Department of Electrical Engineering
Control Robotics and Machine Learning Lab
Technion - Israel Institute of Technology
המעבדה לבקרה רובוטיקה ולמידה חישובית
Adaptive Methods for Continuous Control
Background
In robotics tasks, the actions that are performed at each timestep are complex high dimensional control commands. E.g., the amount of force (normalized to [-1,1]) to apply for each motor.
While there are algorithms capable of coping in such a setting, the complexity of the action space results in instability and longer training procedures.
Project Goal
Extend the Q estimation to better cope with continuous control tasks. Specifically, you will try to estimate the return using polynomials / adaptive discretization methods, such that finding the optimal action is possible analytically.
Project steps
-
Understand reinforcement learning and deep learning methodologies
-
Learn the basics of continuous control
-
Implement basic methods
-
Research and improve the method, compare to baselines
Required knowledge
-
Strong programming skills.
-
Any knowledge in DL and RL is an advantage.
-
Any knowledge in optimization is an advantage.
Environment
-
PyTorch (Python)
Comments and links
-
See the papers "Deep Deterministic Policy Gradients" and "Twin Delayed DDPG".
-
Potential publication if results are satisfying.
