In addition to the above tests, we tested the various formations against each other, as reported in Table 3. Each entry shows the goals scored for and against when a team using one formation played against a team using another formation over the course of 24 10-minute games. The right-most column collects the total goals scored for and against the team using that formation when playing against all the other teams. In all cases, the teams used flexible positions, but no set-plays.
Table 3: Comparison of our different formations. Entries in the table
show the number of goals scored. Total (and percentage) cumulative
goals scored against all formations appear in the right-most column.
The results show that the defensive formation (8-2-0) does the best. However the total goals scored when using the defensive formation is quite low. On the other hand, the 3-3-4 formation performs well with a high goal total.
This detailed study allowed us to devise an effective formation-switching strategy for RoboCup-97. Our team [22] used a 4-4-2 formation in general, switching to a 8-2-0 formation if winning near the end of the game, or a 3-3-4 formation if losing. This strategy, along with the flexible teamwork structure as a whole, and the novel communication paradigm, helped us to perform well in the tournament, making it to the semi-finals in a field of 29 teams and out-scoring opponents by a total score of 67-9 [18].
We also used this flexible teamwork structure on our CMUnited-97 small robot team which won the RoboCup-97 small-size robot competition, out-scoring opponents by a total score of 13-1 [18]. Although developed in simulation, all of the teamwork concepts apply directly to real robot teams as well. We were able to reuse the code textually from our simulator clients on the robots and immediately achieve variable formations, flexible positions, and position switching.
Unlike the simulated agents, the robots have a global view, seeing the entire field via an overhead camera. Nonetheless, the agent control modules are distributed, enabling the use of the same team member architecture, including locker-room agreement and formation structure, as was originally developed in simulation. Our robotic system is described in detail in [28].