Goals:
- To acquire an adequate impression of state-of-the-art research in multi-agent learning, including topics such as learning and teaching, fictitious play, rational learning, no-regret learning, targeted learning, multi-agent reinforcement learning, evolutionary learning.
- To write a short and organized scientific paper summaries on specific topics in MAL
- To be able to deal with technical questions in MAL and be able to formulate written answers
- To write a small thesis on a specialized topic in MAL, and present it to an audience
- To participate in a scientific discussion, according to social scientific standards.
Assessment
Four take home assignments and two exams.
A retake requires at least a 4 for the original test.
Pre-requisites
The course assumes knowledge of probability theory and game theory.
Knowledge of game theory can be acquired through the period 3 multi-agent systems course INFOMAS.
|
|
This course discusses algorithms for machine learning that typically occur in multi-agent systems, such as: reinforcement learning, no-regret learning, fictitious play, satisficing play, Bayesian learning, learning and teaching, and evolutionary learning (replicator dynamic).
Multi-agent learning (MAL) studies agents that learn and adapt to the behavior of other agents that themselves learn and adapt.
The presence of other learning agents complicates learning, which makes the environment non-stationary (a situation of learning a moving target) and non-Markovian (a situation where not only experiences from the immediate past but also earlier experiences are relevant).
It becomes less beneficent to only adapt to the behavior of other agents, on the pain of being exploited by more steadfast agents that do not follow but instead impose their strategy on others.
Important topics of MAL include (evolutionary) game theory, fictitious play, gradient dynamics, no-regret learning, multi-agent reinforcement learning (MinMax-Q, Nash-Q), leader (teacher) vs. follower (learner) adaptation, and the emergence of social conventions.
Examples of domains that need robust MAL algorithms include manufacturing systems (managers of a factory coordinate to maximize their profit), distributed sensor networks (multiple sensors collaborate to perform a large-scale sensing task under strict power constraints), robo-soccer, disaster rescue (robots must safely find victims as fast as possible after an earthquake) and recreational games of imperfect information such as poker. Indeed, poker and simplified forms of poker are an important topics of research in multi-agent learning.
Course form
Lectures.
Literature
Available on the course site.
|
 |
|