Record-breaking run from Go-Explore on Montezuma’s Revenge: Scores over 2 million points! скачать в хорошем качестве

Record-breaking run from Go-Explore on Montezuma’s Revenge: Scores over 2 million points!

After robustification, Go-Explore has generated a robust, deep neural network policy that can reach level 118 and score over 2,000,000 points on Montezuma’s Revenge, beating both the AI and human world record. A full description is here: eng.uber.com/go-explore Solving the Atari games Montezuma’s Revenge and Pitfall has been a long-standing AI grand challenge. These games represent a broad class of challenging, real-world problems called “hard-exploration problems,” where an agent has to learn complex tasks with very infrequent or deceptive feedback. Previous state-of-the-art algorithms obtained at most 17,500 points on Montezuma's Revenge (barely solving level 1) and no algorithm has managed to score even a single point on Pitfall. This video introduces Go-Explore, a new family of algorithms capable of achieving scores over 2,000,000 on Montezuma’s Revenge and over 23,000 on Pitfall! Go-Explore reliably solves all three unique levels of Montezuma’s Revenge, and then generalizes to the nearly-identical subsequent levels (which only differ in the timing of events and the score on the screen). We have even seen it reach level 159, although that was a different network that did not get as high of a score as the one in the video above. Authors: Adrien Ecoffet, Joost Huizinga, Joel Lehman, Kenneth O. Stanley, and Jeff Clune, all from Uber AI Labs