The following fields, among others, will be treated:
Reinforcement learning with known and unknown models, discrete and continuous dynamic systems, Markov process formalism, Bellman optimality principle, exact and approximate algorithms, proofs of convergence, action policies, MDPs, discounted MDPs, POMDPs, reinforcement learning with temporal difference, Monte Carlo, and Q-learning.
The course also includes components, where students should prepare a lecture as well as develop a laboratory session where other students participate.
The course gives an introduction to the field reinforcement learning. The aim is that students should be acquainted with different methods that are used for learning based on feedback.
On completion of the course, you should be able to:
* identify basic concepts, terminology, theories, models and methods in reinforcement learning
* develop and systematically test a number of basic methods in reinforcement learning
* evaluate different learning algorithms experimentally and interpret and document results of experimental studies
* choose appropriate method to process automatically various types of data as e g sensor data that are used in controlling algorithms
* account for basic methods and limitations in reinforcement learning
* build a toolbox of different algorithms and be able to apply these on real problems
in order to
* be familiar with basic possibilities and limitations for reinforcement learning and thereby be able to assess which problems in e g robot movement regulation and automatic decision-making that can be solved with these technologies
* be able to implement, analyse and evaluate simple systems based on reinforcement learning
* have a broad knowledge to be able to read and profit by literature in the area.
