Skip to content

Latest commit

 

History

History
53 lines (38 loc) · 2.11 KB

CHANGELOG.md

File metadata and controls

53 lines (38 loc) · 2.11 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

As long as the mayor version is 0, the API is not stable and may change with minor versions.

The minor version will be incremented if any of the following changes:

  • The HTTP API
  • The C API
  • The inputs for the neural networks

This means you can reuse the same neural networks between for example 0.2.0 and 0.2.1, but not between 0.1.0 and 0.2.0.

Unreleased

  • changed Default neural nets are now compiled into the executable

0.2.0 - 2023-11-26

  • added Batch inference of neural networks.
  • changed Improved selection of positions for rollouts via self play.
  • changed Different neural networks for contact and race.
  • changed Rollout data is now stored with GnuBG position IDs.
  • changed Use Hardsigmoid instead of ReLU for hidden layers.
  • changed Use AdamW instead of SGD as PyTorch Optimizer.
  • changed Use L1Loss instead of MSELoss as loss function during supervised training.
  • added The C API now supports raw evaluation of positions.
  • changed The C API doesn't need to reload the neural nets for every call.
  • added Documentation for engine and the training process.

0.1.0 - 2023-10-17

Initial release of wildbg.

  • added Move generation.
  • added Finding positions for rollouts via self play.
  • added Rollouts with fixed number of 1296 games for generating training data.
  • added Training of new neural networks with PyTorch.
  • added Inference of existing neural networks with tract.
  • added Comparison of neural networks by playing against each other.
  • added Implementation of the GnuBG position ID.
  • added HTTP API for moves and cubes.
  • added Simple C API for best move in 1-pointer.

Thanks for contributions