How quantitative analysis could change StarCraft

The community likes to think that eSports is on the cutting edge of competitive play, but we still have much to learn from conventional sports. I don’t know much about the production and marketing side of sports, but I do know some statistics, and StarCraft, at least, is lagging behind conventional sports tremendously in quantitative analysis.

The only significant statistics I see from StarCraft are 1) win percentages in various circumstances and matchups and 2) Actions Per Minute (APM).* Win percentages are very broad metrics and not particularly instructive. APM is generally regarded as misleading at best and irrelevant at worst. Granted, StarCraft is a complicated game: the sides are often asymmetric, and game length varies. Sc2gears tracks many more statistics, but these haven’t become standard for broadcasting and analysis, whereas conventional sports broadcasts almost always feature statistics. Even Fantasy StarCraft discussions are pretty fuzzy, whereas fantasy football and baseball really are sports fans geeking out over numbers. Generally, StarCraft analysis is qualitative.

One of the coolest advances in conventional sports is computational, normative analysis. Today, games are tracked with better equipment, and by combining that data with advanced statistics, we can make predictions about what players should do in various circumstances. Because baseball is basically turn-based, it already has advanced sabermetric analysis (link to a reddit discussion about this). Basketball, however, has also been making strides in this area, according to this recent story from Grantland.

Hopefully you’re familiar with basketball, but if you’re not, it’s a 5 on 5 sport played on a (usually) indoor court. On opposite ends of the court, there are hoops, and each team’s goal is to shoot the basketball into their target hoop. On offense, teams design specific plays, and execution is key. On defense, however, teams have general schemes and react to what the other team is doing. Given that, it’s always been assumed that defensive skill is all about experience, “smarts”, and other intangibles.

Well, new analysis is starting to give us more concrete ways to understand defense. A new camera-tracking system in the NBA called SportVU can track where players are, and that data is turned into X-Y coordinates for clean video footage. And it gets even better. With significant computational analysis, the Toronto Raptors have come up with the “ideal” defense that minimizes the expected point value of a play**. You can watch the videos in the Grantland article where there are 2 sets of defenses super-imposed on the play: the actual defenders on the play, and the “ghost” defenders of where the players should be.

Hopefully you’re beginning to see how this analysis can impact StarCraft. Fortunately, we already all of the relevant data for unit positions in replays. If we can figure out how to parse expected outcomes from a large number of these replays, then we can begin to see general trends. Watching professional play, big deathball fights often come down to positioning. Is it safe to fight in this open area? Can you safely attack this base without getting trapped? How should you position your army to get the best engagement? Which units should be in front? These are similar questions to what the Raptors are answering in basketball.

It’ll be a lot of work to make this work. Specifically, it’s very difficult to parse meaningful actions out of a stream of data. The Raptors managed to recognize a pick and roll (one offensive player stands beside another defender, allowing the ball carrier to run around them. The first offensive player then goes in the opposite direction, hopefully resulting in confusion between the 2 defenders and leaving 2 open players). It may sound simple, but that’s darn hard, and I find that amazing.

Anyways, I think there’s a huge opportunity here for growth in eSports and a way for us to remain at the cutting edge of sports analysis, and even Artificial Intelligence at that. And there’s a tremendous amount of really interesting stuff that I would have to investigate and share, if you guys are interested. So before you head off, let me know in the poll below if you would be interested in me writing any of the following.

Which of the following topics should I elaborate on?

  • Just stick with the build orders, buddy (40%, 4 Votes)
  • Machine learning for event parsing and predictions (30%, 3 Votes)
  • Speculation on useful statistics for StarCraft (20%, 2 Votes)
  • Training AI to play StarCraft (this is tangentially related to this post) (10%, 1 Votes)
  • Advanced statistics and sabermetrics from baseball (0%, 0 Votes)

Total Voters: 8

Loading ... Loading ...

* If you know of more, please let me know. I’m interested.

** I’m not 100% sure how they do this in basketball, but I can explain how this is done in baseball in another post if you want