The DeepMind team, led by Max Jaderburg, used reinforcement learning to train their AI agents. Reinforcement learning uses rewards to incentivize an AI agent to achieve a complex goal over many steps. Just as you might reward a child with a gold star every time they do their chores, this algorithm gives the AI agent positive feedback every time it does something right.

Reinforcement learning has proven effective in training AI agents to play complex single-player games in the past, as well as competitive two-player turn-based games like chess and Go. Although AI capabilities have reached "superhuman mastery" for these games, the teamwork aspect of multiplayer games was thought to be too complicated for even the smartest AI.

But in this study, published in the journal Science, Jaderburg and his colleagues let AI agents loose in the game Quake III Arena — specifically, in a "Capture the Flag" scenario where two teams compete against each other to seize enemy flags. The game presents countless different randomly generated environments, so the AI players had a lot to learn.

While past studies have armed AI agents with knowledge of game strategy or models of the environment, the DeepMind study started the AI players with the same information a human player would have. In this case, each agent learned independently based on experience and then applied learned strategies to a new map in the next game — just as any human playing the game would.

But the AI still had a major advantage: the consolidated knowledge of a whole population of bots. In a practice called "population-based training," the scientists had several AI agents play against each other in multiple games and then selected the best players. Based on predetermined parameters for the group, the scientists weeded out underperforming agents and replaced them with "offspring" armed with knowledge from the first generation.

After about 450,000 games, the AI players were able to beat human players — and not just any human players. These bots beat professional gamers, even when the bots' reaction times were slowed down to human levels. Even after the human gamers had hours to practice against the AI, they could only manage to beat the bots 25 percent of the time.

This new technology is good for more than antagonizing gamers who think they're the best of the best. This study demonstrates the potential of "multiagent reinforcement learning" — reinforcement learning where several agents learn at once to create a combined knowledge base, rather than having one agent improve its own skills through trial and error. By having multiple agents testing their knowledge in varied environments, this strategy creates scalable learning that can be applied in dynamic scenarios. In other words, the bots are only getting smarter.

Get stories like this one in your inbox or your headphones: Sign up for our daily email and subscribe to the Curiosity Daily podcast.

Go deeper into deep learning with "How Smart Machines Think" by Sean Gerrish. We handpick reading recommendations we think you may like. If you choose to make a purchase, Curiosity will get a share of the sale.