That artificial intelligence is superior to us in some games, the past has already proven. Also first attempts to deal with Atari games are nothing new, so why then this message. The latest system called “Go Explore” has now spoken all the high scores in the complex computer games classics and can now even make world record holders sweat.
Atari games are a popular test to challenge the reaction speed of AI algorithms. They are complex, like to mislead the player and are based on a reward system by means of collected points. A perfect learning environment for artificial intelligence. The machine learns by trying and failing, and here, in the style of reinforcement learning, it tries to achieve the best possible score. If the machine works well, it receives its reward in the form of a high score.
Montezuma’s Revenge Solved by Go-Explore (Sets Records on Pitfall too) (Source: Uber AI Labs)
In games like Pitfall or Montezuma’s Revenge, however, this kind of learning and rewarding has failed so far, because the action and the resulting reward are not immediately recognizable, as in PacMan and co, but only after many more moves. The machine therefore quickly gets stuck and doesn’t get the boost it gets in other games. To counteract this, the “Go Explore” team has now developed an AI that can fall back on a kind of memory. It remembers certain states and moves and “thinks” back to them to come to new conclusions.
The new system works, having led to record-breaking performance in previously unsolvable Atari games, and it has scientists excited. This has implications not only for the way machines play, but also for how they react in everyday life. They can always take a step back and – if they ever reach a dead end – rethink and start again. This will allow systems, such as those for automated texting, to work even more efficiently, and collaborative robots to cooperate even better with humans. All thanks to success in a few classic computer games.
