PytorchRL
Project no longer kept up to date - have a look at https://github.com/chozabu/UnrealPyTorchAgents instead!
This is a project trying out RL with pytorch and ue4
setup
uses this branch of UE4Python: git@github.com:InterdimensionalGamesInc/UnrealEnginePython.git for 4.24 compat clone it into the projects Plugins folder (this should really be a sub-repo)
in windows, using global python env, packages:
tensorflow-tensorboard==0.4.0rc3
pytorch, using command from pytorch website
more?
usage
Working with this currently requires editing the python scripts to configure if there should be an attempt to load a model and continue training You will probably need to disable file loading. Files are currently saved to a folder like "/Epic Games/UE_4.24/Engine/Binaries/Win64/NNModels"
video
https://i.gyazo.com/803b7bd236e042cf321c58652eaa4428.mp4 - bipedal walker, doing fairly well
info
training is done "on the fly" in a background thread from physical instances of an agent, all sharing the same neural network
Network is TD3 based on this implementation: https://github.com/nikhilbarhate99/TD3-PyTorch-BipedalWalker-v2
interesting reads/projects
https://spinningup.openai.com/en/latest/
Reinforcement Learning: An Introduction. Second edition, in progress. Richard S. Sutton and Andrew G. Barto
unity-ml-agents