DeepMind has beaten the world's best at chess, Go and Starcraft. Now, it's added another victory to its collection; Quake III.
Researchers pitted the Google-owned AI against professional gamers at matches of the videogame's popular capture the flag mode, where players play in teams, guard their own team's flag, and try to capture their enemy's.
The result was emphatic, DeepMind's AI showing capabilities beyond that of humans at a collaborative game.
Videogame-playing AI "agents"
In a recently released paper for Science, DeepMind researchers reported that they had programmed automated "agents" to play the capture the flag game mode in Quake III.
These agents showed humanlike behavior and were able to "adapt to teammates with arbitrary skills," DeepMind researcher Wojciech Czarnecki told The New York Times.
So not only were the AI agents able to play the game effectively, they were able to adapt to other players, whether they were human or AI.
They learned skills displayed by human players, such as assigning an agent to defend a flag, while others went on the offensive to capture the opponent's flag and bring it home. They did so much more effectively than their human counterparts.
Beating humans at their own game
DeepMind's AI already beat the world's strongest chess-playing computer program, Deep Blue, which itself beat the world champion of chess, Garry Kasparov, in 1997. It also beat world-best programs at shogi (Japanese chess) and Go. It did this after just a few days of playing against itself, using reinforcement learning.
The AI, owned by Google's parent company Alphabet Inc., also took on professional Starcraft II players after accumulating the equivalent of 200 years of knowledge on the game. It won 10 consecutive games, losing only once.
Now, DeepMind's new research paper shows that it can also beat the best at Quake III, a game that requires increased coordination as it is in the first-person perspective. On top of that, the AI does so while playing a mode focused on collaboration.
The difference in ability is so large, in fact, that even AI "agents" programmed to be slower still had a 79% win rate against human players. The DeepMind research team tested the AI with a response-delay lag of a quarter of a second — similar to the reaction time of humans — to find the results.
The research team's work is undeniably impressive. Enabling a program to acquire 200 year's worth of knowledge in a fraction of the time and showing those learnings applied effectively in the virtual world is no small feat.
However, the question remains: what are the real-world benefits of a videogame-playing AI?
Many researchers believe that capabilities learned by AI in videogames can be translated into the real world.
A DeepMind blog post about the Quake III research says, "this work highlights the potential of multi-agent training to advance the development of artificial intelligence." It will, they say, enable the "development of robust agents that can even team up with humans."
Big promises in the field of AI are increasingly common. Just this week, MIT researchers revealed a new algorithm that showed promise for allowing safer work environments. The algorithm, in principle, allows robots to move more effectively around factory floors while avoiding collisions with human workers.
It is easy to imagine DeepMind's AI used for similar applications. Robots could effectively work as a team without needing preprogrammed inputs, allowing for a much more efficient form of automation.
Other, lesser reported DeepMind branches do also focus on real applications in medicine and science. AlphaFold, for example, has shown impressive capabilities in protein folding, a crucial practice in developing new drugs. The AI has also been shown to detect over 50 eye diseases as accurately as a doctor.
However, reservations remain. Mark Riedl, a professor at Georgia Tech College of Computing who specializes in AI, told The New York Times what the "agents" are doing in Quake III isn't really teamwork. They have such a programmed understanding of what is happening in the game, he claims, that their actions merely look like collaboration.
Max Jaderberg, an AI researcher for DeepMind, countered by saying, while definitions of teamwork are debatable, "one agent will sit in the opponent’s base camp, waiting for the flag to appear, and that is only possible if it is relying on its teammates.”
The Google-owned DeepMind AI is showing impressive capabilities in the virtual world, but there is a long road ahead before we see these fully translated into real-world benefits.