Home |  DEUTSCH |  Contact |  Imprint |  Login |  KIT

Inproceedings3384

Aus Aifbportal

Wechseln zu: Navigation, Suche

(This page contains COinS metadata)

Learning from recorded games: A scoring policy for simulated soccer agents




Published: 2004

Buchtitel: Proceedings of the ECAI 2004, 16th European Conference on Artificial Intelligence, Workshop 8: Agents in dynamic and real-time environments
Seiten: 43-48
Verlag: IOS Press
Referierte Veröffentlichung
BibTeX

Kurzfassung
This paper outlines the implementation of a new scoring policy for the agents of the Simulated Robot Soccer team from the University of Koblenz, called RoboLog. The applied technique is capable of acting in real time in the dynamic environment of the RoboCup Simulation League and uses data obtained from prerecorded soccer games for supervised neural network learning. The benchmark used for testing this approach is the Optimal Scoring Problem stated as finding the point in the goal where the probability of scoring is the highest when the ball is shot to this point in a given situation. Goalshot situations from numerous logfiles are extracted and employed for the training of two independent multi layered perceptrons. Beside the usage as training patterns the gained data is evaluated statistically and provides interesting general insights into goalshots carried out lately in Simulated Robot Soccer. The results obtained after extensive testing of the new policy are presented. Furthermore, general issues of learning from observed logfile data and starting points for future work are discussed.

Download: Media:Rettinger04-Learning_from_Recorded_Games.pdf
Weitere Informationen unter: Link



Forschungsgruppe

Wissensmanagement