Learning Hierarchical Policies in Dynamic Environments

Advisor: Ruslan Salakhutdinov, School of Computer Science, Carnegie Mellon University

• Proposed a hierarchical RL and meta RL based framework for solving sparse rewards tasks in dynamic environments.
• The agent first learns a generic representation of a set of skills over a distribution of environments using meta learning. These skills are then fine-tuned to the given environment with a few gradient updates and a high level policy over these skills is learned for solving the required task.

The project report is available here.