His name is Agent57. DeepMind, the Google subsidiary dedicated to artificial intelligence, has just developed an AI that learned to play 57 Atari arcade video games. To do this, the algorithm is based on deep reinforcement learning. Editor's note) which allows it to do better than most humans.
The AI even manages to beat its counterparts who have struggled in some games in the past. This is a major achievement because the latter require adopting very different strategies depending on the case and therefore having a real capacity for adaptation.
A nice step forward even if the AI remains perfectible
To achieve this feat, the researchers built on their previous work. Deep-Q had already overcome several Atari games for the first time in 2012. As the MIT Technology Review, AI has a form of memory that allows it to base decisions on things it’s seen in the past.
This is particularly important because current AI systems that work with deep learning often lack versatility. We all know about the exploits of AlphaGo, which excels in the game of go and turns out today almost invincible for its human adversaries. However, the algorithm is highly specialized in this discipline but not or hardly operational in others. The development of AI capable of multitasking is therefore very good news.
However, Agent57 has a long way to go, by the admissions of its designers. In particular, they plan to reduce the amount of computing that AI needs to run. The latter is huge and can only learn to play one game at a time. They also want to improve the algorithm for some of the simplest games in the collection, which other much less successful systems get better results.