[rule110] { peter bentley has done it again }

  • From: "Martin Schneider" <martin@xxxxxxxxxxx>
  • To: list@xxxxxxxxxxx
  • Date: Mon, 28 Jun 2004 17:14:10 +0200 (CEST)

{{ peter bentley has done it again }}

{ editors note }

If you are interested in real life applications of artificial creativity
and  evolutionary design there is no way to get around Peter Bentley.
He started out with his PHD on GaDes - a system to evolve designs for
things like tables and racing cars back in 1996.
GaDes was rather a proof of concept, but finally evolutionary design seems
to be adopted by the industry.
Check out the WIRED-article below ...

Story location: http://www.wired.com/news/autotech/0,2554,63900,00.html


{ Breeding Race Cars to Win  }

By Michelle Delio

A technology that allows robots to rebuild themselves and computer programs
to evolve and become better on their own is now being used to breed
super-fast Formula One race cars.

F1 cars, which can hit speeds of 150 mph on a curvy track, are open-cockpit,
single-seat vehicles." Racing teams invest millions each year in each car to
install the latest technology and to fine-tune the machines' performance.

Even minor changes -- in things like the cars' wing height (F1 cars rely
heavily on aerodynamics), suspension stiffness or type of tire rubber used
on a particular day at a particular track -- can give a car a fraction of a
second's edge in speed, which often means the difference between a win and a
loss.

Formula One teams pride themselves on their mechanical tweaking skills. But
the Digital Biology Interest Group at University College London discovered
that they can boost performance by using computers to "breed" the cars.

But there was no dating, no wooing, not even a messy oil wet spot in this
survival-of-the-fastest experiment. The breeding was done solely with
computer-generated simulations using genetic algorithms -- programs that
combine Mother Nature's laws and computer science to mimic the natural
process of evolution.

Using this sort of programmed procreation, the Digital Biology Interest
Group has made self-healing battlefield surveillance robots -- gadgets that
look like robotic snakes that can figure out how to wiggle home even when
severely damaged, unlike less-evolved robots that typically just give up
when one of their critical components goes out of commission.

The group is currently working on creating crash-proof computers that could
write and repair their own operating systems and program code to suit users'
needs.

For the race-car research project, likely car designs were generated and
then tested using a racing simulation designed by Electronic Arts, with
virtual replicas of various Formula One racetracks.

The researchers configured 68 parameters in the simulation car, which
affected suspension, engine performance, tire and brake pressure, fuel
consumption and steering control.

The cars that performed especially well were treated as though they had
their own genetic code and were then bred by the computer to produce the
next generation, which combined the best features of both parent cars.
Unlike the products of more standard breeding activities, though, the worst
or weirdest features could be extracted from the genetic mix.

The process continued until eventually the ultimate Formula One vehicle
evolved, said Peter J. Bentley, leader of the University College London
digital biology group and author of popular science book Digital Biology.

Bentley said some of the cars that evolved were "clearly on the limits of
drivability -- only the computer or Michael Schumacher could have driven a
car set up in some of the solutions."

According to simulations run with the best and most drivable cars, on
various tracks, it is possible to shave 88/100th of a second per lap using
genetic algorithms to tune the cars. In an industry where 1/100th of a
second really matters, that's significant.

The most evolved car was then tested in a race against a computer-generated
default car, two cars tuned by a human racing expert, and a car designed by
a research team member. They then ran the simulation on the U.K.'s
Silverstone track.

The evolved car came in first with a time of 1:20.349 per lap. The expert
setting came second, 0.879 seconds slower. The car tuned by research team
member Krzysztof Wloch came in third with a time 1.09 seconds slower. The
default car came last, 2.42 seconds behind. In real life, the fastest lap
ever clocked at Silverstone in 2003 was 1:21.209.

While Bentley's team stands behind its research, the work hasn't been tested
in the real world. The entire process has been performed though simulations,
as the research team didn't have access to an actual Formula One car.

"Formula One teams are rather too secretive to allow us to do this kind of
work and publish it," said Bentley. "And unfortunately, the cars are too
expensive for us to borrow. We also can't afford to pay a test driver. So we
did the work on a very good software simulator. This modeled the cars and
racetracks in enormous detail, and allowed us to judge how good each
evolving solution was, simply by having the computer drive the virtual car
around a track and watching the lap time."

He said the real test would be to use the system in an actual Formula One
car.

"Using our system you could evolve the car setup while the racing is going
on. So if a car was damaged, at the next pit stop you could optimize the
settings to offset whatever has gone wrong," he said. "You could even beam
changes to the car while it is on the track, but somehow I don't think
racing authorities would go for that."


{ Peter Bentley in the fuzzylib }

Books from Peter Bentley available at the rule110 fuzzylib:

On Growth Form and Computers (Peter Bentley, Sanjeev Kumar)
Creative Evolutionary Systems (Peter Bentley, David Corne)
Evolutionary Design by Computers (Peter Bentley)

http://www.rule110.org/readingclub/fuzzylib.html
---------------------------------------------------------------------------
list archives  : //www.freelists.org/archives/rule110
change options : http://www.rule110.org/contact.html
or sign off by mailing to list-request@xxxxxxxxxxx - subject: 'unsubscribe'

Other related posts:

  • » [rule110] { peter bentley has done it again }