Foundations of Deep Reinforcement Learning: Theory and Practice in Python

15. Actions

An action is an output from an agent that changes an environment by causing it to transition into the next state. A state is perceived; an action is actuated.

Action design is important because it gives an agent the ability to change its environment. How actions are designed affects whether the control of a system is easy or hard, and therefore directly impacts the difficulty of the problem. A particular control design may make sense to one person but not another. Fortunately, there often exist multiple ways to perform the same action. For example, the transmission of a car can be controlled manually or automatically, but people usually find automatic transmission easier to use.

Many of the lessons from Chapter 14 also apply to action ...