I am only familiar with the basics of Reinforcement Learning and came across two terms that I don't understand: path-wise cost function and the terminal cost for the given system. Are these terms analogous to those in the Bellman Equation? Attaching a snippet of the paper for context.
What is the path-wise cost function and what is the terminal cost in Double Deep Q-Network Learning?
Viewed 18
1 Answers
These are not RL terms, but rather just some arbitrary naming convention from the paper you linked. From RL perspective both these would be just called rewards, and authors choose to discriminate between a reward obtained on a decision point ("path-wise"), so rewards of form R(s_t, a), and rewards that are obtained in the final state R(s_T). Traditionally in RL we would attach rewards only to actions, and thus the "terminal cost" would just be folded to whatever is the action that leads us to terminal state. Or equivalently we would just have rewards attached to states only. The differences are really minor and mostly end up being about the preference of an author. And same here, authors preferred to give separate names for the objects, as in their storytelling and/or math it was relevant. There is no general split in RL though.
Related