AlphaStar AI Achieves Grandmaster Rating After Beating the Best Human Players in Video Game StarCraft II
Wed, April 21, 2021

AlphaStar AI Achieves Grandmaster Rating After Beating the Best Human Players in Video Game StarCraft II

StarCraft II is a popular science fiction online game, wherein players compete in real-time. / Photo by: SobControllers via Flickr

 

An artificial intelligence dubbed AlphaStar, which was developed by Google AI company Deepmind has recently achieved a grandmaster rating after it defeated the best human players of video game StarCraft II.  British multidisciplinary scientific journal Nature shared that the AI was included in the top 0.15 percent of the 90,000 players in the region.

Competing at the Highest Level of the StarCraft II

StarCraft II is a popular science fiction online game, wherein players compete in real-time. The game is often played one-on-one in intergalactic warfare. It also involves three species: The (1) Terrans or the humans, the (2) Zerg or the species of assimilated lifeforms, and the (3) Protoss or the technologically-advanced species that have powers. 

The DeepMind team, which previously developed artificial intelligence that can play chess and Go, also developed AlphaStar in the hopes of making a machine that is capable of understanding and learning any task that humans can do. It considered StarCraft II because of its rapid pace and strategy complexity.

Why StarCraft II Posed a Tougher Challenge for the AI

The developers said that StarCraft II posed a tougher challenge for artificial intelligence than other board games because the opponents are usually hidden from the view. Research group head Dave Silver said that one of the things they were excited about is that the video game presents a lot of challenges that are also seen in the real-world.

Players of the game begin with few resources so they have to gather gasses and minerals to create technologies and make new buildings in the platform. In the same way, they can invest their time in increasing the worker units. Gamers are given a small map to view and one of the challenges they will encounter is the number of choices in the game. With more than a hundred trillion possible game moves and thousands of choices, players should make a move before it becomes obvious who has won and overwhelmed others' structures.

The challenge, therefore, is to multitask by executing combat maneuvers, managing the resources, and strategizing against the opponents.  

 

StarCraft II posed a tougher challenge for artificial intelligence than other board games because the opponents are usually hidden from the view. / Photo by: metamorworks via Shutterstock

 

Training AlphaStar 

DeepMind trained Al to recognize patterns. The team first trained three neural networks for each race of aliens it would encounter in the game. Then, they tapped into the database of past StartCraft II games that were also provided by Blizzard, the video game company behind StarCraft. They used the database to train the artificial intelligence agents to mimic the moves of the strongest game players.

These agents were then copied to compete against each other, improving their skills. In machine learning, this is called reinforcement learning. BBC’s technology desk editor Leo Kelion, who is not a part of the DeepMind team, narrated how the Google AI group also built the exploiter agents. The purpose of these agents is to expose the weaknesses of the strategies used by the main agents so they can find the means to correct the weaknesses.

AlphaStar project co-lead David Silver likened the exploiter agents as “sparring partners" because they were forcing the main agents to choose robust strategies. Training the AI just took the DeepMind team 44 days but it represents about 200 years if compared to human gameplay.

AlphaStar’s Grandmaster Status

The entire training resulted in the Grandmaster status of AlphaStar for the alien races it fought in the game. Nevertheless, there remain about 50 to a hundred professional StarCraft II players who outperformed AlphaStar during the game. The researchers emphasized that it is not about developing a technology that can fight wars but for general-purpose intelligence. 

Some gamers considered the AI’s performance “impressive.” Meanwhile, other AI scientists and experts are not fully convinced that AlphaStar can claim the game’s complete victory because it still has a few weaknesses. Memorial University of Newfoundland’s AI researcher Dave Churchill, for example, said that StarCraft is “nowhere near being solved” although he acknowledged that it is the strongest AI system for the StarCraft game.

AI and Gaming Industry

AI has been changing the gaming industry in major ways, including how video games are played and developed. Game and eSports analytics and market research company Newzoo shared in its market report that the video game industry is far bigger than the music and film industry put together. 

Combined across PC games, video games, and mobile games, it already reached 2.3 billion gamers worldwide. In 2018, the gaming video content market was $5.2 billion and forty-five percent of US gamers were women.

AI and machine learning are also used in game development, such as by playing as NPCs (non-player characters) and responding to the user’s actions in unexpected and unique ways. AI further makes more realistic interactions as neuro-linguistic programming creates more realistic responses and conversations for the AI. Even game developers are using machine learning and AI to make their games more beautiful and less pixelated for the players.