Course Contents:

Machine learning methods are presented which allow artificial systems to learn successful action policies. The artificial agent could be a robot, a character in a computer game or an Internet browser. In general there is no teacher available, who could tell the agent which action would be optimal in a given situation. Instead, the agent just gets occasional "rewards" or "punishments", and has to find out on his own how much each action of a sequence contributed to a reward. From this information the agent has to develop efficient strategies for future tasks. Applications will be shown that mainly demonstrate learning algorithms for humanoid robots (on which the institute is currently working in a joint research project together with EPFL Lausanne) and for quasi-commercial computer games.

Reinforcement Learning algorithms (http://reinforcementlearning.ai-depot.com/Main.html) have been particularly successful for solving problems of this kind. Therefore we will concentrate on this learning approach during the lecture and discuss both the theoretical background (dynamic programming, Markov decision processes) and applications.


Discussed Topics:

Genetic Algorithms

In this lecture we will also cover genetic algorithms (often called evolutionary algorithms, see http://www.aic.nrl.navy.mil/galist/), which is another interesting approach to machine learning of successful policies. Here the computer simulates evolution by randomly mutating and crossing-over different promising strategies. The "fittest" of the newly generated policies are selected and evolution proceeds on this new population.


Discussed Topics