Skip to content

Neural MMO v1.3

Choose a tag to compare

@jsuarez5341 jsuarez5341 released this 01 Jul 02:06

Prebuilt IO Libraries

  • Blade: We have improved and streamlined the previously unstable and difficult to use IO libraries and migrated them here. The new API provides framework-agnostic IO.inputs and IO.outputs functions that handle all batching, normalization, serialization. Combined with the prebuilt IO networks in Ethyr, these enable seamless interactions with an otherwise complex structured underlying environment interface. We have made corresponding extensions to the OpenAI Gym API to support variable length actions and arguments, as well as to better signal episode boundaries (e.g. agent deaths). The Quickstart guide has been updated to cover this new functionality as part of the core API.
  • Trinity: Official support for sharding environment observations across multiple remote servers; performance and logging improvements.
  • Ethyr: A Pytorch library for dynamically assembling hierarchical attention networks for processing NMMO IO spaces. We provide a few default attention modules, but users are also free to use their own building blocks -- our library can handle any well defined PyTorch network. We have taken care to separate this PyTorch specific functionality from the core IO libraries in Blade: users should find it straightforward to extend our approach to TensorFlow and other deep learning frameworks.
  • Embyr: Agents now display additional information overhead, such as when they are immune to attacks or when they have been frozen in place.
  • A reasonable 8-population baseline model trained on 12 (old) CPU cores in a day.
  • Improved and expanded official documentation
  • New tutorials covering distributed computation and the IO API
  • The Discord has grown to 80+! Join for active development updates, the quickest support, and community discussions.