Multi-Agent Systems

Course Objective

After successfully completing this course, the student

- Has a solid understanding of concepts from elementary and intermediate
game theory, such as Nash equilibria for simultaneous games, backward
induction and subgame perfect equilibria for sequential games and the
Shapley value in cooperative games.
- Understands various principled approaches to balance exploration and
exploitation;
- Has a solid understanding of the tabular solution methods for (single
agent) reinforcement learning;
- Is able to explore and digest current research on deep reinforcement
learning and multi-agent reinforcement learning.

Dublin descriptors:
1 Knowledge and Understanding,
2. Applying Knowledge and Understanding:

Course Content

In Multi-agent systems (MAS) one studies collections of interacting,
strategic and intelligent agents.
These agents typically can sense both other agents and their
environment, reason about what they perceive, and plan and carry out
actions to achieve specific goals. In this course we introduce a number
of fundamental scientific and engineering concepts that underpin the
theoretical study of such multi-agent systems. In particular, we will
cover the following topics:

- Beliefs, desires, and intentions (BDI)
- Introduction to non-cooperative game theory
- Introduction to coalitional game theory for teams of selfish agents
- Principles of Mechanism Design
- Exploration versus Exploitation
- Markov Decision Processes
- Reinforcement learning for a single agent
- Introduction to multi-agent reinforcement learning

Additional Information Teaching Methods

Two lectures (1h45) and one recitation class (1h45) per week.

Method of Assessment

There will be weekly homework assignments that will be graded (5 of which are done in groups and 1 is done individually). In addition, there will be a final exam that will test the student's ability to apply the course material to new and concrete problems.

The final grade will be a weighted average of the grades for the homework assignments (50%) and the final exam (50%).

There is a resit offered for the final exam and the individual homework assignment. There is no resit for the group homework assignments.

Literature

Recommended reading:

Yoav Shoham, Kevin Leyton-Brown: Multiagent Systems
Publisher: Cambridge University Press (15 Dec. 2008)
ISBN-10: 0521899435
ISBN-13: 978-0521899437

R.S. Sutton, A.G. Barto, F. Bach: Reinforcement Learning: An
Introduction
Publisher: MIT Press; second edition edition (23 Nov. 2018)
Language: English
ISBN-10: 0262039249
ISBN-13: 978-0262039246

Additional Information Target Audience

Master Artificial Intelligence

Recommended background knowledge

Basic calculus, probability theory and linear algebra. Fluency in a
programming
language.