The best offense has always been a good defense. Until now.
Thanks to a team of researchers at Disney Research, California Institute of Technology and STATS, we may soon find ourselves saying something more along the lines of, “The best offense is a defense informed by advanced machine learning technologies.”
In a paper titled, “Data Driven Ghosting using Deep Imitation Learning,” a team of researchers introduced a machine learning approach that can help sports teams better prepare for specific opponents by calculating that opponent’s most likely response to a hypothetical attack.
Yisong Yue, an assistant professor at CalTech and coauthor of the paper in question, says the approach begins by asking a simple question:
“Given that this is the location and speed and acceleration or whatever of the current players, can we predict where all these players will go?” Yue asks.
To do so the approach takes demonstrated behavioral data for every player, builds individual models of fine-grain behavior for each, uses deep imitation learning to predict a player’s most likely response to a proposed game situation and visually renders that most likely response as a “ghost.”
Thus the title of the research paper. Let’s break that down.
The colloquial definition of artificial intelligence (AI) describes any system that in some way mimics the cognitive behavior of a human being.
Automated assistants like Siri, Google Home and Alexa, Google’s AlphaGo and self-driving cars are all examples of modern AI. They behave like we would expect another human to behave.
AI is usually divided into subsets, one of which is machine learning.
Machine learning describes any system that can self-adjust during a particular process and supervised learning is in turn a subdivision of machine learning that describes any system that maps a given input to a specified output.
Basically, supervised learning is a subdivision of artificial intelligence that enables a computer to predict future events based on past events.
So you can think of this ghosting approach as a sort of supervised machine learning approach. It takes player tracking data as its input and is directed to find an output (the ideal position for the next couple of steps).
And imitation learning?
Imitation learning is simply the process of training an agent to repeat a demonstrated behavior. In this particular case the researchers sought to train an AI agent (in this case, a simple dot on a computer screen to represent a player) to behave like the sports players in the demonstrations (the tracking data).
All together now:
“At a high level, AI is simply about designing computer programs that can behave in increasingly complex ways,” Yue wrote in an email. “In this case, we want the trained AI agent to mimic the decision making of professional soccer players as represented by trajectories. Coaches do this all the time when they draw up plays on a chalkboard, and ask their players to execute certain trajectories in different situations.”
So, take a step and consider the approach as it works from beginning to end.
STATS provides Yue and his team with tracking coordinates for players across an entire season. These 2D coordinates come with human annotations that provide some context (i.e., if the player is shooting or passing the ball).
And there’s a lot of data — about 10 frames per second, per player. That means for a single play during a soccer match (a shot on goal, for instance) there could be 2,200 data points for just 10 seconds of play.
The approach then takes each data point and compares it to the relevant situation.
If a particular player tends to back away from an attacker rather than confront him nine out of ten times, the supervised learning approach will most likely determine that in the future that player will step back rather than forward when an attacker approaches.
Run that approach tens of thousands of times, take into account the movement of other players on the field and their data, combine the results and you end up with something that looks like this.
“The hope is that well over time using this algorithm and leveraging some very powerful deep learning techniques underneath I’m going to be able to learn some sort of decision making mechanisms,” said Hoang Le, a PhD candidate at CalTech and coauthor of the paper.
Those decision-making mechanisms, the digital equivalent of looking at an opponent’s defensive playbook, ultimately allow teams to better prepare for matches.
Currently this ghosting technology is limited to sports that involve little or no contact — a limitation that stems from insufficient player tracking technologies.
“In heavy contact sports where there’s a lot of collisions, the tracking technology is not as accurate,” Yue said.
There’s also the issue of an inability to implement 3D tracking, which limits ghosting technology even further.
“In basketball your body posture, your jumping ability, your height matters more. We’re not tracking that. We’re not tracking their pose. We’re just tracking their location,” Yue said. “In soccer this is less of a concern because there’s a much larger field, the game is in some sense a slower pace, it’s more based on geometry and formations.”
In other words, current ghosting technology can only determine where on the field a player will go next. It can’t determine when and how high a basketball player will jump, if a hockey player will try for a hip check or a stick check or if an offensive lineman will go on the offensive or pull back.
Nonetheless, Yue is confident that as tracking technologies advance ghosting will become more accurate and applicable to a wider range of sports and Patrick Lucey, director, data science at STATS and a coauthor of the paper, says they intend to capitalize on any and all opportunities to come.
“I can’t go into the specifics about what we are doing next at STATS in terms of our R&D,” Lucey wrote in an email. “I can say though, that we are wanting to utilize STATS’ trove of data and passion for data science helps teams find the winning edge from an analysis perspective. Where sports data was once simple event collection, now we track 6,000,000 data points per game and using machine-learning techniques – we can maximize the value of this data to help teams create a common language around analytics and ask the ‘what if questions’ needed to find and capitalize upon its unique winning edge.”
Image and thumbnail via Sporttechie.com (via Yisong Yue)