angelopetraglia
07-10-2016, 08:23 AM
Article from The Australian. Interesting insight into how much technology is impacting the analysis and decisions in the game.
Western Bulldogs gain from artificial intelligence
A Victoria University student analyses player data during a Dogs training session. Picture: Darrian Traynor.
Did artificial intelligence help the Western Bulldogs win the AFL flag? The Sons of the West certainly played a pack-centred style of football, in contrast with the more flowing play of opposing teams.
On grand final day they excelled at catapulting the ball out of packs like there was no tomorrow. After two or three lightning handpasses, the ball would be with a Bulldog in the open, and on the way to a goal. Coach Luke Beveridge can take credit for forming a “handball club” to teach players to punch ballistic missiles. The Bulldogs have been the No 1 handball*!ing side out of stoppages.
But don’t discount the role of computer machine learning — a form of artificial intelligence — in helping shape the Doggies’ game play. They are among a few AFL clubs going hi-tech to gain an edge.
Luke Dahlhaus of the Bulldogs handballs while being tackled during the 2016 AFL Grand Final
Sam Robertson, a specialist in machine learning and data analytics, and a senior research fellow at Victoria University, is a key member of senior Bulldogs staff. He says he has performed machine learning analyses of player positions around a pack.
“Absolutely, yes. If we want to take some x, y and time co-ordinates on each of our players around a stoppage, we can do that, and we can model that into some kind of situation where we match an ideal shape or ideal pattern, something we call the dominant region. We have an idea of what we want the stoppage to look like, we obviously try to get dominance around that area. Where the machine learning comes in, instead of it being one size fits all, it’s almost a what-if analysis: if the opposition (does) this, then how do we counteract that so we maintain our stoppage dominance?
Western Bulldogs sports scientist Dr Sam Robertson.
“We’re also making sure we have an understanding of what that does to the rest of the ground. If you move a player or change the set-up in one place on the ground it can also (have an) influence down the field. The machine learning side is the flexibility in decision-making.”
Machine learning is about computers making sense of the screeds of data that can flow from matches and training sessions. In AFL game play, that could mean analysing the movements of players in a match and simulating the outcome if they had acted differently. The Bulldogs can analyse who is owning different parts of the ground based on their location — a concept they’ve borrowed from soccer. Another is to measure congestion.
At the moment, Western Bulldogs are limited by not having the data of the opposition. “We know what we’re doing from a tracking player perspective, and we can then model that, but to obtain the information on the opposition is more difficult,” Robertson says.
Player position data used by The Western Bulldogs
For about five years the Bulldogs and other club players have worn a tracking device with a GPS or local positioning system for Melbourne’s Etihad Stadium, accelerometer and gyroscope sewn into the top rear section of their jumpers. Data can be collected about the location of every player at every moment of a match, in the way a Fitbit or Garmin wearable device can track exercise.
Fast computers can merge each player’s data into a single time and motion study of the entire match. It can identify position, velocity and acceleration at any moment. Staff then can look at alternative game plays and model outcomes. By linking two players they can observe tagging and how well their movement is coupled.
Analysis sometime involves small bits of game play. “We will look at one-minute, two-minute and five-minute increments, and to go back and create a retrospective model.”
So far Robertson has offered this analysis after games on a Monday. “The next stage is to use that in the coach’s box. It’s something we’ll look to implement in 2017. Even if that model adds another 5 per cent of understanding that the coaches don’t have, it’s going to be useful for us in a live sense as well.”
Norm Smith medallist Jason Johannisen shows what the Bulldogs ‘handball club’ does in a contest.
Game play is only one area where Robertson values machine learning. He has used it to look at the types of players on the Bulldogs list who may have deficiencies, what combination of players work best together and what types of players they need from the draft.
Player maintenance, injury management and player game day selection are other applications.
“Player selection is a really big one, we’re doing a lot of work in it. I’m talking from draft selection to scouting to team selection, we’ve got a dedicated PhD student in operations research at the moment, he’s working solely on player evaluation with a focus on player selection,” Robertson says.
