Observation

Scope: The perceptual outcome in the POMDP formulation of Kaelbling, Littman, and Cassandra (1998) ¹, rather than all uses of “observation” in science.

In AIMA’s broader Agent framework, a percept is the content received through sensors, and a percept sequence is the complete perceptual history. ² The POMDP observation variable below adds a particular probabilistic model; it is not a definition of every sensory input.

In the POMDP model

The decision-maker does not directly know the underlying state. After taking an Action, it receives an observation drawn according to an observation model. An observation is therefore evidence about a state, not necessarily the state itself. In this model, writing $O(o\mid s^{\prime },a)$ denotes the probability of receiving observation $o$ after action $a$ when the resulting state is $s^{\prime }$. Here $O$ is the observation model, $o$ is one possible report, $s^{\prime }$ is the resulting state, and $a$ is the chosen action; the notation presumes specified sets of states, actions, and observations and a probability distribution over reports. It does not assert that every sensor behaves this way in practice.

Boundary example

If a navigation sensor says “wall nearby,” that report is an observation even if the wall is absent because of sensor noise. A map coordinate inferred from several such reports is an estimate, not another raw observation. Equating the report with the hidden location would discard the very uncertainty that partial observability models.

Relations

• Follows: an Action precedes the observation in the modeled step.
• Conditions: the new observation, together with the action, is used to update a Belief state.
• Guides indirectly: a Policy over belief states can use new evidence without equating a report to the true state.
• Contrast: text fetched by Retrieval-augmented generation is retrieved context for a language model; calling it a POMDP observation additionally requires a defined hidden state and observation model.

Related reading: Context window distinguishes the input available to an LLM from information stored elsewhere.

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. ↩