The Hanabi Challenge : From artificial teams to mixed human-machine teams
Motivation and Scope
Multi-Agent Reinforcement Learning (MARL) is becoming increasingly
popular to allow multi-agent cooperation and synchronization in various
tasks, especially multi-agent games. The Hanabi Challenge recently
proposed is an excellent benchmark to test MARL agent in a cooperative
setting with limited communication. Algorithms developed for this
challenge consider so far exclusively teams of artifical agents
evolving and learning together, exhibiiting some type of Theory of Mind
by modeling the other agent's learning process. In this project, we are
interested in evaluating whether the algorithms, trained with a team of
artificial agents, can be transferred to mixed teams, made up of humans
and artificial agents.
Specific Tasks
In this project, the task is as follows:
1. adapt the existing Hanabi Challenge benchmark and the associated GUI
to allow human players to interact with artificial agents
2. implement and train state of the art Hanabi-playing MARL algorithm.
3. validate the developped approach with human users.
The code and models developed during this thesis will be reused in future research at AASS. Appropriate acknowledgement to the student will be given in case of any publication using the developed code.
Necessary Skills
- Good programming skills and experience in Python.
- Basic knowledge of Machine Learning, if possible Reinforcement
Learning.
- Willing to learn and conduct user-experiments.
Benefits
During this project, you will learn how to work in a challenging
Reinforcement Learning problem. You will also learn how to perform
tests with human users. Depending on the project results and student's
motivation, a publication in a peer-reviewed venue is possible.