Train your first Decision Transformer

Hugging Face · 4 years ago · 48

Hugging Face released a tutorial for training Decision Transformer models, a technique that uses transformers to solve reinforcement learning problems by generating action sequences instead of fitting value functions. The training takes approximately 40 minutes on Google Colab using the halfcheetah-expert offline RL dataset with 120 epochs and a batch size of 64. Users can now train offline RL agents from scratch using the custom data collator and trainer classes provided in the accompanying notebook.