Without detailed data on opponents, he has modelled the Bulldogs’ opposition based on general attributes. For example, he regards Swans and Giants as among sides that move the ball in a certain way.
Predicting injury, monitoring recovery and assessing the risk of fielding a player is “the buzz area” and machine learning offers a way to model player risk based on history. Robertson says he has accumulated two years of data on players.
He says a specific model is needed for each player that takes into account family history, say in terms of susceptibility to knee injuries. “If your model says this athlete has a one in three chance of being injured in this session, you have to make a decision from that,” he says.
As to the future, he says the ability of machines to ingest and analyse a match video will have a huge impact. Quantifying exactly what every player is doing, including how many sprints and jumps they do in a match, is beyond what current sensors offer. Video also could automate keeping stats of when players kick or mark the ball, handball and tackle.
Fans & media might get access to player movement data, says Robertson.
Alternatively, there are tattoo-like ankle sensors that could measure kicking. “I think video is a more realistic output, but it’s very hard to set up a permanent camera system for Australian football because of the size of our grounds,” Robertson says.
He believes machine learning systems will become more powerful when teams begin to share match day data. Champion Data, which provides stats for the AFL, and Catapult, which provides tracking sensors, have proposed data sharing to the 18 clubs.
“We could do thousands of things which we can’t do now,” he says. It’s being discussed, and he expects clubs to agree to it by next year. “They’ve done it in the NBA and they’ve done it in soccer.”
Across time he expects game data will be available to the media and, eventually, fans will gain access to it. “There’s a whole community of amateur analysts out there in football,” he says.
Robertson isn’t overstating the role machine learning played in the Bulldogs’ eventual victory, although it contributed to honing the team’s style of football. He cites the passion and remarkable commitment of Beveridge and the team, their determination and hunger for victory, and the return of key players from injury.
It’s early days yet for machine learning in AFL, but within five years artificial intelligence could wield a huge influence in elite sports in Australia.
Come post-grand final presentations, maybe the club computer, too, will get to wear a premiership medal around its neck.
Bob Murphy and Easton Wood hold up the premiership cup after the Bulldogs’ historic win.
Western Bulldogs gain from artificial intelligence
A Victoria University student analyses player data during a Dogs training session. Picture: Darrian Traynor.
Did artificial intelligence help the Western Bulldogs win the AFL flag? The Sons of the West certainly played a pack-centred style of football, in contrast with the more flowing play of opposing teams.
On grand final day they excelled at catapulting the ball out of packs like there was no tomorrow. After two or three lightning handpasses, the ball would be with a Bulldog in the open, and on the way to a goal. Coach Luke Beveridge can take credit for forming a “handball club” to teach players to punch ballistic missiles. The Bulldogs have been the No 1 handball*!ing side out of stoppages.
But don’t discount the role of computer machine learning — a form of artificial intelligence — in helping shape the Doggies’ game play. They are among a few AFL clubs going hi-tech to gain an edge.
Luke Dahlhaus of the Bulldogs handballs while being tackled during the 2016 AFL Grand Final
Sam Robertson, a specialist in machine learning and data analytics, and a senior research fellow at Victoria University, is a key member of senior Bulldogs staff. He says he has performed machine learning analyses of player positions around a pack.
“Absolutely, yes. If we want to take some x, y and time co-ordinates on each of our players around a stoppage, we can do that, and we can model that into some kind of situation where we match an ideal shape or ideal pattern, something we call the dominant region. We have an idea of what we want the stoppage to look like, we obviously try to get dominance around that area. Where the machine learning comes in, instead of it being one size fits all, it’s almost a what-if analysis: if the opposition (does) this, then how do we counteract that so we maintain our stoppage dominance?
Western Bulldogs sports scientist Dr Sam Robertson.
“We’re also making sure we have an understanding of what that does to the rest of the ground. If you move a player or change the set-up in one place on the ground it can also (have an) influence down the field. The machine learning side is the flexibility in decision-making.”
