Researchers for Facebook’s AI labs recently entered their bot into a StarCraft competition. Here, we’ll talk about how they did, and why it’s important.
When I was a kid, video games were for kids. For the most part, that was the paradigm, and every video game ad in print or video knew it.
As I grew up, though, all of that changed. Games for adults hit the market, and this has branched out into the competitive sphere with the eSports phenomenon (check out these tips on being a better competitive gamer here).
See the disparity here? I think there’s an obvious different age group being targeted.
Now, video games are the inspiration for many of today’s technologists, and in a wonderful twist of fate, they are being used as training grounds for the future of AI technology. Every tech giant from Google to Apple has dipped their fingers into the AI boom that has come with Industry 4.0.
That includes social media giants as well, mind you. Facebook has their hands on some AI research, you may remember this fiasco from July when Facebook’s chat AI invented its own language.
Software has been using games for training for a long time. For example, IBM’s computer Deep Blue beat a world champion in chess back in 1997.
Games are a great environment to teach AI how to make judgement calls, especially when it comes to human behavior. In fact, judgement is something that has become a key component of contemporary AI research. You can see it in everything from AI that can help diagnose cancer patients to self-driving cars that don’t get confused by decals.
So far, though, the best game I’ve seen AI play yet has been Blizzard’s StarCraft.
A StarCraft Tournament With a Twist
Teaching AI to play games is some pretty serious business. That being said, it’s also pretty fun, especially when it comes in the form of a StarCraft tournament for AI called The AIIDE StarCraft AI Competition.
It’s an interesting tournament, where the real competition isn’t exactly the game itself but how well the programmers have designed their automated players. So, it’s like eSports, but the players don’t wear shirts (so they don’t have to take them off).
And the best part? Facebook quietly slipped into the competition unannounced with an AI called CherryPi.
Despite their recent big announcement of support for the Canadian Institute for Advanced Research (CIFAR), we haven’t heard too much about Facebook’s own AI research until now.
Now that we’ve seen how CherryPi did at AIIDE, we know a bit more. The bot came in sixth place overall (out of 28), but competitors did remark that it had unique reactions.
Dan Gant, a software hobbyist whose bot PurpleWave placed second, mentioned that CherryPi had a keen understanding of movement speeds within the game. While most bots advanced or retreated based on relative numbers, CherryPi was different.
“CherryPi appeared to know when it could move fast enough to sneak around an enemy to attack its base,” said Gant.
So, okay, it didn’t win, but that comment is pretty interesting because it shows us that CherryPi ‘thinks’ a bit unconventionally, at least when StarCraft is concerned.
This is where things get really interesting because it shows us how this little competition could have big implications for the direction that AI research is going in.Is there a better training ground for AI than StarCraft? #no #AIIDE #CherryPiClick To Tweet
What StarCraft has to do With Predictive AI
Real-time strategy games like StarCraft require a lot of snap judgement. StarCraft took up a lot of my teenage years, so I like to think that I know what I’m talking about.
Snap judgements, as researchers have come to learn, are very hard to program. To put it into perspective, when Deep Mind wins at chess, or Google’s AI AlphaGo wins at Go (read about that here), they have time to run through all the variables.
When you play a game like StarCraft, you have to make decisions on the fly. As I can attest, that can put stress on your ability to plan things out. For me, at least, the only way to get better is to play more, so you can get used to dealing with situations on the fly.
Many of us learn kinesthetically, or by touch and with experience, like with driving cars. Learning how to do these things comes naturally to us, but not so much for AI. That’s why deep learning platforms have become so important to AI research.
I got pretty good at StarCraft, myself, but that was after years of practice. Maybe that’s why I’m pretty impressed with CherryPi, even though it finished in sixth place in its first ever contest.#FacebookAI #CherryPi 6th place in AIIDE StarCraft competition, 2,049 wins out of 2,966 matches.Click To Tweet
See, CherryPi isn’t like the other bots in the competition. It’s learning, and I wouldn’t be surprised if it comes back with a better showing next year–just like AlphaGo got way better after just a year.