next up previous
Next: Conclusion Up: The CMUnited-98 Champion Simulator Previous: Ball Handling



In order to test individual components of the CMUnited-98 team, it is best to compile performance results for the team with and without these components as we have done elsewhere [10]. However, competition against other, independently-created teams is useful for evaluating the system as a whole.

At the RoboCup-98 competition, CMUnited-98 won all 8 of its games by a combined score of 66-0, finishing first place in a field of 34 teams. Table 1 details the game results.

Table 1: The scores of CMUnited-98's games in the simulator league of RoboCup-98. CMUnited-98 won all 8 games, finishing in 1st place out of 34 teams.

From observing the games, it was apparent that the CMUnited-98 low-level skills were superior in the first 6 games: CMUnited-98 agents were able to dribble around opponents, had many scoring opportunities, and suffered few shots against.

However, in the last 2 games, the CMUnited-98 strategic formations, communication, and ball-handling routines were put more to the test as the Windmill Wanderers (3rd place) and AT-Humboldt'98 (2nd place) also had similar low-level capabilities. In these games, CMUnited-98's abilities to use set-plays to clear the ball from its defensive zone, to get past the opponents' offsides traps, and to maintain a cohesive defensive unit became very apparent. Many of the goals scored by CMUnited-98 were a direct result of the opponent team being unable to clear the ball from its own end after a goal kick: a CMUnited-98 player would intercept the clearing pass and quickly shoot it into the goal. In particular, two of the goals in the final game against AT-Humboldt'98 were scored in this manner. On the other hand, the CMUnited-98 simulator team was able to clear the ball successfully from its own zone using its ability to execute set-plays, or pre-compiled multi-agent plans. Rather than kicking the ball up the middle of the field, one player would pass out to the sideline to a second player that would then clear the ball up the field. After a series of 3 or 4 passes, the ball was usually safely in the other half of the field.

Another strategic advantage that was clear throughout CMUnited-98's games was the players' abilities to maintain a coherent defensive unit exploiting the offsides rule, and conversely, its ability to get through the defense of other teams. Often, the opposing teams were unable to get anywhere near the CMUnited-98 goal because of the defenders' ability to stay in front of some of the opposing forwards, thus rendering them offsides and prohibiting them from ever successfully receiving the ball.

In order to deal with opposing teams that tried to use a similar technique, the CMUnited-98 forwards would kick the ball towards the offensive corners of the field (the ``sending'' skill described in Section 4.7) and then either get to the ball before the defenders or intercept defenders' clearing passes. CMUnited-98 scored several nice goals after such kicks to the corners.

In addition to the strategic reasoning that helped the team win its final two games, the fine points of the dribbling and goaltending skills also came into play. Using their predictive, locally optimal skills (PLOS--see Section 4), the CMUnited-98 players were occasionally able to dribble around opponents for shots. At a crucial moment against the Windmill Wanderers, the CMUnited-98 goaltender made a particularly important save: winning 1-0 near the end of the game, a shot got past the goaltender, but it was able to turn and catch the ball before the ball entered the goal.

next up previous
Next: Conclusion Up: The CMUnited-98 Champion Simulator Previous: Ball Handling

Peter Stone
Mon Nov 30 20:08:29 EST 1998