After the game of chess or Go, artificial intelligences Deepmind attack video games. After having proven in the field over the years, Deepmind today revealed that one of its AI managed to complete the 57 games released on the Atari 2600, even the most complex.
Artificial intelligence gaming
Will artificial intelligence soon be able to surpass the best players in the world even in video games renowned for their difficulty, such as Sekiro: Shadows Die Twice? For now, she’s only mastering the second generation of video games, but her current progress is historic. DeepMind, Google’s AI laboratory, has managed to develop AI capable of mastering the 57 games released on Atari 2600, one of the first video game consoles in the 1970s and 1980s, and to do better than human being. As DeepMind proudly points out in its report, this AI named Agent57 is the “first IT agent in deep reinforcement learning to score above the human average across all 57 games in the Atari 2600.”
To successfully beat and finish many more games than its predecessors, the Agent57 used a decision-making model that encouraged them to observe the game and test a variety of options before formulating a winning strategy. The rendering of the latter was then automatically judged by a “meta-controller”, in a way responsible for weighing the pros and cons before continuing with new observations or a new test. The goal for DeepMind is to get as close as possible to the design of true general artificial intelligence.
This means that, to get there, its scientists and engineers must program and train the most versatile and versatile AI possible, one that will prove effective in the widest variety of tasks possible. Video games, even the most primitive, are one of the best ways to try it out: if the technical constraints are the same, their operation differs on each title. Agent57 therefore revealed that it was not only as capable or even better than human intelligence in certain games, but managed to move from one game to another without losing efficiency.
Agent57 playing the 57 games of the Atari 2600
Agent57 facing its four biggest gaming challenges
DeepMind notes, finally, that Agent57 was the hardest hit by four specific games. Montezuma’s Revenge and Pitfall would have challenged him to test a large number of strategies before achieving success and mastering the best method possible. Likewise, Solaris and Skiing would have handicapped AI in their lack of direct feedback between test and result. Agent57 therefore took longer to learn the right techniques.
“Agent57 has been able to improve by increasing the number of simulations: the longer they train, the higher their score,” said DeepMind. The concern, however, notes MIT Technology Review, is that Agent57 had to re-adapt to each game it played, whether or not it was already tried, as if it was the first time that ‘she faced him. Human beings, on the other hand, can master several tasks or games at the same time without losing experience or dexterity.