A set of reinforcement learning algorithms have been shown to be better at playing classic video games than human gamers or other artificial intelligence systems .
The algorithms have been developed by a team of researchers from Uber AI Labs in San Francisco.
Reinforcement learning algorithms learn to do things by synthesizing the information provided by a large set of data: they recognize patterns and use them to make guesses about new data. But such algorithms tend to have problems when they encounter data that doesn’t match other data. Problems that have been corrected in this new development.
To do this, they have added an algorithm that remembers all the paths that a previous algorithm has taken while trying to solve a problem. When it finds a data point that doesn’t appear to be correct, it goes back to its memory map and tries another route.
The researchers tested their new approach by adding rules from a video game and a goal : score as many points as possible and try to achieve a higher score each time. They then used their system to play 55 Atari games . The new system beat other artificial intelligence systems 85.5% of the time. It did particularly well in Montezuma’s Revenge , scoring higher than any other artificial intelligence system and even breaking the record for a human.
The researchers believe their algorithm could be carried over to other applications, such as image or language processing by robots.