Swedish web site
Reinforcement Learning2026VT, 6.0 credits
|
|
Course plan
No of lectures
9 lectures are planned. This may change.
Recommended for
PhD students interested in reinforcement learning, learning how to act.
The course was last given
VT2024.
Goals
At the end of the course the students should have a broad and deep knowledge of reinforcement learning including value-based methods, policy-based methods, model-based methods, actor-critic methods, multi-agent RL, multi-objective RL and hierarchical RL. They should also be able to implement common methods and understand research papers in the field.
Prerequisites
Practical skills of programming in Python.
Working knowledge of deep learning, especially deep neural networks.
Understanding of common concepts in computer science and optimization.
Organization
Lectures that present the material and labs where the students learn about the algorithms and experiment with the methods.
Content
See lectures.
Literature
Deep Reinforcement Learning by Aske Plaat, 2022 ISBN 9789811906374 also available online
Lectures
1. Introduction to reinforcement learning
2. Value-based RL methods
3. Policy-based RL methods
4. Model-based RL
5. Actor-Critic methods
6. Multi-agent RL
7. Multi-objective RL
8. Hierarchical RL
9. Inverse RL
Examination
4 Labs.
Lab 1: Implement and compare tabular Q-learning and SARSA on at least two
environments in the OpenAI Gym.
Lab 2: Implement and compare a deep value-based, e.g. rainbow DQN, and a deep
policy-based approach, e.g. REINFORCE on at least two environments in the
OpenAI Gym.
Lab 3: Implement and evaluate an actor-critic method such as TRPO, PPO or SAC
in at least two different environments in the OpenAI Gym.
Lab 4: Implement and evaluate one of the more advanced RL methods as presented
in lecture 6-9 in at least two environments in the OpenAI Gym.
Examiner
Fredrik Heintz
Credits
6 ECTS
Comments
Page responsible: Anne Moe
