a human

a human

Deep Q-Network Plays Atari 2600 Pong

27m ago
SOURCE  

Description

I implemented Deep Q-Network (DQN), a deep neural network system for reinforcement learning recently proposed by Google DeepMind's researchers, using Caffe. After a day of learning, DQN (right) can successfully play Pong from raw visual inputs. In this video DQN is configured to make a completely random move with a probability of 5% for each frame, still outperforming the computer opponent (left). The scores of the three trials in the video are 16, 13 and 19, while the score of a human expert is reportedly only -3. The source code is available at https://github.com/muupan/dqn-in-the-caffe. See the original paper http://www.cs.toronto.edu/~vmnih/docs/dqn.pdf for the algorithm details.