策略学习(Policy-Based Reinforcement Learning)

1. Policy Function (a|s)

1. Policy function (a|s) is a probability density function.
2. It takes state s as input.
3. It output the probabilities for all the actions, e.g., (left|s) = 0.2; (right|s) = 0.1;(up|s) = 0.7.
4. The agent performs an action a random drawn from the distribution.

2. Policy Network (a|s,)

Policy network: Use a neural net to approximate (a|s).

1. Use policy network (a|s,) to approximate (a|s).
2. : trainable parameters of the neural net.

关于上图：

1. 
2. Here, ={"left", "right", "up"} is the set of all actions.
3. That is why we use softmax activation.

 

References

[1] https://www.youtube.com/watch?v=qI0vyfR2_Rc