- Staged training: training can now start in an easier environment and get harder over time, until it reaches the environment you configured. - Scenario variation: the environment can be rearranged during training, so agents learn general rules instead of memorizing one map. - Test maps: agents are measured on maps they never saw during training, showing how well they really learned. - ONNX export: trained agents can be exported in ONNX, the standard format for neural networks, ready to use in your own projects. - Network diagrams: every export now includes images that show how the agent's neural network is built. - Five new languages: French, German, Russian, Simplified Chinese and Japanese.
Improvements
- Every card in the training details now explains what it means when you hover over it. - Exports are now delivered as one compressed file per agent type. - Visual improvements to menus and lists.
Fixes
- Fixed the game opening in English for some players whose Steam is set to another language. - General stability fixes.
Posted by the game’s developer or publisher to its Steam community hub and published by Valve through Steam’s public news API. Sifter reproduces it unedited and writes none of it.