How does ML decide value for continuous action space?

Hello, I am new to Unity ML.
I tried several experiments on my ragdoll project. I attempt to make the ML agent could control the ragdoll to chase the target. Basically, this project is edited from the example project Walker.

But several experiments gave the result that my agents seldom make "Motion", such as striding the both legs to run. The agents usually seems to decide a "Form" and only use slight change to perform action. For example, in my video, the agent only use its left leg to jump, and never try to stride its right leg. And even the left leg is not making full stretching.

So I am wondering how does ML decide value for continuous action space?

The first hypothesis is it just randomly tries actions for all training and keeps learning from the results.
The second hypothesis is when agent found an action set that looks good, it would start to modify from this action set like +-0.2 on actions and makes learning from the results.

I want to ask which one is correct? or none of these hypothesis are correct?
And how could I make my agent more agile in motion instead of being rigid in specific pose?

Another question is my project only need [-1, 1] as range for action.
My project uses Math.Clamp(action, -1, 1) for all action values as project Walker.
But does it mean that agent actually tries the value that over the range of [-1, 1]?
Does this make agent explore action space less efficient? Could I limit the range of action space for ML?

Many thanks,
Sherlore

 

Thank you for your clear and useful explanation. It helps me a lot!