Google DeepMind AI Goes On Quake III Killing Spree With Human-Like Skills

Quake III
Researchers and computer scientists far and wide are leveraging artificial intelligence for all kinds of tasks, everything from weather prediction and automating dangerous tasks, to finding cures for diseases and solving complex social problems. Oh, and gaming. In fact, Google trained its DeepMind system to play Quake III, and apparently it's kicking some human butt.

Why train an advanced AI system to play a computer game? There are actually practical applications, and they have nothing to do with winning an esports tournament or bragging rights.

"Billions of people inhabit the planet, each with their own individual goals and actions, but still capable of coming together through teams, organizations and societies in impressive displays of collective intelligence. This is a setting we call multi-agent learning: many individual agents must act independently, yet learn to interact and cooperate with other agents. This is an immensely difficult problem— because with co-adapting agents the world is constantly changing," Google explains. "To investigate this problem we look at 3D first-person multiplayer video games."

What's interesting here is that Google's AI has to learn from scratch how to see, act, cooperate, and compete in unseen environments, and all from a single reinforcement signal per match. In other words, Google is not feeding DeepMind any instructions on how play Quake III. Instead, it has to compete against itself until it learns how to play and what strategies work best.

"Each agent in the population learns its own internal reward signal, which allows agents to generate their own internal goals, such as capturing a flag. A two-tier optimization process optimizes agents’ internal rewards directly for winning, and uses reinforcement learning on the internal rewards to learn the agents’ policies," Google explains.

Google DeepMind Quake III
Click to Enlarge (Source: Google)

After doing so, Google held a tournament with 40 human players, randomly matching up humans and AI agents in games on two-player teams, both as opponents and teammates. Out of all the teams, the ones that consisted entirely of bots (FTW) performed the best.

Google has bigger aspirations in mind than dominating in video games. It wants to continue improving its current reinforcement learning model and population-based training methods to further develop AI. Those concepts can then be applied to more meaningful winning Jeopardy.