Deep Q Network

• 4.1 DQN 算法更新
• 4.2 DQN 神经网络
• 4.3 DQN 思维决策
• 4.4 OpenAI gym 环境库

Notes

Deep Q-learning Algorithm

This gives us the final deep Q-learning algorithm with experience replay:

There are many more tricks that DeepMind used to actually make it work – like target network, error clipping, reward clipping etc, but these are out of scope for this introduction.

The most amazing part of this algorithm is that it learns anything at all. Just think about it – because our Q-function is initialized randomly, it initially outputs complete garbage. And we are using this garbage (the maximum Q-value of the next state) as targets for the network, only occasionally folding in a tiny reward. That sounds insane, how could it learn anything meaningful at all? The fact is, that it does.

Extension

• Using Keras and Deep Q-Network to Play FlappyBird | Ben Lau
• Demystifying Deep Reinforcement Learning
  • The above post is a must-read for those who are interested in deep reinforcement learning.

Learning Notes: Morvan - Reinforcement Learning, Part 4: Deep Q Network