A deep dive into reinforcement learning

Reinforcement learning concerns a distinct branch in the field of machine learning. While supervised and unsupervised machine learning techniques learn based on a set of instances, reinforcement learning concerns an agent that learns by interacting with an environment. To illustrate this setting, consider a video game, where the agent identifies as the player of the game. The environment concerns the video game that can be controlled with a gamepad and observed by looking at the screen. The agent receives feedback from the environment in terms of rewards, for example via a game score. Based on this feedback and the observations of the environment, a reinforcement learning agent aims at learning to play the game optimally, by interacting repeatedly with the environment. In this course, we will study both the reinforcement learning problem and the algorithms that can be used to approach such problems.

Over the last years, several important milestones were achieved using the reinforcement learning framework, such as learning to play ATARI 2600 video games, learning to play the board game Go to beat the number one human player and learning to reach grandmaster performance in the real-time strategy game Starcraft II. However, reinforcement learning is no longer limited to games, as many real-world problems can be formulated as a reinforcement learning problem. Recent studies used reinforcement learning to solve challenging problems such as wind farm control, epidemic mitigation, and robot control.

In this course, we will convey the fundamentals that underlie reinforcement learning. To achieve this, we will start from intuition and applications to move towards formalisms that can implement a reinforcement learning agent. To move from formalisms to practice, we will explore these concepts in a grid-world environment using the python programming language.

• 10:00-12:00 Lecture
• 13:00-15:00 Practical Session