Google's DeepMind retackles Atari games
After beating opponents in Chess, Go and even at cards, Google is preparing its DeepMind artificial intelligence (AI) to beat 49 classic Atari games thanks to a new algorithm.
Enter the Elastic Weight Consolidation (EWC) algorithm, which essentially allows DeepMind to retain the knowledge of games that it has beaten.
Despite having beaten these 49 game before in February 2015, the researchers found themselves having to reprogram DeepMind for each of the games and found that the AI quickly forgot how to beat a game when they moved on to the next challenge.
This problem was dubbed ‘catastrophic forgetting’; this occurs because AIs like DeepMind are only capable of learning rules and tasks when they are presented all at once. A normal human mind can learn a new task sequentially.
The usage of this new algorithm promises to help researches teach programs to learn in a more flexible manner and also helps us understand better how a human mind works.
Source: The Verge