Among the many other things artificial intelligence excels at, it is now better than the average human when it comes to playing 57 different Atari games. This is the latest achievement by DeepMind, the AI-focused company owned by Google’s parent, Alphabet.
Using Atari games to measure artificial intelligence agents is not new. In fact, it is commonly referred to as “Atari57” with the general premise being that the wide selection of games are varied enough in skill and challenge to make them ideal for reinforced learning. This is also not the first time that DeepMind has taken on the gaming world with the AI previously proving its mettle at Go and StarCraft.
DeepMind has now not only confirmed the creation of its Agent57 reinforcement learning agent, but also that it was able to score “above the human baseline” on every one of the 57 Atari 2600 games. As well as explaining the results in a detailed blog posting, DeepMind has also released a YouTube playlist showing Agent57 playing every one of the Atari games used for testing.
Does It Matter If AI Is Better Than You At Gaming?
While it seems somewhat dystopian to compare the results of AI to humans specifically, this is common with human performance typically used as a way to identify whether a non-human entity has performed well at something. At a fundamental level, being equal to or better than humans is simply a measure of how well the AI is doing at a task. More specifically, how much better than chance it is doing.
On a more general note, the performance in gaming does matter in many respects. While AI is not ultimately designed to play games, the act of reinforced learning is an important one. For example, regardless of what the algorithm(s) go on to do once graduating from sitting at home playing Atari games, the skills learned during the Atari57 experience will be ones that can be applied elsewhere. The whole premise of Atari57 is for these agents to learn different strategies and skills to overcome the varied challenges posed by the different games. These skills learned can then be applied in other real-world use cases when AI comes up against challenges or tasks it is less familiar with.
It is this last point that’s so important here as the Atari57 data can often be misleading due to agents sometimes performing exceptionally well on some tasks (games), but less so on others. In DeepMind’s case, the AI appears to have performed sufficiently well across all games. In other words, it is not just better than you at some of the games, or even most, but every Atari game it is has played. It is this ‘general inteligence’ achievement that seems to set the Agent57 artificial intelligence apart from the other AIs that have come and tried their luck at the Atari game lineup before.
