DeepMind agents show broader skills in simulated game world

DeepMind researchers built AI agents that learned to handle many games in a 3D virtual world and then used those skills in unfamiliar games, according to DeepMind. The result matters because the ability to transfer learning across tasks is one of the main barriers between today’s narrow AI systems and more general-purpose software.

The work took place in XLand, a simulated environment where software agents see from a first-person view, DeepMind said in a research paper and blog post. The agents learned by trial and error across a large set of simple games, including tasks such as moving objects with certain colors or shapes to specified places, or hiding an object from another player’s view.

According to Fortune’s Jeremy Kahn, many of the games involved two or three players, with some designed for competition and others for cooperation. After training, the agents could use learned behaviors and strategies in new XLand games without being trained for each one individually.

Why the work drew attention

DeepMind has previously built systems that reached very high performance in board games and Atari video games, Fortune reported. In those earlier cases, the algorithm had to be trained again for each game, rather than carrying a reusable set of strategies across multiple games.

Max Jaderberg, the DeepMind researcher who led the project, told Fortune the approach, which DeepMind calls “open-ended learning,” represents “one of the essential steps” toward building agents that can perform many tasks. He also said researchers should “stop chasing absolute performance in one narrow domain” and focus more on broad performance across many tasks, even if that means giving up expert-level results in any single one.

Chris Nicholson, chief executive of Pathmind, told Fortune he saw the result as “the G in AGI,” referring to artificial general intelligence. Pathmind sells AI systems for industrial operations, and Nicholson said the DeepMind methods could help create software and robots that are easier to train and more adaptable than current systems.

Nicholson did not describe those systems as human-like, according to Fortune. He said they could still be more capable and less costly than existing tools in some commercial settings.

The broader AGI race

Artificial general intelligence, or AGI, refers to software that can perform many tasks rather than one specific job, Fortune reported. The publication said the current AI boom remains centered on narrow systems built for defined uses.

Major technology companies have continued to fund AGI-related research despite the difficulty of the goal, according to Fortune. Alphabet has invested billions in DeepMind, whose stated aim is AGI, and has also funded Google Brain, while Microsoft invested $1 billion in OpenAI, another organization focused on AGI, Fortune reported.

Fortune also reported that Elon Musk, who helped found OpenAI, has spoken often about AGI, and that Tesla’s robot plans may reflect continued interest in the field. Facebook has said it does not really believe in AGI, though Fortune reported that the idea remains of interest to some leading machine-learning scientists at the company.

Wojciech Czarnecki, another DeepMind scientist involved in the research, told Fortune that general agents could be useful when a task lacks enough data or a good simulator. In that case, he said, a system with broad skills might perform adequately at first and improve with limited training and experience.

Fortune described AGI as still potentially years or decades away, and possibly unattainable. DeepMind’s XLand work does not settle that question, but the research gives AI developers a clearer example of skills transferring across tasks inside a controlled world.

This story draws on original reporting from Fortune.