ABOUT THE PROJECT
Vezhnevets et al. recently introduced the Feudal Network, a Hierarchical Reinforcement Learning architecture capable of learning options without the manual specification of subtasks. The approach represents sub-goals as changes to a learned state representation, and in doing so converts the traditionally challenging task of sub-goal discovery into a representation learning problem that can be solved effectively using neural networks. In this project, we implemented the Feudal Network and tested its performance against competitive baseline agents.
This project is meant to serve as an open sourced implementation of a complex Machine Learning solution. By giving people access to cutting edge models, we hope to accelerate the pace of development with Deep Hierarchical Learning.
Note this project is still under active development, updates to come soon!
This project was developed for CS234 at Stanford University.