Before we get too far into the next NHL season, let’s look at an instant replay of game six of the 2008 Stanley Cup Final:
In those last 40 seconds, things get pretty intense. The Detroit Red Wings have the puck out of the zone, and Pittsburgh Penguins' Marian Hossa is coming down fast on the ice, lobbing a backhander at goalie Chris Osgood. With only 10 seconds left Osgood moves, and the puck hits his arm. Before it even starts rolling over the side of the net, the horn sounds and the time is up. The Red Wings have won 3-2. The season is officially over, but this is where the software-based analytics begins.
Ryan Lilien doesn’t look like a hockey player. There’s neat dark hair instead of a mullet, and glasses instead of a facemask. As a graduate of Cornell University and someone who earned a Ph.D. as well as an MD, you might assume he’s more egg-head than puckhead. But when you talk to him about Canada’s favourite game, you’d also find it hard to believe he wasn’t born here.
“When I was in graduate school (at Dartmouth College), I got hooked on hockey,” he says. “When I came to the University of Toronto about two years ago, my focus was on computational biology, but this is Canada. I wanted to see if we could do something that could combine both interests.”
The result is a project called the Computational Analysis of Ice Hockey Gameplay. The goal is to develop a system that will learn how hockey is played and help a team improve their performance. Or better yet, help a team like the Penguins understand how Osgood knew just where and how to move in those last 10 seconds, and how they could outthink him.
“Tracking the puck is hard. It’s small,” he says. “As a hockey fan, you don’t always see the puck, but you know based on the position of the players where the puck is and where they’re moving. The computer could do the same thing.”
Lilien’s approach is to use “machine vision,” a subfield of engineering that encompasses computer science, optics, mechanical engineering, and industrial automation. Today it’s mostly used to monitor and inspect packaged goods in a manufacturing setting, but the U of T team project will apply it to video footage of NHL games. The software will study where players move, their habits and play styles. Specially-developed algorithms will then attempt to reason, under uncertain conditions, what kind of patterns emerge and what relationships they have to winning or losing a game.
“It’s not simply to say this shot was taken from this position, but what led up to that situation?” he says. The concept is not far removed from how enterprises use business intelligence (BI) software. Based on sales, customer requests and other information, they try to determine what kind of cause-and-effect relationships their actions have on a company’s growth, and adjust their strategy accordingly. Like hockey teams, they are studying outcomes. Or forecasting.