Game AI perception and memory without omniscient agents

Separate truth from knowledge

A game world knows where every character is. An individual guard should know only what its designed senses, communications, or scripted knowledge provide. If decision code reads a hidden target's live transform through a remembered reference, the guard can become omniscient despite convincing sight-cone graphics.

Observations are records

Unreal's AI Perception supports sensory updates and configurable stimulus aging, including settings governing stale-data forgetting. [1] Its stimulus representation includes location, strength, age, and sensing status. [2] These fields support an observation record; they do not automatically define the game's confidence model or make signal strength a probability of truth.

Worked example

In an illustrative stealth design, a guard hears a noise at time 10 s, records its location, and assigns an initial confidence 0.8. Choose a deliberately simple design rule: confidence = max(0, 0.8 - age / 10) with age measured in seconds. At 12 s, age is 2 s and confidence is 0.6. At 16 s, confidence is 0.2.

Set an investigation threshold of 0.3. The earlier record can justify investigating; the later one cannot under this rule. These numbers are tuning choices, not experimentally calibrated probabilities. Require a current successful sight observation for Attack, even while an older sound still permits Investigate. Search around the recorded sound location rather than continuously following the source actor's hidden position.

Implementation workflow

Store source identity, sense type, observation time, observed position, and expiration policy. Distinguish event time from processing time if sensing is queued. Resolve identity safely when actors disappear, and copy relevant observation data instead of retaining unrestricted access to world truth.

Define how a new sight observation updates an old sound belief and how contradictory evidence changes action eligibility. Treat team reports as explicit observations with a sender and age. Schedule expensive sensing within a budget, but avoid interpreting an unperformed query as confirmed absence. Display the knowledge record alongside the actual world state during debugging.

Limits and behavior checks

Confidence decay does not prove that an area is safe or that a target stopped moving. A missed observation, expired memory, and explicit negative evidence are different events. Keep these distinctions visible in decision rules.

Test occlusion immediately after detection, destroyed targets, delayed reports, and a guard hearing its own effects. Inspect why each action was selected and which observation justified it. Good memory preserves believable continuity while retaining uncertainty; it should not serve as a back door to information the agent never acquired.

Sources

[1] AI Perception — Epic Games.

[2] FAIStimulus reference — Epic Games.