Markov Decision Processes

absorbing state

An absorbing state is a place you can check in to but never leave. Once the agent enters it, every action loops straight back to the same state and pays nothing more; the game is simply over. The checkmated board, the crashed car, the reached goal: each is a dead end the agent can never escape and from which no further reward flows.

This little trick is what lets an episodic task be written as a single tidy MDP. By making the terminal state absorbing with zero reward, an episode that ends is really just an infinite stay in a state where nothing happens, so the same infinite-horizon machinery and the same Bellman equations apply without special cases. In practice the bookkeeping marks the step as terminal and resets the environment for the next episode.

Also called
terminal state