ddpg:

1. state and action belongs to different data, and should be input at different layers.

2. use batch normalization

3. use the same one layer for the state in actor and critic.