Modeling Team Strategies using Deep Imitation Learning
Current state-of-the-art sports statistics compare players and teams to league average performance, such as “Expected Point Value” (EPV) in basketball. These measures have enhanced our ability to analyze, compare and value performance in sport. But they are inherently limited because they are tied to a discrete outcome of a specific event. For example, EPV for basketball focuses on estimating the probability of a player making a shot based on the current situation. In this work, we explore how teams control time and space by examining sequential decision making.
We have developed an automatic "ghosting" system which illustrates where defensive players should have been (instead of where they actually were) based on the locations of the opposition players and ball. We employ a machine learning technique called deep imitation learning, and modify standard recurrent neural network training to consider both instantaneous and future losses, which enables ghosted players to anticipate movements of their teammates and the opposition. Our approach avoids the man-years of manual annotation need to train existing ghosting systems, and can be fine tuned to mimic the behavior of specific teams or playing styles.
Peter Carr is a Research Scientist at Disney Research, Pittsburgh. His research interests lie at the intersection of computer vision, machine learning and robotics. In particular, he has focused on computer vision algorithms for camera calibration and object tracking, as well as machine learning techniques for understanding spatio-temporal trajectory data. Peter joined Disney Research in 2010 after receiving his PhD from the Australian National University.