Inverse reinforcement learning for strategy extraction
(Inverses Verstärkungslernen für die Strategieentwicklung)
In competitive motor tasks such as table tennis, mastering the task is not merely a matter of perfect execution of a specific movement pattern. Here, a higher-level strategy is required in order to win the game. The data-driven identijcation of basic strategies in interactive tasks, such as table tennis is a largely unexplored problem. In order to automatically extract expert knowledge on effective strategic elements from table tennis data, we model the game as a Markov decision problem, where the reward function models the goal of the task as well as all strategic information. We collect data from players with different playing skills and styles using a motion capture system and infer the reward function using inverse reinforcement learning. We show that the resulting reward functions are able to distinguish the expert among players with different skill levels as well as di erent playing styles.
© Copyright 2013 Machine Learning and Data Mining for Sports Analytics ECML/PKDD 2013 workshop. Veröffentlicht von Department of Computer Science, KU Leuven. Alle Rechte vorbehalten.
Schlagworte: | Wettkampf Tischtennis Fertigkeit Taktik motorisches Lernen Bewertung Motion Capturing |
---|---|
Notationen: | Naturwissenschaften und Technik Trainingswissenschaft Spielsportarten |
Tagging: | Markov Ketten Bewegungsmuster Strategie Spielweise |
Veröffentlicht in: | Machine Learning and Data Mining for Sports Analytics ECML/PKDD 2013 workshop |
Herausgeber: | A. Zimmermann, J. Davis, J. Van Haaren |
Veröffentlicht: |
Leuven
Department of Computer Science, KU Leuven
2013
|
Dokumentenarten: | Artikel |
Sprache: | Englisch |
Level: | hoch |