Department of Electrical Engineering
Control Robotics and Machine Learning Lab
Technion - Israel Institute of Technology
המעבדה לבקרה רובוטיקה ולמידה חישובית
Advanced Approaches in Off-Policy learning
Background
Most of the recent breakthroughs in Deep Reinforcement Learning has been through advanced methods to collect and process data. This can be seen in the initial DQN which through the experience replay enabled the first DRL solution to a vast range of ATARI games. Later on, methods such as Prioritised Experience Replay, Distributed Experience Replay and more unique data collection methods such as Go-Explore - each of these methods has led to state-of-the-art performance in various games.
Project Goal
-
Implement advanced methods for experience storing and collection.
-
Assess the performance of each method.
-
Publish results.
Project steps
-
Understand the DQN framework.
-
Understand advanced experience replay methods (prioritized / distributed).
-
Implement new storage and collection methods.
-
Suggest your own ideas!
Required knowledge
-
Strong programming skills.
-
Knowledge in Deep Learning.
-
Any knowledge in RL is an advantage.
Environment
-
TensorFlow.
Comments and links
-
See the DQN paper by Google.