next up previous contents index
Next: RoboCup-99 Up: RoboCupシミュレーションリーグの歴史 Previous: RoboCup-97   Contents   Index

RoboCup-98

The second international RoboCup championship, RoboCup-98, was held on July 2-9, 1998 in Paris, France [#!AK99!#]. It was held in conjunction with the ICMAS-98 conference.

The winning teams were entered by:

  1. Stone et al. (Carnegie Mellon University)
  2. Burkhard et al. (Humboldt University)
  3. Corten and Rondema (University of Amsterdam)
  4. Tambe et al. (ISI/University of Southern California)

Unlike in the previous year's competition, there was no team that exhibited a clear superiority in terms of low-level agent skills. Games among the top three teams were all quite closely contested with the differences being most noticeable at the strategic, team levels.

One interesting result at this competition was that the previous year's champion team competed with minimal modifications and finished roughly in the middle of the final standings. Thus, there was evidence that as a whole, the field of entries was much stronger than during the previous year: roughly half the teams could beat the previous champion.

The 1998 scientific challenge award was shared by Electro Technical Laboratory (ETL), Sony Computer Science Laboratories, Inc., and German Research Center for Artificial Intelligence GmbH (DFKI) for "development of fully automatic commentator systems for RoboCup simulator league."

To encourage the transfer of results from RoboCup to the scientific community at large, RoboCup-98 was the first to host the Multi-Agent Scientific Evaluation Session. 13 different teams participated in the session, in which their adaptability to loss of team-members was evaluated comparatively. Each team was played against the same fixed opponent (the 1997 winner, AT Humboldt'97) four half-games under official RoboCup rules. The first half-game (phase A) served as a base-line. In the other three half-games (phases B-D), 3 players were disabled incrementally: A randomly chosen player, a player chosen by the representative of the fixed opponent to maximize "damage" to the evaluated team, and the goalie. The idea is that a more adaptive team would be able to respond better to these.

Very early on, even during the session itself, it became clear that while in fact most participants agreed intuitively with the evaluation protocol, it wasn't clear how to quantitatively, or even qualitatively, analyse the data. The most obvious measure of the goal-difference at the end of each half may not be sufficient: some teams seem to do better with less players, some do worse. Performance, as measured by the goal-difference, really varied not only from team to team, but also for the same team between phases. The evaluation methodology itself and analysis of the results became open research problems in themselves. To facilitate this line of research, the data from the evaluation was made public at: http://www.isi.edu/~galk/Eval/


next up previous contents index
Next: RoboCup-99 Up: RoboCupシミュレーションリーグの歴史 Previous: RoboCup-97   Contents   Index
Hidehisa Akiyama 2004-11-21