Learning strategies in table tennis using inverse reinforcement learning.

Area of Science:

• Robotics and Artificial Intelligence
• Human-Computer Interaction
• Game Theory

Background:

• Learning complex interactive tasks like table tennis requires both motor skills and strategic decision-making.
• Identifying and modeling strategies from demonstrations in interactive tasks remains an underexplored area.
• Existing approaches often lack the ability to capture nuanced, expert-specific strategic information.

Purpose of the Study:

• To develop a computational framework for representing and inferring strategies in table tennis.
• To enable data-driven identification of basic strategic elements guiding striking movements.
• To distinguish players based on skill level and playing style using inferred strategies.

Main Methods:

• Utilized a Markov decision process (MDP) framework to model task goals and strategic information.
• Employed model-free inverse reinforcement learning (IRL) to discover the reward function from match demonstrations.
• Validated the approach using data from players with diverse styles and playing conditions.

Main Results:

• The developed framework successfully identified fundamental elements influencing striking movement selection.
• The inferred reward function captured expert-specific strategic nuances.
• The model effectively differentiated players across various skill levels and playing styles.

Conclusions:

• The proposed computational model provides a robust method for inferring and representing strategies in complex interactive tasks.
• Inverse reinforcement learning from demonstrations is a viable approach for discovering strategic decision-making in sports like table tennis.
• This work contributes to advancing AI capabilities in understanding and replicating human strategic behavior.