Summary of the Car Racing Competition held at the 2008 IEEE World Congress on Computational Intelligence, organized by Daniele Loiacono, Julian Togelius, Pier Luca Lanzi
More information available at http://cig.dei.polimi.it
1. Car racing competition(s):
lessons learned and
future directions
Julian Togelius, Daniele Loiacono, Pier Luca Lanzi
2. Car Racing Competition:
2007, 2008, 2009...?
• We want to make this a recurring event,
increasing both the quality of submitted
controllers and of the competition
organization
• Last year: used the simplerace game
(lightweight Java implementation)
• 5 entries for CIG, 12 for CEC
3. Comparing TORCS
to simplerace
• More advanced/realistic (e.g. better
dynamics and collision handling, gear shifting)
• Harder (in a sense)
• Completely deterministic (no noise)
• Slower. Much slower...
• Not completely cross-platform
• Not designed for learning algorithms
4. Not designed for
learning algorithms...
• Overhead from restarting
• Memory leak
• Not simple for client to select track
• Instant shutdown from excessive car damage
• Exploits (degenerate strategies possible)
• crossing the start line backwards!
5. However...
• All of the problems (except memory leak)
have been solved with client- or server-side
patches
• Taken together, TORCS is the best
alternative we’ve found
6. The future of the car
racing competition
• We want to make this a recurring event,
continuously improving the quality of both
competition and entries
• Next iteration confirmed for CIG 2008
• Several questions regarding in which
direction to evolve the competition...
• we want your input!
7. The future of the car
racing competition
• What can we improve?
• Measuring learning rather than design
• Accessibility and participation
• Validity and generality of results
• Dissemination
8. Measuring learning
rather than design skills
• How do we measure the power of learning
algorithms and representations rather than
the competitors’ programming skills?
• Varying the task (e.g. tracks, cars)
between training and scoring
• Automatic, track-specific learning phase
after submission
• Is this important?
9. Accessibility and
participation
• Last year we had much higher participation
second time around (same software)
• How can we make it easier to participate?
• Interfaces in more languages? (which?)
• More example trainers / controllers?
• Should we reach out to other communities?
(classical RL people, game developers etc.)
10. Validity and generality
(what can we learn?)
• That a controller based on algorithm X
wins, does not prove that algorithm X is
better than others for car racing...
• How do we improve the validity of the
competition results?
• ...it also does not prove that algorithm X is
good for any other (car control) tasks
• How do we ensure generality?
11. Dissemination
• More people will submit better controllers
if they can get a publication out of it
• Last year’s competitions became a 37-page
GPEM paper...
• Is there a better publication format? Special
issues? Workshop proceedings?