For a while now, Google's company DeepMind has been working on an artificial intelligence (AI) which plays Atari games better than you remember your older brother playing them in the 1980s. The AI is not only extremely proficient at playing these cult classics but has also learned to play 49 of them completely on its own. Despite this impressive feat, the DeepMind's creation isn't perfect and some games have simply proved to be too complicated for it to learn them on its own, Montezuma's Revenge being one of them. However, the Google-owned company has recently been hard at work correcting the flaws in its AI which has finally mastered the unforgiving 1984 platformer developed by the now-defunct Utopia Software.
As its developers explain it, they had to make the AI "curious enough" for it to want to actually win the game. How does one program artificial curiosity? According to DeepMind – by applying the principles of conditional teaching and learning. More specifically, the developers have started handing out rewards to the AI whenever it decided to explore more of the Montezuma's Revenge 8-bit world. Initially, the AI has only managed to beat the first level of the game and would always fail during the second one. However, after the process of "incentivizing" it, the DeepMind's creation completed the first level in only four tries and has reached the 15th out of 24 levels of Montezuma's Revenge in total.
As the developers revealed in their latest study, their reward system has "significantly improved" the AI's exploration capabilities in quite a large number of challenging games. While you may not remember Montezuma's Revenge as a notoriously hard game, especially in the context of other video games of that era, the platforming genre is actually one of the toughest ones for AIs to tackle due to the fact that no one has yet come up with a reliable method to make the computer plan for traps in advance on its own.
How does one even reward an AI? You can find that out by checking out the video below in which DeepMind's artificial intelligence is playing the aforementioned Atari platformer. For the unaware, this same AI is also the one which has recently managed to beat the South Korean Go grandmaster Lee Sedol in an extremely impressive manner, triumphing in four games out of five.