-
Language:Italian language. English slides and textbook.
-
Textbooks:-) Textbook: "Reinforcement Learning: An Introduction", Sutton-Barto, free download at incompleteideas.net/book/the-book-2nd.html.
-) Course slides. -
Learning objectives:Introduction to the basic RL principles, with a particular emphasis to their applications to combinatorial games.
LEARNING OUTCOMES
KNOWLEDGE AND UNDERSTANDING
At the end of the course the student should:
-) understand agent-environment interaction in MDP;
-) recognize the main differences among different RL principles;
-) know the most important RL algorithms.
APPLYING KNOWLEDGE AND UNDERSTANDING
At the end of the course the student should be able to:
-) understand whether a certain problem is well-suited for RL;
-) model a decision task as MDP;
-) work in the model-free case with both MC and TD methods;
-) implement from scratch a RL pipeline able to learn a simple combinatorial game.
COMMUNICATION SKILLS:
At the end of the course the student should be able to communicate RL concepts with a proper and sound language.
LEARNING SKILLS:
At the end of the course the student should be able to read and partially understand textbooks and research papers on RL. -
Prerequisite:None.
-
Teaching methods:Lectures.
-
Exam type:Final written exam, with optional additional oral communication.
-
Sostenibilità:It does not address issues related to environmental sustainability.
-
Further information:E-mail: parton@unich.it.
Mobile phone: 349-5323-199.
This course aims to describe some of the recent progresses made in AI thanks to Deep Reinforcement Learning techniques.
We will learn - time permitting - to describe real-life problems as Markov Decision Processes (MDP), and to deal with them using dynamic programming or Reinforcement Learning, according whether a full distribution model, or only real experience or a sample model, is available.
1) The reinforcement learning problem: interation between agent-environment.
2) The reinforcement learning setup: Markov decision processes.
3) Prediction and control in dynamic programming.
4) Prediction and control in the model-free case: TD and MC methods.
5) Prediction and control with approximation.
6) Policy gradient methods.
7) Exploration versus exploitation: multi-armed bandits.
8) Reinforcement learning in perfect information two-player zero-sum games.
Scopri cosa vuol dire essere dell'Ud'A
SEDE DI CHIETI
Via dei Vestini,31
Centralino 0871.3551
SEDE DI PESCARA
Viale Pindaro,42
Centralino 085.45371
email: info@unich.it
PEC: ateneo@pec.unich.it
Partita IVA 01335970693
