[robocup-small] Fwd: Stephan Chalup: [RoboCup-SIG-learn] Summary of the SIG meeting on Learning in Multi-Agent Systems at RoboCup 2005]

Brett Browning brettb at cs.cmu.edu
Sun Jul 24 08:41:11 EDT 2005


Hi All,

Below is the summary of the Multi-agent learning SIG.

Brett

-------- Original Message --------
Subject: Summary of the	SIG meeting on Learning in Multi-Agent Systems at RoboCup 2005
From: Peter Stone <pstone at cs.utexas.edu>


Thanks to Stephan Chalup for leading the discussion at the RoboCup SIG
on Multiagent Learning.  The meeting summary is below.

This is the last message on the topic to the general lists.  If you
have something to add, or would like to follow the discussion, please
subscribe to the mailing list at

     http://sserver.sourceforge.net/SIG-learn/

There was one other SIG meeting held, on multiagent modeling.  That
one was relatively small and focussed on possible changes to the
simulation on-line coach competition.

Cheers,
	Peter



------- Forwarded Message

To: sserver-sig-learn at lists.sourceforge.net
Subject: [RoboCup-SIG-learn] Summary of the SIG meeting on Learning in
Multi-Agent Systems	at  RoboCup 2005
Date: Sun, 24 Jul 2005 11:50:21 +1000

Summary of the SIG meeting on Learning in Multi-Agent Systems at RoboCup 2005

The 'SIG on Learning in Multi-Agent Systems' met during the RoboCup
2005 Symposium in Osaka on Monday, July 18th 2005, 5:00-6:30pm. About
30 researchers from all leagues except the Small Size League and
RoboCup Jr. attended the meeting. The discussions and comments
addressed a variety of topics with focus on the following two main
questions:

(I) Different variations of multi-agent learning have been
successfully applied at RoboCup in the simulation league (e.g. by the
Brainstormers, by Peter Stone's team, and others). How could
multi-agent learning be implemented in the Four-Legged League or other
'real robot' leagues? Can the results form the simulation league be
`transferred' to the real robot leagues? Here some outcomes of the
associated discussion:

a) The Milan RoboCup Team (MRT, Andrea Bonarini) had started to
implement a simulator for the mid size league a few years ago. There
is a publication in 2000 addressing their work. They have achieved
some interesting learning results in simulation. E.g. they found that
under some constraints the number of goals scored by a team where only
one robot was adapting its behaviour could be significantly increased
through learning. Currently MRT is about to transfer some of their
learning approaches to their real robot team.

b) Some of the rescue simulation teams reported on successful learning
results on selected subtasks (would be good to have the literature
references).

c) It was noted that some teams of the small size league achieved
learning results with multi-agent learning (e.g. Michael Bowling,
http://sserver.sourceforge.net/SIG-learn/learning-approaches.htm#SimultaneousAdversialMultirobot).

d) The NUbots presented their behaviour simulator (Stephen Young). It
was discussed how it could be used for learning tasks in the four
legged league. It was suggested to continuously update the simulator
parameters with statistical information gathered from the real
environment (e.g. ball grabbing failure rate). It was noted that other
teams have experimented with AIBO simulators as well (e.g. UChile,
German Team , etc*)

e) It was reported and discussed how team play can be achieved using
teams which are composed of robots which originate from different
four-legged league teams (Michael Quinlan). A related experiment was
conducted by the NUbots and rUNSWift in this year's Four-Legged League
Open Challenge. It was reported that similar concepts were already
implemented and tested by members of a German simulation team
(Humboldt) in about 1998/1999. However, their approach and results
were never published.

f) Peter Stone has publications and software for the keepaway tasks in
the soccer simulation league available at
http://www.cs.utexas.edu/users/AustinVilla/sim/keepaway/

Note: Some previous learning approaches are listed at
http://sserver.sourceforge.net/SIG-learn/learning-approaches.htm



(II) The second major question addressed in the SIG meeting was: What
could be a suitable learning (or adaption) challenge for the
Four-Legged League's technical challenge competition in 2006?  (This
questions was raised by Peter already some time ago:
http://sourceforge.net/mailarchive/forum.php?thread_id=3202004&forum_id=3089)

The discussion and brainstorming session at the RoboCup 2005 SIG
meeting came up with several proposals. The following list contains a
selection/summary of our discussions and is open for further
extensions and refinement:

a) Evaluate locomotion in unknown environments. For example, measure
the walking speed of the AIBO on different surfaces and when climbing
over small obstacles. The speed could be measured by an overhead
camera (although this would not be necessary in the setting we
discussed). For the challenge a race course could be set up (e.g. a
round course on the field). The robot (or the robots of several teams)
would need to race, e.g. 10 times around the course and a significant
part of their score could be how they improve between the first and
the last round (other components of the score could be the total time
they required for the task). It was noted that this would probably
better be classified as an adaptive walk challenge than a multi-agent
learning challenge.

b) Watch or observe team behaviour. Then adapt to improve your play
(this is similar to the last point of this list below).

c) Individually differently skilled players wear different
uniforms. Can the opponent teams' robots recognise the different
skills and adapt their strategy accordingly?

d) Learning within your own team. Can we develop an allround team
player?

e) Role position assignment based on observation of individual skills
and weaknesses.

f) Multi-agent team behaviour adaption: Play a 2 vs. 2 match where one
team (given by the committee) has a fixed initially unknown very
specific behaviour. Can your team of two (or four) AIBOs recognise
this and adapt accordingly to increase its scoring rate? That is, the
challenge is to perform successful behaviour adaption/learning to
improve your scoring rate against a given fixed specialist team (the
latter is kept small so that behaviour can be more clearly
identified)? The challenge could be performed on several fields in
parallel to save time. In a 5min observation run your team could
observe the fixed teams behaviour. In a second 5 min game your team
should have adapted and be able to score more goals. The goal
difference between the fist and the second half would reflect how well
your team has adapted.

Thanks to everyone who attended the meeting.

Stephan Chalup


- ------------------------------------------------------------------------------------
Stephan Chalup, Ph.D.
School of Electrical Engineering & Computer Science
The University of Newcastle
Callaghan NSW 2308, Australia
Phone: +61 2 492 16080, Fax: +49 89 1488 235410
E-mail: chalup at cs.newcastle.edu.au


- -------------------------------------------------------
SF.Net email is sponsored by: Discover Easy Linux Migration Strategies
from IBM. Find simple to follow Roadmaps, straightforward articles,
informative Webcasts and more! Get everything you need to get up to
speed, fast. http://ads.osdn.com/?ad_id=7477&alloc_id=16492&op=click
_______________________________________________
Sserver-sig-learn mailing list
Sserver-sig-learn at lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/sserver-sig-learn

------- End of Forwarded Message

_______________________________________________
robocup-legged mailing list
robocup-legged at cc.gatech.edu
https://mailman.cc.gatech.edu/mailman/listinfo/robocup-legged



-- 

-------
Brett Browning, Ph.D.
Robotics Institute, Carnegie Mellon University
phone: 412 268 6021, lab: 412 268 2601
fax: 412 268 4801, web: http://www.cs.cmu.edu/~brettb



More information about the robocup-small mailing list