v0.25.0-rc2

• Increased upper limit for maximum collision events.
• Allow negative values for some of the extended moves left head parameters.
• Fix a critical bug in training data generation for input type 3.
• Fix for switching between positions in uci mode that only differ by 50 move
  rule in initial fen.
• Some refinements of certainty propagation.
• Better support for c++17 implementations that are missing charconv.
• Option to more accurately apply time management for uci hosts using
  cuteseal or similar timing techniques.
• Fix for selfplay mode to allow exactly book length total games.
• Fix for selfplay opening books with castling moves starting from chess960 fens.
• Add build option to override nvcc compiler.
• Improved validity checking for some uci input parameters.
• Updated the Q to CP conversion formula to better fit recent T60 net outputs to
  expectations.
• Add a new optional experimental time manager.
• Bug fix for the Q+U in verbose move stats. It is now called S: and contains
  the total score, including any moves left based effect if applicable.
• New temperature decay option to allow to delay the start of decay.
• All temperature options have been hidden by default.
• New optional cuda backend convolution implementation. Off by default for
  cudnn-fp16 until an issue with cublas performance on some gpus is resolved.

v0.25.0-rc1

• Now requires a c++17 supporting compilation environment to build.
• Support for Moves Left Head based networks. Includes options to adjust search to favour shorter/longer wins/losses based on the moves left head output.
• Mate score reporting is now possible, and move selection will prefer shorter mates over longer ones when they are proven.
• Training now outputs v5 format data. This passes the moves left information back to training. This also includes support for multiple sub formats,
  including the existing standard, a new variant which can encode FRC960 castling, and also a further extension of that which tries to make training
  data cannonical, so there aren't multiple positions that are trivially equivalent with different network inputs.
• Benchmark now includes a suite of 34 positions to test by default instead of just start position.
• Tensorflow backend works once more, almost just as hard to compile as it used to be though.
• --noise flag is gone, use --noise-epsilon=0.25 to get the old behavior.
• Some bug fixes related to drawscore.
• Selfplay mode now defaults to the same value as match play for --root-has-own-cpuct-params (true).
• Some advanced time management parameters are now accessed via the new --time-manager parameter instead of individual parameters.
• Windows build script has been modernized.
• Separate Eigen backend option for CPU.
• Random backend no longer requires a network.
• Random backend supports producing training data of any input format sub type.
• Integer parameters now give better error messages when given invalid values.

v0.24.1

• Fix issues where logitq was being passed as drawscore and logitq wasn't passed to some GetQ calls. Causing major performance issues when either setting was non-default.

v0.24.0

• New parameter --max-out-of-order-evals-factor replaces --max-out-of-order-evals that was introduced in v0.24.0-rc3 and provides the factor to multiply the maximum batch size to set maximum number out-of-order evals per batch. The default value of 1.0 keeps the behavior of previous releases.
• Bug fix for hangs with very early stop command from non-conforming uci hosts.

v0.24.0-rc3

• New parameter --max-out-of-order-evals to set maximum number out-of-order evals per batch (was equal to the batch size before).
• It's now possible to embed networks into the binary. It allows easier builds of .apk for Android.
• New parameter --smart-pruning-minimum-batches to only allow smart pruning to stop after at least k batches, preventing instamoves on slow backends.

v0.24.0-rc2

• All releases are now bundled with network id591226 (and the file date is old enough so it has a lower priority than networks that you already may have in your directory).
• Added a 'backendbench' mode to benchmark NN evaluation performance without search.
• Android builds are added to the official releases.

v0.24.0-rc1

• Introduced DirectX12 backend.
• Optimized Cpuct/FPU parameters are now default.
• There is now a separate set of CPuct parameters for the root node.
• Support of running selfplay games from an opening book.
• It's possible to adjust draw score from 0 to something else.
• There is a new --max-concurrent-seachers parameter (default is 1) which helps with thread congestion at the beginning of the search.
• Cache fullness is not reported in UCI info line by default anymore.
• Removed libproto dependency.

v0.23.3

Single bugfix release that should reduce the frequency of incorrect instamoves when the search tree becomes large in a prior move.

v0.23.2

• Fixed a bug where odd length openings had reversed training data results in
  selfplay.
• Fixed a bug where zero length training games could be generated due to
  discard pile containing positions that were already considered end of game.
• Add cudnn-auto backend.

v0.23.1

• Fixed a bug with Lc0 crashing sometimes during match phase of training game generation.
• Release packages now include CUDNN version without DLLs bundled.