Hands-On Reinforcement Learning for Games

MAML-RL

The learn2learn repository holds another great example of how to use their library for a few variations of this method. A good method for us to look at will be an implementation of Meta-SGD, which further extends MAML by adopting per-parameter learning rates using vanilla policy gradients, and is often referred to as MetaSGD-VPG. This concept was originally presented in the paper Meta Reinforcement Learning with Task Embedding and Shared Policy, which was itself presented at IJCAI-19.

Make sure that you have completed all the installation steps from the last exercise before proceeding. If you have troubles running the sample, repeat the installation in a new virtual environment. Some issues may be related to the version of PyTorch ...