Performance

Documentation Unreal Engine AI Performance

What each system costs, how often it runs, and how to read stat SEC on your own hardware.

MovementEvaluatorComponent direction sampling dominates SEC CPU time. Role assignment, action polling, awareness and reactions run on timers, on events, or only while a feature is active.

Cost Model

Every SEC system runs on a known cadence over a known loop. What that costs in milliseconds depends on your CPU, so the shape is documented here and the numbers come from your own machine through stat SEC.

The slider models linear scaling from a per-enemy figure you supply. Read one enemy's cost from stat SEC in PIE, put it in, and the model tells you where a crowd lands.

Output Log

Active SEC enemies20 enemies

1 duel50 horde150 stress

Illustrative scaling — run stat SEC in your level for real numbers

MovementEvaluator - Total

0.070 ms

MovementEvaluator - Direction Scoring

0.063 ms

MovementEvaluator - Avoidance

0.004 ms

RoleSubsystem - Role Distribution

0.001 ms

ThreatDetection - Tick

0.008 ms

Estimated SEC overhead

0.083 ms

Share of 60 fps frame budget (16.6 ms)0.50%

Low share at this count

At 60 fps you have 16.6 ms per frame. Movement direction sampling dominates the plugin's share; everything else runs on a timer, on an event, or only while a feature is active.

What Runs When

Cadence is how often the work happens. The multiplier is what makes one pass more expensive.

System	Cadence	Work each pass	Multiplier
Movement evaluator	Every tick while `EvaluateAndMove` or AI Movement runs	Scores one direction per sample and picks a winner	`NumSamples` (default 16)
Avoidance	Inside the movement pass, overlap refresh at 10 Hz	Sphere overlap, then a penalty per neighbour found	Pawns packed inside the avoidance radius
Nav sampling	Inside the movement pass, cache refresh every `NavCheckInterval` (0.1 s), staggered across `NavStaggerStride` (4) frames	Projects sampled directions onto the navmesh	`NumSamples`, navmesh density
Detour query	On the detour recheck interval while hybrid movement runs	One path query compared against straight-line distance	Path length
Action scoring	`EvaluationCooldown` (0.1 s) on the native brain, or each Poll Action	Hard gates, then a score for every action still standing	Actions on the set, times scorers per action
Reaction scoring	Once per trigger rule that matches an incoming stimulus and clears its own interval, conditions and chance roll	Gates and scores the reactions that rule allows	Rules matched per stimulus, times reactions per rule
Awareness	Timer at `EvaluationInterval` (0.1 s)	Advances a meter per remembered actor and ages the rest	Actors perceived at once
Brain tick	Every tick, gated inside by `EvaluationCooldown`	Picks a phase, then drives movement and action scoring	Contains both, see the warning below
Role subsystem	`RoleReassignmentInterval` (8 s)	Fitness score per combatant per candidate role	Combatants, times roles
Threat detection	Every tick by default	One dot product against the focus actor	Fixed small cost
Melee trace	Every tick between `StartTracing` and `StopTracing`	Swept trace per socket per frame	Active trace sockets
Approach window	Every tick while an approach window is open	Steers the pawn toward its approach target	One pawn at a time
Territory	Only when something asks	Route arithmetic, or one navmesh query for a wander spot	Posts on the route
Combat tokens	On claim and release	Slot bookkeeping on the target	Fixed small cost

Territory, tokens and reactions do no work until something calls them, and the brain skips its own movement pass while it is walking a leg, waiting out a post's dwell, or looking around a spot it just reached, so an idle enemy at a post costs only the brain tick.

Tuning for Scale

Details

Movement Evaluator Component (MovementEvaluatorComponent)

Movement Evaluator

Num Samples

Enable Avoidance

Enable Nav Aware Sampling

Nav Sampling

Nav Check Interval

0.1

Num Samples scales the direction-scoring loop. Avoidance and nav sampling both refresh at 10 Hz by default.

