|
This,
ultimately, is the goal of the simulated matches - the development
of algorithms that can be ported into physical robots. RoboCup
president Hiroaki Kitano hopes that a team of such robots
will beat a team of humans by the year 2050. This apparently
frivolous goal belies the deeper intent of the RoboCup researchers,
since the same cooperative strategies which help robots win
a soccer game can have wide applications beyond the match.
For example, RoboCup includes "robot rescue competitions,"
where the robots are taken into simulated disaster areas and
required to seek out dummies or put out mock fires. Building
computers that can beat a chess master at his own game has
motivated startling advances in artificial intelligence. We
can only guess what advances in robotics and A.I. will result
from moving human/machine competition to the physical realm.
Says Stone, "…(beating a human team) is a monumental task.
If we can get robots to do it, we can get them to do a lot
of other things as well."
View
some actual RoboCup matches!
The
matches from the 1999 and 2001 championships clearly demonstrate
how the strategies of the simulated RoboCup teams evolve over
time. "From year to year, the games look more and more like
real soccer," notes Stone.
An
example of this realism occurs seconds into the 1999 match
between Carnegie-Mellon University's CMUnited (red) and Albert-Ludwigs-University's
MagmaFreiburg (blue). (Watch
clip - 540k) The ball goes out of bounds, and after
bringing the ball in, red nearly scores a goal based on a
pre-determined set play (Stone's "locker-room-agreement").
Red then quickly performs a similar play on the other side
of the field, this time making the goal. As the score increases
for CMUnited, the predetermined strategy shifts to more defensive
play. No communication between the players is necessary to
organize this shift, which is a pure result of the locker-room
agreement. In fact, relying on communication would be a hindrance,
since the realistic constraint of limited information exchange
would slow down the play. The 1999 match illustrates a highly
successful application of the locker-room agreement approach,
as CMUnited wins the match 7-0.
By
the 2001 championship match between Tsinghua University's
Tsinghuaeolus (blue) and Karlsruhe University's Brainstormers
(red), play has become even more advanced, as the teams improve
the code. In 1999 and before, typical play is dominated by
dribbling straight up the middle towards the goalie, but by
2001 we see more complex stategy - frequent passing to the
wings and crossing into the center.
One
clip in particular provides a good example of this general shift
in strategy. "Starting at cycle 592, Tsinghua players 10 and
11 pass back and forth. (Watch
clip - 520k) At cycle 645, player
10 appears to have a breakaway but nonetheless passes to player
11 on the wing for a cross to the center," Stone points out.
An
effective trick used in this match by the Brainstormers has
since been made illegal for simulated RoboCup matches, just
like in real soccer. When red is on the defensive and a player
gets the ball, under the right conditions the player kills
the play by passing back to its goalie. A good example of
this tactic can be seen around cycle 1600, (Watch
clip - 460k) following an extended attack by blue.
Less obvious is the new approach Karlsruhe takes to the artificial
intelligence of its players. The Brainstormers use a reinforcement-learning
algorithm whereby the players learn from their successes,
and adjust their behavior accordingly. Since the effects of
any action may not be immediately apparent, the players have
to figure out exactly which combination of actions lead to
a beneficial result. The enhancements of both teams make them
quite evenly matched. In a close game, they head into overtime,
where Tsinghua finally wins 1-0. 
-
- - - - - - - - - - -
2
pages: | 1 | 2 |
Diagram: John Sear
|
