Modul

You will get familiar with state-of-the-art data-driven representations and algorithms for controlling stationary and mobile robots. The first part covers basic concepts of Supervised and Imitation Learning of Deep Neural Networks by means of optimization techniques. In doing so, we dedicate an entire lecture to the practical application of networks in robotics. The second part expands on various approaches to Reinforcement Learning. Accompanying the lecture, we discuss case-studies from Robotics research.

Course objectives:
- Successful participants will have a basic understanding of Machine Learning and of the mathematical optimization techniques (gradient-based methods), used in this context.
- Successful participants will understand representations (Feed-Forward and Recurrent Networks) and algorithms (Back-propagation) in Deep Supervised and Imitation Learning. They will be able to deploy them on problems which are related to learning robot behaviors.
- Successful participants will gain a comprehensive insight in the terminology of Reinforcement Learning, its stochastic foundations (MDP), model-free learning methods (MC, TD, SARSA, Q-), policy-gradient approaches (Actor-Critic, TRPO, PPO) and model-based approaches (global and local models). On that basis, successful participants can develop solutions to learn robot motor skills.

- Introduction and Foundations of Machine Learning
- Optimization for Machine Learning
- (Deep) Supervised Learning Introduction
- Guest Lecture - Innovative Practical Applications
- (Deep) Imitation Learning Introduction
- (Deep) Reinforcement Learning Introduction
- Markov Decision Processes and Dynamic Programming
- Monte-Carlo Learning and Time Difference
- Basic Policy Gradients
- Advanced Policy Gradients
- Model-based Reinforcement Learning