Release 0.19.0

📦 NEW: Use see threshold of +8
📦 NEW: Remove virtual loss (was probably bugged)
📦 NEW: Various threading and performance fixes/improvements.
📦 NEW: Info output improvements
📦 NEW: SCReLU in value net
📦 NEW: Time management

 --------------------------------------------------
 Results of princhess-0.19.0 vs princhess-0.18.0 (8+0.08, 1t, 128MB, 8moves_v3.epd):
 Elo: 79.17 +/- 14.63, nElo: 120.71 +/- 21.53
 LOS: 100.00 %, DrawRatio: 35.20 %, PairsRatio: 3.44
 Games: 1000, Wins: 352, Losses: 128, Draws: 520, Points: 612.0 (61.20 %)
 Ptnml(0-2): [9, 64, 176, 196, 55], WL/DD Ratio: 0.35
 --------------------------------------------------

Release 0.18.0

📦 NEW: Add MultiPV support
📦 NEW: Use generalized cpuct
📦 NEW: 50 move rule damping
📦 NEW: Perspective network
📦 NEW: Policy Net from self play data

elo_check-1  | Score of princhess-0.18.0 vs princhess-0.17.0: 327 - 240 - 433  [0.543] 1000
elo_check-1  | ...      princhess-0.18.0 playing White: 186 - 103 - 211  [0.583] 500
elo_check-1  | ...      princhess-0.18.0 playing Black: 141 - 137 - 222  [0.504] 500
elo_check-1  | ...      White vs Black: 323 - 244 - 433  [0.539] 1000
elo_check-1  | Elo difference: 30.3 +/- 16.2, LOS: 100.0 %, DrawRatio: 43.3 %

Release 0.15.2

📦 NEW: Add MultiPV support

Release 0.17.0

📦 NEW: Value net. New Data, 512 hidden layer, king and threat buckets

Release 0.16.0

📦 NEW: New value network from self play data using goober for training

Score of princhess-0.16.0 vs princhess-0.15.0: 472 - 246 - 282  [0.613] 1000
Elo difference: 79.9 +/- 18.5, LOS: 100.0 %, DrawRatio: 28.2 %

Release 0.15.1

📦 NEW: Increased mate score

Release 0.15.0

📦 NEW: New policy network architecture
📦 NEW: 384 hidden layers in eval
📦 NEW: More history by keeping full ttable between searches and 16 bit move struct
📦 NEW: Add CPuctRoot and tune
👌 IMP: Various other refactorings
🐛 FIX: Correct depth calculation
👌 IMP: Include score in tb move info

Release 0.14.1

🐛 FIX: Use an explicit architecture/target rather than native

Release 0.14.0

📦 NEW: Eval net with increase to 256 nodes in hidden layer and quantization
📦 NEW: Policy net
📦 NEW: Change net inputs to position + threats + defenses with previous move and virtual king mobility
📦 NEW: Correct training data based on tablebases
📦 NEW: Switch to fathom for tablebases
📦 NEW: Add node limit support (use node=1 for policy only play)
📦 NEW: Use FPU in selection
📦 NEW: Add policy temperature

Release 0.13.0

📦 NEW: Only expand nodes on second visit
📦 NEW: Consider visits in move selection
📦 NEW: Eval network
🐛 FIX: Use <empty> as default for SyzygyPath
🐛 FIX: Error in cp eval and sane minimums for info output

STC:

Score of princhess-0.13.0 vs princhess-0.12.0: 505 - 209 - 286  [0.648] 1000
Elo difference: 106.0 +/- 18.7, LOS: 100.0 %, DrawRatio: 28.6 %