About multi-task reinforcement learning. #11
Description
Thank you for your work on single-task reinforcement and transfer learning!
But I think the most meaningful part of this paper is transfer learning on multi-task reinforcement learning.
If the teacher models are in different envs(for example, one is in CartPole-v1,others is in Acrobot-v1), the state space and action space of them are totally different. It is necessary to pay attention to how to design the input layer and output layer of the student model.
"About 90% of parameters are shared, with only 3 small MLP “controllers” on top which are task specific and allow for different action sets between different games." It's in the paper. But I don't know the details