Unity-Technologies/ml-agents 0.8.0

ML-Agents Beta 0.8.0

on GitHub

New Features

• Enable training with multiple concurrent Unity simulations.
• Ability to exchange custom protobuf messages.
• Support for 2D ray casting.
• Enable tracking and logging training metrics.
• Splitting ML-Agents package into mlagents.trainers and mlagents.envs.
• Ability to also use RenderTexture for Visual Observations.

Fixes & Performance Improvements

• Various improvements to the documentation.
• Various other bug fixes.

Known Issues

• Inference does not work with multi-discrete action when training with LSTM.

Acknowledgements

Thanks to everyone at Unity who contributed to v0.8.0, as well as: @LeSphax, @pyjamads, @tkggwatson, @malmaud, @supercurious and @rafvasq. Special thanks also go out to our partners Jam City who helped us evaluate the performance gains of this release on a real game, Snoopy Pop.