0

Bought by Google in 2014 for 400 million pounds, DeepMind is an AI company best known for beating the world champion in the game Go.

Now, the AI agents developed by Google’s DeepMind subsidiary have beaten human pros at StarCraft II — a first in the world of artificial intelligence. In a series of matches streamed on YouTube and Twitch, AI players beat the humans 10 games in a row. In the final match, pro player Grzegorz “MaNa” Komincz was able to snatch a single victory for humanity.

Unlike Chess or Go, StarCraft II is not a game where you take turns and watch all the pieces and have time to figure out your next move. StarCraft II happens in real time.

But DeepMind didn’t seem to have a problem with that. It beat the pro “TLO” and then beat “MaNa”. Overall, DeepMind won 10 games but they streamed a live event recently where MaNa took 1 game for the humans.

The AI created to play StarCraft II was dubbed “AlphaStar”.

Professional StarCraft commentators described AlphaStar’s play as “phenomenal” and “superhuman.” AlphaStar was particularly good at what’s called “micro,” short for micromanagement, referring to the ability to control troops quickly and decisively on the battlefield.“PHENOMENAL UNIT CONTROL, JUST NOT SOMETHING WE SEE VERY OFTEN”

To ensure fairness, Alphastar was nerfed, or hobbled if you will, in certain areas. For instance, it was not allowed to perform more clicks per minute than humans would be able to. One advantage it did keep was to be able to see the whole map at once vs. having to scroll around the map manually as humans have to.

One other note, the computer only played as one race (StarCraft II has three). The pros that it played were sometimes playing with races they are not most proficient at. Can the AI learning be spread successfully to all races and work no matter what race the human plays? This will probably happen, but how much longer will that take? How well would Alphastar play – or more precisely – how fast would it learn and take to become a champ as another race?

A graphical representation of AlphaStar’s processing. The system sees whole map from the top down and predicts what behavior will lead to victory. 
A screenshot from the games in December, showing AlphaStar facing off against TLO. 

In order to train AlphaStar, DeepMind’s researchers used a method known as reinforcement learning. Agents play the game essentially by trial and error while trying to reach certain goals like winning or simply staying alive. They learn first by copying human players and then play one another in a coliseum-like competition. The strongest agents survive, and the weakest are discarded. DeepMind estimated that its AlphaStar agents each racked up about 200 years of game time in this way, played at an accelerated rate.

What is DeepMind?

Founded in London in 2010 by Demis Hassabis, Shane Legg, and Mustafa Suleyman, DeepMind has developed machine learning systems that uses deep neural networks, reinforcement learning, and systems neuroscience-inspired models. The startup was purchased in January 2014 by Google for a reported 400 million, with Hassabis remaining CEO of DeepMind.

Many other AI programs like IBM’s DeepBlue, which defeated Garry Kasparov in chess in 1997, have used explicit, rule-based systems that rely on programmers to write the code. However, machine learning enables computers to teach themselves and set their own rules, through which they make predictions.

In March 2016, DeepMind’s AlphaGo program beat world champion Lee Sedol in 4 out of 5 games of Go, a complex board game—a huge victory in AI that came much earlier than many experts believed possible. It did this through combining “Monte-Carlo tree search with deep neural networks that have been trained by supervised learning, from human expert games, and by reinforcement learning from games of self-play,” according to Google.

But beyond mastering games, deep learning has other more practical applications. In 2012, it was used to recognize a million images with a 16% error rate—which is now at about 5.5%. Deep learning is also used in text-based searches and speech recognition. According to founder Mustafa Suleyman, it achieved a “30% reduction in error rate against the existing old school system. This was the biggest single improvement in speech recognition in 20 years, again using the same very general deep learning system across all of these.”

Deep learning is also used for fraud detection, spam detection, handwriting recognition, image search, speech recognition, Street View detection, and translation. According to Suleyman, deep learning networks have now replaced 60 “handcrafted rule-based systems” at Google.

Why does DeepMind matter?

While Google DeepMind’s accomplishments in the gaming world are impressive, the implications of its machine learning platform are far-reaching. An announcement that DeepMind was able to slash Google’s electricity bill by increasing energy-efficiency has big implications in both economic and environmental realms.

Also, DeepMind’s partnership with the National Health Service, part of DeepMind Health, employs machine learning in spotting critical conditions in eye health. It hopes to eventually use algorithms to personalize health care treatments, determining which work best on patients, given their previous medical history.

Pin It on Pinterest