An artificial intelligence developed by British firm DeepMind has achieved ‘Grandmaster’ status in the real-time, sci-fi strategy game ‘StarCraft II’.
StarCraft II is one of the world’s most lucrative and popular esports, in which players control different alien races to build up forces and defeat their opponents.
With each battle coming with thousands of possible moves at any given moment, the video game presents a challenge that surpasses traditional tests like chess or Go.
The AI — dubbed ‘AlphaStar’ — proved its mettle in a series of online battles against human opponents, coming out above 99.8 per cent of players in the rankings.
This makes AlphaStar the first ever AI to reach the top tier of human performance in a professionally-played esport, without needing simplifying the game first.
AlphaStar’s achievement is a milestone for machine learning, with the research behind it having the potential to translate to other complex real-world problems.
Scroll down for video
An artificial intelligence developed by British firm DeepMind has achieved ‘Grandmaster’ status in the real-time, sci-fi strategy game ‘StarCraft II’
DeepMind machine learning expert Oriol Vinyals and colleagues developed the AlphaStar program, which is a type of AI that scientists dub a ‘multi-agent reinforcement learning algorithm.’
This means that AlphaStar learns through the experience of several competing deep neural networks — complex programs inspired by humans brains — which act to test out and constantly refine strategies and counter-strategies for in-game success.
These competing agents were started off with data on the strategies used by professional players in past games.
Having engaged human players in online games, AlphaStar was able to reach ‘Grandmaster’ level — the top of seven in-game ranks which also include ‘Bronze’, ‘Gold’ and ‘Master’ — for all three of the uniquely-skilled playable alien races.
AlphaStar then competed against human players in a series of online games, where it reached Grandmaster level for all three of the StarCraft races.
Only around 50–100 people in the game’s online platform rank higher than the AI.
‘This is a dream come true. I was a pretty serious StarCraft player 20 years ago, and I’ve long been fascinated by the complexity of the game,’ said Dr Vinyals.
‘AlphaStar advances our understanding of AI in several key ways,’ he added.
‘Multi-agent training in a competitive league can lead to great performance in highly complex environments, and imitation learning alone can achieve better results than we’d previously supposed.’
‘I’m excited to begin exploring ways we can apply these techniques to real-world challenges, such as helping improve the robustness of AI systems.’
DeepMind, like Google, is owned by California-based firm Alphabet Inc.
StarCraft II is one of the world’s most lucrative and popular esports, in which players control different alien races in real-time to build up forces and defeat their opponents
The sheer complexities of the game — which include how players can only a fraction of the game map at any given time — have meant that mastering StarCraft II has emerged as a grand challenge for AI developers.
Previous attempts to develop an AI that can rival top human players at the game have all failed — even when the rules were simplified, the AI was allowed to execute actions at superhuman speeds or action sequences were pre-programmed in.
‘The history of progress in artificial intelligence has been marked by milestone achievements in games,’ said DeepMind reinforcement learning expert David Silver.
‘Ever since computers cracked Go, chess and poker, StarCraft has emerged by consensus as the next grand challenge.’
‘[Its] complexity is much greater than chess, because players control hundreds of units; more complex than Go, because there are 1026 possible choices for every move; and players have less information about their opponents than in poker.’
‘We addressed this challenge using general-purpose learning methods, rather than handcrafting a solution, and played under the same conditions that humans themselves face.’
‘I’ve found AlphaStar’s gameplay incredibly impressive,’ said Dario Wünsch, a professional StarCraft II player for ‘Team Liquid’ who goes by the handle ‘TLO’.
‘The system is very skilled at assessing its strategic position, and knows exactly when to engage or disengage with its opponent.
‘And while AlphaStar has excellent and precise control, it doesn’t feel superhuman — certainly not on a level that a human couldn’t theoretically achieve. Overall, it feels very fair — like it is playing a ‘real’ game of StarCraft.’
With each battle coming with thousands of possible moves at any given moment, the video game presents a challenge that surpasses traditional tests like chess or Go
‘AlphaStar is an intriguing and unorthodox player – one with the reflexes and speed of the best pros but strategies and a style that are entirely its own,’ added Diego ‘Kelazhur’ Schwimer, who plays professionally for ‘Panda Global’.
‘The way AlphaStar was trained, with agents competing against each other in a league, has resulted in gameplay that’s unimaginably unusual.’
‘It really makes you question how much of StarCraft’s diverse possibilities pro players have really explored.’
‘Though some of AlphaStar’s strategies may at first seem strange, I can’t help but wonder if combining all the different play styles it demonstrated could actually be the best way to play the game.’
The full findings of the study were published in the journal Nature.
‘Ever since computers cracked Go, chess and poker, StarCraft has emerged by consensus as the next grand challenge,’ said DeepMind’s David Silver, pictured here centre
WHAT IS DEEP LEARNING?
Deep learning is a form of machine learning concerned with algorithms which have a wide range of applications.
It is a field which was inspired by the human brain and focuses on building artificial neural networks.
It was formed originally based on brain simulations and to allow learning algorithms to become better and easier to use.
Processing vast amounts of complex data then becomes much easier and allows researchers to trust algorithms to draw accurate conclusions based on the parameters the researchers have set.
Task-specific algorithms which exist are better for specific tasks and goals but deep-learning allows for a wider scope of data collection.