Multi-agent Learning Seminar

Professor Ann Nowé

Assistant Peter Vrancx

Assistant Tim Brys

Tuesday, 13h CoMo Lab (10G711)

Tuesday 27/09/2011: Short introduction to the course
Bi-weekly meetings Tuesday 13-16h.

This year's course will deal with Robotic Soccer. Robocup is an international competition aimed at developing autonomous robot soccer players. Teams from across the globe compete in a number of different leagues (humanoid, small scale, simulation,…) in order to develop the best robotic soccer team. The CoMo lab currently has contacts with the Tech United Robocup team from Eindhoven.

Image source: Wikipedia

In the multi-agent learning seminar we will examine some of the AI challenges involved in developing Robocup players. Students will also have the opportunity to develop and evaluate strategies in a simulated robosoccer environment. If a suitable strategy results, the possibility exists to try the algorithm on a real robot team at the University of Eindhoven.

• the official RoboCup page
• SoccerBots. A High level java simulation of robot soccer.
• Video of Robocup AIBO match.
• PhD thesis on learning in RoboSoccer.
• Layered Learning in RoboSoccer
• Wright Eagle Simulation Team
• Tech United. The RoboCup team of the Technische Universiteit Eindhoven (2nd place RoboCup 2010).
• CoMo visit to Tech United (MAL Seminar 2010)
• Also see the CoMo Robotics Wiki

Course material will be posted on the PointCarre page of the course.

During the second semester students will work on a project which will count for the final evaluation of the project. The idea is to perform your own robosoccer experiment in the official simulator. Projects should be completed in groups of 3 students. Each group should select a research paper on robotic soccer to serve as a starting point for the experiment. This paper will be presented during a seminar session at the start of the second semester. Note that the paper is meant to be the inspiration for the project experiment, but you are allowed to define your own experiments and you are not required to exactly replicate the research in the paper. It is recommended to start from an existing robocup base team (see e.g. Wright Eagle team on the wiki) to implement your experiments. The results of your experiments should be described in a short paper.

• 20/12/2011: Introduction to the simulator.
• 14/1/2012: Mail paper and group members to the assistants.
• 27/02/2012: Each group presents selected paper and proposed experiments.
• 25/05/2012: Send paper and team code to the assistants.
• During final exam session: Present paper and experimental results.

Project groups and topics:

• Wolney de Mello Neto and Thiago Mendonça: Reinforcement learning for RoboCup-soccer keepaway. Application of Semi-Markov Decision Process Sarsa(lambda) to the KeepAway strategy.
• Mikael Lenaertz, Thierry Janssen and Guy Nagels: Learning Decision Trees for Action Selection in Soccer Agents
• Daniel Silva Chaltein de Almeida, Joseph Chege and Nguyen Cu: An Architecture for Action Selection in Robotic Soccer. Analyze passing to positions, not teammates.