Action

Scope: An agent’s choice in a partially observable Markov decision process (POMDP), as formulated by Kaelbling, Littman, and Cassandra (1998) ¹. This is not a definition of every physical or computational action.

In the broader Agent framework, actions are how an agent acts on its environment. ² The transition probabilities in this entry belong to the specified POMDP, rather than to the general definition of an agent.

In the POMDP model

An action is one of the choices available to the decision-maker. The model specifies how the selected action and the current (possibly hidden) state affect the distribution of the next state; the agent then receives an Observation. The action is the choice, not the resulting state change or the evidence subsequently received. A useful one-step shorthand is $s\stackrel{a}{\to }{s}^{\prime }$: $s$ is the state before the step, $a$ the selected action, and $s^{\prime }$ the state afterward. The arrow represents a possible transition, not a deterministic guarantee. The paper treats action choices, transition probabilities, observations, and rewards as parts of a planning problem, not interchangeable labels.

Boundary example

A robot choosing move north has selected an action even if it slips and stays in place. Its later sensor report is an observation, not a second action. Conversely, deciding to consult a document collection can be modeled as an action in a separately specified decision process, but retrieval alone is not automatically an action in the particular POMDP defined by this paper.

Relations

• Chosen by: a Policy specifies which action to take using the information available to the agent.
• Informs: the distribution of the next Observation depends on the action as well as the resulting state.
• Updates: an action and a subsequent observation jointly constrain the next Belief state.
• Contrast: Retrieval-augmented generation incorporates retrieved text into generation; describing that retrieval as an action requires an explicit decision model beyond the original RAG formulation.

Footnotes

1. Kaelbling, Leslie Pack, Michael L. Littman, and Anthony R. Cassandra. “Planning and Acting in Partially Observable Stochastic Domains.” 1998. Artificial Intelligence 101 (1–2): 99–134. https://doi.org/10.1016/S0004-3702(98)00023-X. Author-hosted paper PDF. ↩

2. Russell, Stuart J., and Peter Norvig. Artificial Intelligence: A Modern Approach. Fourth edition, Global Edition. Pearson Education Limited, 2022. Chapter 2, §2.1, pp. 54–56. Official edition page. ↩