Machine learning is about computers making sense of the screeds of data that can flow from matches and training sessions. In AFL game play, that could mean analysing the movements of players in a match and simulating the outcome if they had acted differently. The Bulldogs can analyse who is owning different parts of the ground based on their location — a concept they’ve borrowed from soccer. Another is to measure congestion.
At the moment, Western Bulldogs are limited by not having the data of the opposition. “We know what we’re doing from a tracking player perspective, and we can then model that, but to obtain the information on the opposition is more difficult,” Robertson says.
Player position data used by The Western Bulldogs
For about five years the Bulldogs and other club players have worn a tracking device with a GPS or local positioning system for Melbourne’s Etihad Stadium, accelerometer and gyroscope sewn into the top rear section of their jumpers. Data can be collected about the location of every player at every moment of a match, in the way a Fitbit or Garmin wearable device can track exercise.
Fast computers can merge each player’s data into a single time and motion study of the entire match. It can identify position, velocity and acceleration at any moment. Staff then can look at alternative game plays and model outcomes. By linking two players they can observe tagging and how well their movement is coupled.
Analysis sometime involves small bits of game play. “We will look at one-minute, two-minute and five-minute increments, and to go back and create a retrospective model.”
So far Robertson has offered this analysis after games on a Monday. “The next stage is to use that in the coach’s box. It’s something we’ll look to implement in 2017. Even if that model adds another 5 per cent of understanding that the coaches don’t have, it’s going to be useful for us in a live sense as well.”
Norm Smith medallist Jason Johannisen shows what the Bulldogs ‘handball club’ does in a contest.
Game play is only one area where Robertson values machine learning. He has used it to look at the types of players on the Bulldogs list who may have deficiencies, what combination of players work best together and what types of players they need from the draft.
Player maintenance, injury management and player game day selection are other applications.
“Player selection is a really big one, we’re doing a lot of work in it. I’m talking from draft selection to scouting to team selection, we’ve got a dedicated PhD student in operations research at the moment, he’s working solely on player evaluation with a focus on player selection,” Robertson says.
Without detailed data on opponents, he has modelled the Bulldogs’ opposition based on general attributes. For example, he regards Swans and Giants as among sides that move the ball in a certain way.
Predicting injury, monitoring recovery and assessing the risk of fielding a player is “the buzz area” and machine learning offers a way to model player risk based on history. Robertson says he has accumulated two years of data on players.
He says a specific model is needed for each player that takes into account family history, say in terms of susceptibility to knee injuries. “If your model says this athlete has a one in three chance of being injured in this session, you have to make a decision from that,” he says.
As to the future, he says the ability of machines to ingest and analyse a match video will have a huge impact. Quantifying exactly what every player is doing, including how many sprints and jumps they do in a match, is beyond what current sensors offer. Video also could automate keeping stats of when players kick or mark the ball, handball and tackle.
Fans & media might get access to player movement data, says Robertson.
Alternatively, there are tattoo-like ankle sensors that could measure kicking. “I think video is a more realistic output, but it’s very hard to set up a permanent camera system for Australian football because of the size of our grounds,” Robertson says.
He believes machine learning systems will become more powerful when teams begin to share match day data. Champion Data, which provides stats for the AFL, and Catapult, which provides tracking sensors, have proposed data sharing to the 18 clubs.
“We could do thousands of things which we can’t do now,” he says. It’s being discussed, and he expects clubs to agree to it by next year. “They’ve done it in the NBA and they’ve done it in soccer.”
Across time he expects game data will be available to the media and, eventually, fans will gain access to it. “There’s a whole community of amateur analysts out there in football,” he says.
Robertson isn’t overstating the role machine learning played in the Bulldogs’ eventual victory, although it contributed to honing the team’s style of football. He cites the passion and remarkable commitment of Beveridge and the team, their determination and hunger for victory, and the return of key players from injury.
It’s early days yet for machine learning in AFL, but within five years artificial intelligence could wield a huge influence in elite sports in Australia.
Come post-grand final presentations, maybe the club computer, too, will get to wear a premiership medal around its neck.
Bob Murphy and Easton Wood hold up the premiership cup after the Bulldogs’ historic win.