Synthesizing NBA defenses with deep imitation learning
Encoding and/or constructing generative models of collective behavior is a complex task due to the dynamic nature of the couplings between agents. The natural state space of endogenous factors, such as group topology, and exogenous factors, such as the presence and position of other objects, agents, and signals, is far too high dimensional to be experimentally probed within a laboratory setting. New techniques in reinforcement and imitation learning, however, provide an avenue for progress using large quantities of trajectories from observational video data.
In this talk, we will examine these challenges within the context of professional basketball. The goal is to construct unsupervised models which are capable of synthesizing realistic, responsive NBA team defensive behaviors. We will present compelling progress towards this goal and discuss the key technical insights that have made this possible. Specifically, we will focus on issues of data routing, feature selection, network architecture, and multi-model training.
Speaker Bio:
Andrew Hartnett is a physicist, ecologist, and educator. His research interests center on extending principles from information theory and machine learning to problems in collective behavior. He received his Ph.D. in 2017 from Princeton University where he studied the mechanisms of coordinated movement and consensus decision-making in animal groups. As a postdoc at Disney Research, Andrew focused on understanding collective behavior in team sports: developing deep recurrent models for encoding and predicting player trajectories in basketball. Currently, he is working on understanding and modeling the complex interactions encountered by autonomous vehicles as an engineer at Argo AI.
Date and Time
Location
Hosts
Registration
- Date: 18 Apr 2018
- Time: 06:30 PM to 08:00 PM
- All times are (GMT-05:00) US/Eastern
- Add Event to Calendar
- Starts 13 March 2018 07:30 PM
- Ends 17 April 2018 05:00 PM
- All times are (GMT-05:00) US/Eastern
- No Admission Charge