Computer Science Professor, HockeyTech Team Up To Transform Business Of Player Development

Pascal Poupart, a professor at the David R. Cheriton School of Computer Science at the University of Waterloo in Canada has spent the last year working alongside HockeyTech, a data provider, to automate recognition of events during a hockey game using machine learning techniques.

It takes substantial effort to accurately track puck possession, time on ice and even the number of shots a player takes. It requires even more manpower to render that information useful.

Scorekeepers manually score each game, assistant coaches break those games down into individual events for analysis, and scouts rate and review individual players.

It’s an expensive and time-consuming process.

“What we are doing is making it possible for teams to capture more events with fewer staff,” HockeyTech CTO Cary Moretti said. “Our (machine learning) application automates this capture so that more data is available. Scouts can then use that data to identify prospects and even supplement their reports. Better, general managers can use that additional data to make better recruiting and trade decisions.”

**

Machine learning can be described as any system that can self-adjust during a particular process.

“Today we’ve got lots of data about all kinds of things that can be measured by sensors and so on, and that data is usually just raw, low-level measurements,” Poupart said. “One way of thinking about machine learning is that we design some algorithms that can go through this data and extract useful information.”

Poupart went on to explain machine learning as relative to a recent paradigm shift in the way computers are programmed.

Historically, computers were fed a list of instructions and they would follow those instructions step by step, just like you would follow a recipe for chocolate cake.

But now we all want cakes so intricate and individualized that no human can generate a comprehensive recipe that allows everyone to receive a personalized dessert.

The same reasoning applies to our data.

Imagine that we want a computer to differentiate between a shot and a pass. It sounds simple enough, but consider this play.

It would be exceptionally difficult to explain to a computer why this would qualify as a deflected shot and not a pass to the player in front of the goal. No human could generate the necessary instructions.

So we have turned to machine learning, where instead of telling the computer what it needs to do, we provide it with examples of what should be done.

In this case, instead of trying to explain why the aforementioned was a deflected shot — not a pass — the computer is simply told that this is an example of a deflected shot and allow the computer to generate its own rules as to why.

With enough examples (known as training data) the computer can learn to accurately recognize and classify future events.

**

Similar machine learning techniques have been applied to baseballsoccer and American football, but never in hockey. Moretti explained why.

“It starts actually with the tracking of the data, that’s the real issue,” Moretti said.

Indeed, HockeyTech is one of few organizations gathering hockey data. Not for lack of interest, but because hockey does not easily lend itself to data gathering.

First and foremost, baseball, soccer and American football are much slower than hockey. They’re also played on a big, green field, which is the perfect background for visual identification of objects (contrary to what you might think, solid white backgrounds, like a sheet of ice, are very difficult to track against).

There’s also the added limitations of two-dimensional tracking.

The tracking data Poupart works with is a set of simple, X/Y coordinates for every player. The data doesn’t reveal the player’s orientation or stick position. It doesn’t tell you where the player is looking, if he is standing tall or crouched or if he is winding up to take a shot, etc.

But machine learning techniques can help analyze those data points as a sequence – almost like the data version of a claymation video – and subsequently infer events as intended.

***

So what comes next?

Poupart and his team need to make sure their system is accurate by proving their result data out against human-collected data. Then, they can integrate their new algorithms into their existing platforms.

“I believe that we should be seeing some of this ready in the next six months or so and then it’s a matter of experimenting to see how it works in the real world,” Moretti said.

It’s important to note that HockeyTech isn’t developing these algorithms for big spenders like the NHL. Rather, their goal is to augment data collection and analysis among smaller teams and leagues.

“This particular product tracking is not an NHL product,” Moretti said. “We’re very, very interested in the development of players, and we believe that if we can build a system that’s accessible in the right financial model, i.e. that they can afford, and we can get this out into the junior rinks and the major juniors and even the ECHL and the American Hockey League, those levels of hockey below the NHL, then it will truly change the game.”

Michael Oke, general manager of the Peterborough Petes in the Ontario Hockey League, said the applications Poupart and his team are developing could prove instrumental in identifying players who might otherwise go unnoticed as well as in comparing players from different leagues.

Furthermore, Oke said these technologies can help level the playing field and make proper statistical analysis affordable and accessible to all.

“The statistics in the NHL, the statistics in the American League and the statistics in major junior and college, they’re all pretty accurate, but when you get to some of the levels the statistics aren’t accurate,” Oke said. “The NHL and some of the professional teams and some of the college teams probably have some significant budgets where they can put some resources in to make sure that they have up-to-date, accurate statistics. But when you look at the NHL draft, (consider) a high school in Minnesota. They can’t afford to have that type of technology, and then if a team in the Ontario Hockey League can, how do you compare the guys in Minnesota with the guys in Ontario? There’s such a discrepancy in the accuracy in the information.”

Image and thumbnail via Sporttechie.com

Alex Siegman