DeepHack.Game 5vision team
- 1. Deep Neural Networks for Atari 2600 games 5 vision team
- 2. Atari 2600 games The Atari 2600 is a home video game console released on 1977
- 3. Deep Q-learning Neural Networks for Atari games (DQN) Images and Scores as inputs No prior knowledge about game rules
- 4. Google results • DQN outperforms the best existing methods on 43 games • DQN outperforms human professional human on 29 games
- 5. • Google DQN • Pre-training stage(ISA) • Reward normalization • Minibatch sampling • Reduced fully connected layers • Drop-outs • Actor-critic (1 net for actor 1 net for critic) Our Solution
- 6. Pre-training stage • Motion images sampling • Insertion of extracted images • Subsidiary NN training • Extraction of middle layer weights • Input the weights into first layer of convolution NN
- 7. Reward normalization Reward score used for training Google Our solution Positive reward +1 In range 0.7 : 1 Zero reward 0 0 Negative reward -1 In range -1 : -0.7
- 8. Results 0% 20% 40% 60% 80% 100% 120% 140% Hero Kung-fu master Space-invaders google our model human