Dear Vahid Ghodrati

These links might all the dots together, have a look

1. http://incompleteideas.net/book/first/ebook/node28.html

2. https://www.kdnuggets.com/2018/03/5-things-reinforcement-learning.html

3. https://wiki.pathmind.com/deep-reinforcement-learning

Kind Regards

Qamar Ul Islam

Hi. See the link that might help you:

http://incompleteideas.net/book/first/ebook/node28.html

My understanding is this boundary depends on the goals of the given task and it would be determined by the breakdown of the goals to subgoals noting the facts that how well can these subgoals be solved (i.e., subgoals have robust and efficient solutions) and what reliability you expect. Generally, if sub-goals are solvable and reliable with high probability, then it is more advantageous to convert them into actions and draw the boundary in a higher level to significantly reduce the action-space needed to explore. For instance, if the task aims to train humans to make a better decision on braking or accelerating given the current traffic on the road, then the action at the level of “brain meets body” seems appropriate. If the task aims to train human to make better traveling plan, the actions at the level of “choices of where to drive” seems appropriate. If the tasks aim to assess and optimize the quality of the rubber, the actions where “rubber meets the road” seem appropriate.