[机器学习：李宏毅]28.Deep Reinforcement Learning

1、Deep Reinforcement Learning：AI=RL+DL

2、对比Supervised v.s Reinforcement

Supervised:向老师学习；

Reinforcement：从经验学习，最大化期望；（eg：Alpha GO（机器互下）:supervised+reinforcement）

3、应用方面：

Space invader

Play yourself: http://www.2600online.com/spaceinvaders.htm

• How about machine: https://gym.openai.com/evaluations/eval_Eduozx4HRyqgTCVk9ltw

4、Reinforcement Learning的难点

• Reward delay
•  Agent’s actions affect the subsequent data it receives 

5、Asynchronous Advantage Actor-Critic(A3C)

引言：基于policy和基于value（eg：Alpha GO：policy-based+value-based+model-based）

6、Policy-based Approach：learning an actor

----------------------Step1---------------------

------------------------------------Step2-------------------------------------------

    

---------------------------------------step3---------------------------------------

Add a Baseline：

7、Valued-based Approach：learning a critic

8、Actor-Critic

A3c Deepmind Demo1：https://www.youtube.com/watch?v=nMR5mjCFZCw

A3c Deepmind Demo2：https://www.youtube.com/watch?v=0xo1Ldx3L5Q