DeepMind builds universal AI agents in an open-ended game

Researchers from Britain’s DeepMind have created an extensive gaming environment, XLand, to train universally capable artificial intelligence agents. They can perform tasks they have never encountered.

Reinforcement learning typically trains & tests agents on the same game. New work shows how our team trains generally capable agents on huge game spaces, resulting in agents that generalise to held-out test games, & learn behaviours like experimentation https://t.co/fQ9UrOFkMb 1/ pic.twitter.com/UAPX5VgMmO

— DeepMind (@DeepMind) July 27, 2021

Rather than training agents on a fixed set of tasks, the team defined an entire universe of situations that can be procedurally generated.

The XLand universe comprises a suite of multiplayer three-dimensional games. Within its bounds, agents conducted 700,000 unique sessions across 4,000 worlds. Each agent in the latest generation completed 200 billion learning steps as a result of performing 3.4 million unique tasks.

The researchers say the resulting algorithm can successfully solve a broad spectrum of tasks—from simple object search to complex games such as hide-and-seek or capture the flag—that were not encountered during training, the scientists said.

“Currently our agents can participate in all procedurally generated evaluation tasks, with a few situations that not even humans could solve,” the study says.

According to the scientists, the agents exhibit general heuristic behaviour that is widely applicable to many tasks.

“This new approach marks an important step toward creating more universal agents capable of adapting quickly to a constantly changing environment,” they added.

As noted in June, DeepMind researchers said that achieving artificial general intelligence requires only reinforcement learning .

In July, AI lab specialists assembled and published the most comprehensive database of human protein structures , created by the AlphaFold neural network.

Follow ForkLog’s AI news on Telegram: ForkLog AI — all the news from the AI world!