IBM Watson AI Will Calculate The Overwatch League’s Data-Driven Power Rankings In 2021

Everyone loves power rankings. Well, unless your team stinks enough to land at the bottom of most people’s lists. Power rankings give fans something else to obsess over and argue about, especially in the offseason or between match days.

The 2021 Overwatch League season starts on Friday and many in the Overwatch community, including the OWL casters and analysts on the Plat Chat podcast, have been busy hashing out their own power rankings of all 20 OWL teams.

But what if performance data could be used to objectively figure out which teams are the finest beyond mere series wins and map differentials? And maybe there’s a way to determine the best player in the world at a given time based entirely on stats, regardless of which roles, heroes or maps they play.

To do just that, the Overwatch League has teamed up with IBM, which is using Watson to crunch a vast volume of data and figure out who really is the best of the best.

This builds on a deal that Blizzard and IBM struck last year to use IBM’s cloud and artificial intelligence technology to bolster OWL analytics and interactive content. It was IBM’s first esports deal.

"Every player in the league now gets an overall score,” Overwatch League VP Jon Spector told me last week. “For the first time, heading into our fourth season, we'll have this ability for our fans, our broadcast talent and others interested in Overwatch League, to actually go and finally, definitely answer who's the best player in the world of Overwatch."

"[You have] three roles per team. You have the 32 characters to choose from," Corey Shelton, an IBM digital strategist and the leader of IBM’s Overwatch League project, told me. “When you look at that from a numbers perspective, we're looking at, essentially, nearly two million data points in a given match that have to be processed from a statistics perspective. That's a very interesting challenge and one that IBM is absolutely thrilled about trying to tackle with the Overwatch League.

“What it becomes is a really powerful demonstration of what's possible with data and AI today. These are technologies that we use in other industries, like banking and airlines and retail. But this opportunity gives us the chance to show Overwatch fans what our tech can do."

To build the AI model in time for the start of the 2021 season, IBM harnessed its Area 631 program. It’s an incubator-style initiative that brings together six experts — including data scientists, AI specialists and software developers — and gives them three months to come up with one solution. IBM and OWL’s analyst team worked together to create Power Rankings with IBM Watson (to give its full name) in just two months.

The AI model analyzes more than 360 different stats from OWL matches. Those include everything from accuracy and healing rate to how quickly players build up their ultimate.

The Area 631 team first used IBM Cloud Pak for Data to aggregate, sort and prepare a massive volume of data for analysis. By employing advanced analytics to figure out the correlations between the metrics and the outcome of a match, the team was able to figure out how each stat impacted whether a team won or lost.

The machine learning and AutoAI capabilities of IBM Watson Studio were then used to assign a weighting to each metric. Although all of the stats have an impact on the Power Rankings, around 30 of them are heavily weighted and are the most significant in how teams are graded. AutoAI “simplifies the process of building an AI model,” according to Shelton.

The Power Rankings, which Spector said was the culmination of months of hard work between the two teams, will be updated after each week’s matches. IBM will receive information about Overwatch League games directly from the OWL data center and feed that into Watson. It takes only around an hour for Watson to crunch an enormous volume of data every week. 

Spector said the OWL team values that fast turnaround time. If fans see someone performing exceptionally well in a match, they might rush to find out how that affects the player’s ranking. “What happens to their ranking week over week, I think, is gonna be one of the more fun stories for our broadcast to focus on, for fans to argue about,” Spector said. “We’re planning to come back after matches have wrapped every weekend and pretty rapidly update Power Rankings for fans.”

One aspect that’s especially interesting about the Power Rankings system is that all players, no matter their role, will be included on the same leaderboard.

Comparing DPS (i.e. damage) players against each other is relatively straightforward. You can look at stats like final blows or damage per 10 minutes to figure out how they match up directly against their rivals. The comparisons get more complex when you try to measure players in different roles against each other. Is a certain flex support player having a bigger impact on their team’s success than a main tank?

The AI will give each player an overall score, depending on their impact in that weekend’s matches. Unsurprisingly, the player with the highest rating will reach the top of the rankings. You’ll be able to browse that list and check out each player’s score, as well as rankings of players by role.

You’ll still be able to see more detailed stats for players elsewhere, but the Power Rankings with IBM Watson will focus on that overall number. Think of it a bit like the skill rating and the Top 500 leaderboards, but for the best pros in the world. The model can even account for the rare cases where Overwatch League players flex over from one role to another.

“Putting my fan-of-the-league hat on, it’s been a tremendous amount of fun, as the algorithms were being developed, built and refined, just looking at ‘Okay, this is the best player in the world in 2020,’” Spector said. He added that the AI model has been disproving some of his views “in what’s been this very fun experience over the last couple of months and one that I’m really excited for our fans to participate in soon.”

To test the model, the team used data from the 2020 OWL season. While some may believe DPS is the most important role in Overwatch, the stats from last season indicated otherwise. San Francisco Shock support stalwart Minki "Viol2t" Park was top of the rankings for 2020.

“I think the biggest surprise or what will challenge most people’s assumptions is that when you look at the top ten, it won’t be full of DPS players,” Shelton said. “That speaks to this idea that the game is made up of three roles. All three are very important in terms of the success of any team. You’ll see support and tank highly represented within the rankings.”

“Our reigning league MVP [Shanghai Dragons DPS star Byung-sun "Fleta" Kim] comes in 10th in the 2020 rankings,” Spector added. “I think part of that gets to the point that, with damage players, it can be easier to follow their impact, to see the incredible headshots or the ‘pop-off’ performance where they get four eliminations in the span of a second.

“A lot of what the tanks and support players, the best in the world, do is more subtle in terms of their positioning or saving a teammate at the right time, or great cooldown management or other types of things that can be harder for even me to pick up when I watch Overwatch matches. IBM Watson has been able to quantify that in a way that has really helped us better tell those stories, [in terms of] the impact the less-flashy players sometimes have.”

I’m excited to see how the Power Rankings shift from week to week, and especially between tournament cycles as new balance patches come in to play. Meta shifts always play a critical role in how each Overwatch League season shakes out. A team that rules the roost for the first couple of tournament cycles might slide way down the Power Rankings by the time playoffs roll around, and a struggling team might find itself in a favorable meta at exactly the right time.

There could be more OWL insights from Watson on the way, but for now, it's all about those Power Rankings. You’ll be able to see them on the Overwatch League and IBM websites each Monday once the season starts, as well as on OWL broadcasts. Right now, you can check out the rankings for last season.

There’s another side to the Power Rankings, as team owners, general managers and coaches will be able to see how players stack up against the competition. It seems the OWL broadcast team has quickly taken the Power Rankings to heart as well.

"The IBM folks ran a session for a lot of our broadcast talent to explain to them more about how this works," Spector said. "Every single one of those people have had positive feedback. They've loved what we're doing. They think it's gonna make their jobs easier, that it's gonna help us put out a better product for fans."

The stats don't lie, but that doesn't mean we can't still debate which teams and players are the best.

This is esports, after all. We’ll always find something to argue about.