AI beats human players in StarCraft. Here’s why it matters

DeepMind made history as it introduced AlphaStar, the first artificial intelligence that defeated professional players, 10-1, in StarCraft II, a complex real-time strategy game.

I have been watching South Korea’s Global StarCraft II League for over seven years, and I agree with game commentators when they said that AlphaStar showcases impressive strategic thinking and skills that are comparable to the best players in the world.

There's still room for improvement, and it's fair that some passionate fans feel that machines have too much of an advantage with their superhuman reflexes, despite the developers’ attempt to limit the machine's performance to human levels and level the playing field.

However, as much as I love StarCraft, the bigger question here is about AI's significance and impact on technology and the world. So let us explore three profound implications beyond the world of StarCraft: the arrival of real-time AI, the shift beyond supervised machine learning, and the future of mutual learning between AI and humans.

1. Real-time AI has arrived

DeepMind’s AlphaGo ignited the world’s imagination in 2016 by beating Lee Sedol, one of the best Go players in history. In just three years, the DeepMind team has moved from a fully observable, turn-based game to a partially observable, real-time game.

As shown in the illustration below, AlphaStar “looks” at the screen, passes the information through a series of artificial neural networks (the machine’s brain), and executes actions in real time. It fights air and ground battles while building units at the same time, just like what MaNa - its professional human adversary - does.

Unlike Go, where the machine can take minutes to calculate a single next move, a game in StarCraft can be won or lost at a moment's notice. Furthermore, the fog of war limits what players, including AlphaStar, can see. This means that AlphaStar has to make judgment calls every second based on imperfect information, exploration, and control over hundreds of different units and buildings.

Why does it matter?

Making good decisions under time pressure and with imperfect information in hand is one essential ability that differentiates humans from machines. However, AlphaStar shows that AI is closing in on that gap.

From optimizing and reducing traffic congestion in Hangzhou to delivering Amazon packages or searching and collating financial information, the potential applications of real-time AI will transform companies and even countries. Moreover, AI always performs at its peak - it does not get tired, distracted, or injured over time. This is why it can also assist in high-intensity work like surgeries, where precision, focus, and consistent good decision-making are required.

2. The shift beyond supervised machine learning has begun

Most StarCraft AI bots are terrible and employ a supervised machine learning approach. They execute moves from their playbook database, following rules set by human developers, and lack the capacity for big-picture strategic thinking.

They struggle to keep up with the average players in the Gold or Platinum league. The game is too complicated for developers to load the AI with solutions for every possible scenario. This explains why it's a big deal that AlphaStar just defeated MaNa - a professional player in the top 0.05 percent - with a 5-1 score.

The difference in the quality of play between traditional StarCraft bots and AlphaStar is like that of your college basketball team versus NBA players. AlphaStar learns, adapts, responds to challenges, and comes up with new strategies, just like a top-level human player.

Why does it matter?

AlphaStar gives a glimpse into a future where machines are not just following historical trends - they're also actively creating new ideas and responding to changes. When Amazon recommends a book, when Netflix suggests a movie, or when GrabShare matches you with other riders, the machines use massive amounts of historical data on user behaviors, clusters, and even traffic conditions to generate a final result.

While individual preferences exist, most people are not that unique once clustered, and supervised learning works well in these scenarios. In areas like education or healthcare, however, flexible programs designed around students’ learning habits or individualized treatment plans for patients are preferable, so having an AI that automatically adapts to every user's needs could radically transform lives.

3. Mutual learning between AI and humans

It used to take years of dedicated practice, coaching, and support to train the next StarCraft superstar. Now, it only takes 14 days to build 200 years' worth of StarCraft experience, according to the AlphaStar team.

It is interesting to note that AlphaStar learns in a very similar way as human players do. It studies game replays and forms its own AI league to play against itself over and over again, learning from each game. The stark difference here is that only few rookies make it into the professional StarCraft leagues, and most players retire by the age of 30 because they incur injuries like carpal tunnel or their age has caught up with them.

MaNa is a great player at 25 years old, but can he retain his current level of play for the next five years? The odds are against him.

Why does it matter?

Human-computer interaction and learning have always been a core component of AI, but AlphaStar takes it to the next level. By adopting human-centric ways of learning or analyzing human decisions, AI and humans can join forces for greater productivity. AI can store immense amounts of knowledge and skills on a particular subject, making it a great teaching tool to discover and practice new techniques.

For AI scientists, this is an exciting time for the field. Similar to how AlphaGo contributed to deep reinforcement learning, the next few years will likely see considerable leaps in population-based and multi-agent reinforcement learning based on AlphaStar. This is a much better simulation of how the real world works, with agents of different skill levels and motivations all interacting with each other in the same space.

So why should we care about this?

One of the biggest criticisms of AI is the “big baby problem.” The technology is very good in games like Go, Defense of the Ancients (or more popularly known as DotA), and now StarCraft but lacks practical applications in the real world.

Indeed, the current generation of AI is still very narrow in its focus and impractical beyond specific domains. We are also decades away from artificial general intelligence.

However, the technology will continue to improve and evolve. AlphaStar has brought some possibilities and ideas to the public, but its actual impact will be felt years later after some of these concepts are deployed. Only then will we truly appreciate this moment when machines first stepped up their game.

*I first posted this article on Tech in Asia (TIA) as an subscriber exclusive content. For more in-depth news about Southeast Asia Tech and startup scene, do check out TIA.