The machines, it seems, can cooperate to deal with their enemy. Will mastering a video game translate to other applications as well?
Artificial intelligence has already surpassed humans in games like Chess and Go , but a new DeepMind algorithm now seems capable of putting aside its competitive streak to dive into something much more challenging: cooperation.
According to a study published in the journal Science, DeepMind researchers have designed automated “agents” to team up with machines and humans to play a version of the Quake III Arena video game in ‘Capture the Flag’ (CTF) mode, capture the Flag.
A big step
In the field of artificial intelligence, this is a breakthrough. Never before has a machine been competent enough in complex social interactions, to predict how others will behave, to navigate and win in multiplayer games … DeepMind has described the triumph as “human level performance”; but of course even the most complex games are still a long way from real life.
In Quake III’s CTF mode, each of the player groups tries to capture as many of the other team’s flags as possible. After starting at base camps at opposite ends of a map, players dash around an intricate maze of obstacles, attempting to steal the flag from the other side and bring it back to their side of the map. If a player is hit by an opponent, he is forced back to base to “respawn” and start the hunt from scratch.
Over the course of 450,000 games, comparable to four years of CTF matches, AI agents learned to maneuver through the maze while eliminating and evading opponents. The AIs had access to the same rules as a human player and each one learned independently from their own experience. At first, AI agents made random decisions, but through a brute force trial and error technique, each began filing tips and tricks that would increase their success rate.
Machine VS Human
The researchers then tested their newly trained agents IA in a tournament that included 40 human players. In some cases, human players were paired with AI partners; in others, they tried to defeat them; the caveat was that none of the human participants knew whether their teammates or opponents were humans or machines.
However, unlike people, AIs are able to process information and reaction almost instantly during gameplay, which is why DeepMind incorporated a 267 millisecond delay into the program. But this tweak doesn’t completely level the playing field.
On average, the AI-AI and human-AI teams captured more flags than the human-human pairs. And when human players laid out which teammates they preferred, most of them were machines.
However, in a separate test, two professional game testers were able to outperform the AI agents after 6 hours of training on an advanced map.
Since the Quake III Arena project concluded, DeepMind engineers have already designed a second system that has now mastered the multiplayer setting in Starcraft II , a classic strategy game set in space. But the team has its sights set on much more. Over time, they plan to present their AIs with many more challenges.
But AI still has a lot to learn and it will be a long time before they are ready for more advanced (and larger) applications.
Even the agents who are racking up points in Quake III Arena aren’t really collaborating the way humans are, says Mark Riedl, an artificial intelligence expert at Georgia Tech College of Computing who was not involved in the Science study. Each responds individually to the conditions it observes in the game, while human players can communicate with each other through messages or through real conversation.
On a real battlefield, the AI still has a lot of rungs to climb and it is unlikely to be anytime soon. But it is progress. The machine can learn by itself and, for the moment, the goal of DeepMind is not to kill us, but to increase its intelligence to carry out increasingly complex tasks autonomously.
Referencia: Human-level performance in 3D multiplayer games with population-based reinforcement learning. Max Jaderberg,Wojciech M. Czarnecki, Iain Dunning, Luke Marris, Guy Lever, Antonio Garcia Castañeda, Charles Beattie, Neil C. Rabinowitz, Ari S. Morcos, Avraham Ruderman, Nicolas Sonnerat, Tim Green, Louise Deason, Joel Z. Leibo, David Silver, Demis Hassabis, Koray Kavukcuoglu, Thore Graepel. Science 31 May 2019: Vol. 364, Issue 6443, pp. 859-865 DOI: 10.1126/science.aau6249
