Does a percept sequence have to be finite?

At any finite interaction step, the accumulated sequence is finite, though the agent may continue receiving observations indefinitely.

Is a sliding window the full percept sequence?

No. It is a selected recent portion used by an implementation.

Can an agent ignore earlier percepts?

Yes. A simple reflex design bases its action on the current percept, which may be adequate for some tasks.

AI Glossary / Graphs and symbolic AI / Beginner

Percept sequence

Definition

A percept sequence is the ordered history of observations an agent has received up to a particular point. It allows agent behavior to depend on earlier inputs as well as the latest observation.

What it means

The current percept answers what the agent has just observed. The percept sequence adds what it observed before and in what order. That difference is essential when one observation cannot reveal the situation relevant to a decision.

For example, a level reading of 40 units may be normal after readings of 38 and 39, but concerning after readings of 70 and 55. The same latest value occurs in different histories. Interpretation depends on the task and model, not merely on the presence of a trend.

The standard agent-function formulation maps percept histories to actions. This is a specification of behavior; it does not require an implementation to retain an unlimited verbatim recording.

How it works

At each observation step, a new percept extends the existing sequence. If the history was (p1, p2), receiving p3 produces (p1, p2, p3). Order and repeated observations matter: the sequence is not an unordered set.

An implementation may store the full history, a limited window, a recurrent representation or an internal state updated from observations and previous actions. The stored state should preserve information needed for the decisions the agent must make. Compression can lose relevant distinctions if the design is poor.

A percept sequence contains observations. Action history is conceptually separate, although an agent often needs both to interpret a changing environment. In a deterministic agent, earlier actions may be reconstructable from its policy and earlier percepts; other implementations record them explicitly.

In asynchronous systems, arrival order, observation time and missing intervals also need a defined interpretation.

Three practical examples

1. Equipment monitoring: A controller uses successive vibration readings to distinguish a short spike from a persistent change. The historical pattern provides context that the newest measurement alone lacks.

2. Interactive assistance: An agent receives a correction after an earlier request. Remembering that sequence helps it interpret the latest short instruction without treating it as an isolated task.

3. Mobile robotics: A robot combines successive observations with its motion history to update its location estimate. Repeatedly seeing a similar corridor does not guarantee that it has returned to the same physical location.

Think of it this way

Think of a percept sequence as the ordered entries in an observation notebook. Reading only the last line may hide a change that is obvious from several pages.

The analogy has a limit: an agent need not literally preserve a notebook. It may retain a compact state that supports its decisions, and the notebook's contents are still observations rather than an infallible account of reality.

Worked example

A toy alerting agent receives one level reading per minute. Its rule is to alert when the latest three readings show two consecutive decreases of at least 10 units each.

History A ends with (70, 55, 40). The decreases are 70 − 55 = 15 and 55 − 40 = 15. Both meet the threshold, so the rule alerts.

History B ends with (38, 39, 40). The changes are increases of one unit, so the rule does not alert. Both histories have the same current percept, 40, yet lead to different actions.

The agent need store only the latest three readings for this particular rule. That window would be insufficient for a different rule requiring a daily baseline. The example establishes why history matters; it does not show that the chosen thresholds diagnose a real equipment fault.

Illustrative case study

In a fictional project, Maple Service deploys an assistant that repeatedly asks a customer for information already supplied. The implementation passes only the latest message into its decision process, losing earlier answers and corrections.

The team introduces a structured session state that records relevant observations and which values supersede older ones. It tests conversations with repeated, corrected and contradictory inputs rather than assuming all earlier statements remain current.

The case illustrates a memory-design requirement. Retaining more history alone does not guarantee better decisions; the system must interpret sequence, scope and changes correctly.

Common misconceptions and mistakes

A percept sequence is just the latest percept. It includes the observation history up to that point.
Observation order is irrelevant. Reordering readings can change the situation they describe.
A history-dependent agent must store everything forever. A suitable internal state may summarize what the task needs.
The percept sequence automatically includes actions. Observations and actions are distinct, even when both are logged together.
More history always helps. Stale, misleading or irrelevant observations can complicate inference and require careful handling.

Business use

History requirements should be part of workflow design. Decide which past inputs affect present decisions, how corrections are represented and when a session or operational context ends.

For monitoring, preserve timestamps and identify gaps so that a sequence of three messages is not mistakenly treated as three equally spaced measurements. For interactive systems, distinguish a changed instruction from an additional independent request. For robotics, record actions needed to interpret how observations evolved.

Testing should include pairs of histories with the same final input but different appropriate responses. Such tests reveal whether the chosen memory representation preserves distinctions that matter, rather than merely demonstrating that a system can retrieve an earlier sentence.

Check your understanding

Why does the alerting agent act differently on the two histories even though both end at 40?

Answer: Its rule uses the latest three observations. One history contains two sufficiently large decreases; the other does not.

Related terms

Percept
Agent function
Internal state
Belief state
memory
recurrent model
partial observability
temporal reasoning

Learn next

Design two observation histories that end with the same input but require different actions in your chosen task. Identify the smallest useful state that distinguishes them, then test whether a correction or missing observation breaks that representation.

Sources

Russell and Norvig formulate agent functions over percept histories and contrast reflex agents with agents maintaining state: https://aima.cs.berkeley.edu/slides-pdf/chapter02.pdf

The level-alert calculation and Maple Service case are original illustrations. The alerting rule is educational and is not an equipment-monitoring recommendation.

Previous: Percept schema Next: Perception