, DrawRatio: 24.4 %

Improvements

• In this release, an evaluation tuner is added to the engine. All evaluation parameters have been tuned with it using the training positions from the Zurichess dataset quiet-labeled.v7.
• Additional evaluation parameters
  • Pawn structure: passed, isolated, backward and doubled pawns
  • Mobility
  • Bishop pair
• Search
  • Improve root move ordering: whenever a new best move is found, it is moved to the front of the move list
  • Use multiple TT entries for move ordering (if the bucket contains multiple entries with different moves for the given position)
  • Add aspiration windows
  • Include non-capturing queen promotions in quiescence search
  • Add countermove heuristic
  • Stop search after depth 1 if there is only one legal move (only if the option wtime/btime is used)

Bugfix

• Use all/both killer moves (previously, only one move was used)

Match against previous version

Chess960, 2s + 0.02s

Score of Fatalii 0.5.0 vs Fatalii 0.4.0: 309 - 125 - 66  [0.684] 500
...      Fatalii 0.5.0 playing White: 154 - 63 - 33  [0.682] 250
...      Fatalii 0.5.0 playing Black: 155 - 62 - 33  [0.686] 250
...      White vs Black: 216 - 218 - 66  [0.498] 500
Elo difference: 134.1 +/- 30.2, LOS: 100.0 %, DrawRatio: 13.2 %

Improvements

• Search
  • Added null move pruning
  • Added futility pruning
  • Added reverse futility pruning
• Transposition table
  • Now uses a bucket system with 4 entries per bucket
  • Improved replacement scheme based on entry age and depth
• Evaluation
  • Draws by insufficient material are now detected
  • Added a small tempo bonus for the side to move
  • Slightly tuned PSTs by hand (hopefully for the last time, planning to automate this in the future :D)
• UCI
  • info now prints exact mate scores. No more 327.66 scores!

Match against previous version

Chess960, 2s + 0.02s

Score of Fatalii 0.4.0 vs Fatalii 0.3.1: 383 - 65 - 52  [0.818] 500
...      Fatalii 0.4.0 playing White: 194 - 36 - 20  [0.816] 250
...      Fatalii 0.4.0 playing Black: 189 - 29 - 32  [0.820] 250
...      White vs Black: 223 - 225 - 52  [0.498] 500
Elo difference: 261.1 +/- 36.1, LOS: 100.0 %, DrawRatio: 10.4 %

Fixed a bug that affected Chess960 and caused the engine to illegally castle queenside in some positions

Enhancements

• Added support for Chess960
• Improved move ordering
  • Added killer heuristic
  • Added history heuristic
  • Queen promotions are now searched directly after the hash move
  • Underpromotions are searched last
• Improved quiescence search
  • Ignore underpromotions
  • Include all moves while in check
• Use principal variation search
• Evaluate position incrementally
• Minor optimizations

Bugfix

• Detect threefold repetition correctly if the position first occurred after a double pawn push but there was no legal en passant capture

Self-play vs. previous version

Match result at 2s + 0.1s:

Score of Fatalii 0.3.0 vs Fatalii 0.2.1: 164 - 21 - 15  [0.858] 200
...      Fatalii 0.3.0 playing White: 73 - 18 - 9  [0.775] 100
...      Fatalii 0.3.0 playing Black: 91 - 3 - 6  [0.940] 100
...      White vs Black: 76 - 109 - 15  [0.417] 200
Elo difference: 311.8 +/- 65.2, LOS: 100.0 %, DrawRatio: 7.5 %