Tomoichi, Takahashi, AI Magazine
Analyzing Good Plays to Train Player Agents
* Teaching (training, programming, and learning) player agents is the challenging research in the RoboCup simulation games. LOGMONITOR is a tool for analyzing games from game records. LOGMONITORS can be used not only to improve the collaboration among agent players but also to evaluate them.
RoboCup provides challenging research for a multiagent system. Player agents must move autonomously in a dynamic environment. Teaching (training, programming, and learning) player agents is the main theme in simulation games.
Reviewing score books of soccer matches gives game data such as which players scored goals, which made shots, and which made passes. Score books are useful for analyzing games, ranking players, and making strategies for coming games. This situation is assumed to be similar to RoboCup simulation games. Using data in score books helps to program the synthetic player's function. Following RoboCup-97, the LOGMONITOR (Takahashi 2001) was made for automatically extracting good plays in simulation games to train player agents.
RoboCup simulation games are recorded as log files, where the positions of the ball and all players for both teams at each simulation step are written. It is necessary to recognize soccer plays such as shots or passes from the time-sequence data for checking a play's quality. The play recognition and its performance depend on the situation. Thus, it is similar to gesture understanding in a computer-human interface. For example, recognition of a pass requires that the ball move from one player to a teammate. …