RoboCup Special Interest Group (SIG) on
Multiagent Learning
SIG on Multiagent Learning
The primary aim of this SIG is to use the RoboCup domain for
focussed investigations of algorithms and techniques for collaborative and
adversarial learning. It builds upon the IJCAI-97 synthetic agents learning
challenge by expanding it to include multiagent learning challenges in all the
different RoboCup leagues and by bringing together researchers with this common
interest in rapid, focussed discussions.
Specific goals include:
- the creation of benchmark
learning tasks for comparative studies within the different RoboCup
leagues.
- the definition of
mechanisms for transferring and sharing results and techniques across the
different RoboCup leagues, and for studying the degree to which
implemented techniques generalize across the different leagues.
- creation of subtasks
designed to encourage people who have not previously been involved in
RoboCup to test their multiagent learning techniques without having to
create fully-functional, competitive teams. For example:
- a team could learn
only how to play offensively or defensively, rather than needing to
balance the two goals.
- 2 or three players
could learn to score or defend against 1 or 2 fixed opponents.
- players could learn
to keep possession of the ball, without trying to score.
- players could learn
to clear the ball from their defensive zone, without trying to score.
In the simulator, the SIG will aim to support a modified
version of the simulator and publicly available clients to implement such
subtasks. Milestones could then be concretely specified in terms of offensive
or defensive learning-based performance against a fixed opponent.
A specific challenge in the real-robot leagues is to create and test
hardware-independent learning algorithms that could be used by more than one
team. The SIG will encourage hardware-based teams to pair up with other
hardware or software-based teams to investigate common learning algorithms on
their different platforms.
Organizing committee members:
- Peter Stone (Chair),
University of Texas at Austin, pstone@cs.utexas.edu
- Minoru Asada
Osaka University, asada@ams.eng.osaka-u.ac.jp
- Michael Bowling Carnegie
Mellon University, mhb@cs.cmu.edu
- Bernhard
Hengst University of New South Wales, Sydney, Australia bernhardh@cse.unsw.edu.au
- NODA, Itsuki National
Institute of Advanced Industrial Science and Technology, Japan I.Noda@aist.go.jp
- Martin Riedmiller
University of Karlsruhe, riedml@ira.uka.de
- Richard Sutton
AT&T Labs - Research, sutton@research.att.com
- Manuela Veloso Carnegie
Mellon Univeristy, mmv@cs.cmu.edu
- Pejman Iravani Open University, p.iravani@open.ac.uk
General Goals:
RoboCup SIGs are being formed with the following goals:
- to create a forum for
discussion among researchers with common interests ACROSS THE DIFFERENT
RoboCup LEAGUES and from OUTSIDE OF THE CURRENT RoboCup COMMUNITY.
- to define research
milestones so that competition rules and research efforts can be targeted
based on the milestones.
- to make software and
hardware resources available, so that researchers can use such assets for
their own research, particularly to start up new RoboCup teams and
research programs.
- to promote research is
specific area by organizing mailing lists, web pages, and workshops
UPDATED Learning Approaches
List of machine learning
approaches to RoboCup.
Learning
Benchmarks
Here you can find the proposed
benchmarks for the Multiagent Learning Task.
Mailing List:
If you're interested, please join the
mailing list.
Other RoboCup SIGs:
The complete list of RoboCup SIGs and instructions for how
to propose a new RoboCup SIG are available from the main RoboCup SIG website.
Peter Stone
Go back to my homepage
Send me mail