Deep Reinforcement Learning for Sparse-Reward Manipulation Problems

Advisor: Prof. Matt Mason, School of Computer Science, Carnegie Mellon University

• Proposed Prioritized Hindsight Experience Replay for sample efficient reinforcement learning in multi-goal manipulation environments from rewards which are sparse and binary.
• The agent assigns priorities to transitions stored in replay memory based on their temporal difference errors and performs importance sampling of the stored transitions for training its policy network.

The project report can be found here.