Num Samples: Range 4–32. Halving samples roughly halves direction-scoring work. Start at 8 for horde crowds if strafing still looks acceptable.

bEnableAvoidance: Turn off when enemies never pack tightly (solo bosses, sparse open fields).

bEnableNavAwareSampling: Turn off only when you accept ledge/hazard blind spots. Keeps NavCheckInterval and NavStaggerStride from doing nav probes.

Hybrid movement: Distant enemies pathfind; close enemies use tactical sampling. HybridSwitchDistance (default 800 cm) controls the handoff. See Movement System.

Project-wide timers

Role and target timing live in Project Settings → Soulslike Enemy Combat → Combat Roles → Timing:

Timing

Role Reassignment Interval

Min Time In Role

Re-evaluate Targets On Reassignment

Role Reassignment Interval: How often UAICombatRoleSubsystem re-runs global assignment. Default 8 s. Set 0 to drive reassignment manually.

Native brain / Poll Action: EvaluationCooldown on USECBrainComponent or STTask_PollAction (default 0.1 s) spaces action picks. Raise it for background grunts that do not need 10 Hz decisions.

Threat: Call SetTickInterval on UThreatDetectionComponent to skip frames when you only need coarse player-look reactions.

Built-in Profiling

Open the console (~) and run:

stat SEC

Registered cycle stats (from STATGROUP_SEC):

Stat	Scope
`MovementEvaluator - Total`	Full per-AI movement pass, containing the four below it
`MovementEvaluator - Direction Scoring`	Sample loop over `NumSamples`
`MovementEvaluator - Avoidance`	Throttled overlap scoring
`MovementEvaluator - Nav Sampling`	Nav validity cache refresh
`MovementEvaluator - Detour Query`	Path-length detour check
`ActionEvaluation - Scoring`	Gates and scores every candidate action
`ReactionEvaluation - Scoring`	Gates and scores reactions for one stimulus
`Awareness - Evaluate`	One pass over remembered actors
`Brain - Tick`	Native brain tick, containing action scoring and the movement pass
`RoleSubsystem - Role Distribution`	Global role assignment pass
`ThreatDetection - Tick`	Per-tick threat level

Two of these are totals rather than separate costs. MovementEvaluator - Total contains the four movement stats under it, and Brain - Tick contains ActionEvaluation - Scoring and the whole movement pass whenever the native brain drives the enemy. Adding the column up counts that work two or three times. Under a StateTree the same functions run with no enclosing brain tick, so Brain - Tick reads zero and the rest stand on their own.

Melee tracing and the approach window carry no cycle stat. Measure them in Unreal Insights on USECMeleeTraceComponent and USECApproachComponent tick.

Unreal Insights: filter CPU trace events SEC_MovementEvaluator, SEC_DirectionScoring, SEC_Avoidance, SEC_NavSampling, SEC_DetourQuery, SEC_ActionScoring, SEC_ReactionScoring, SEC_Awareness, SEC_BrainTick, SEC_RoleDistribution, SEC_ThreatDetection.

For log-based diagnosis, see Debugging & Troubleshooting.

Optimizations Already in Code

Unit-circle sample cache: trigonometry for direction samples is precomputed when NumSamples changes.
Throttled avoidance: overlap queries run at 10 Hz, not every frame.
Throttled nav cache: NavCheckInterval plus NavStaggerStride spread nav sweeps across frames in a pack.
Timer-based roles: reassignment runs on RoleReassignmentInterval, not per frame.
Hybrid movement: strategic pathfinding only past HybridSwitchDistance (or while detour escalation is active).
Conditional melee tick: USECMeleeTraceComponent starts with tick disabled; tracing enables tick only for active swings.
Conditional action tick: ActionEvaluationComponent ticks only during action execution (hook dispatch), not while idle.

Scaling Guide

What dominates changes with the crowd, and that decides where tuning pays off.

Scenario	Active SEC enemies	What dominates
Soulslike duel	1 to 5	Nothing worth tuning. One enemy's movement pass is noise against a frame.
Action RPG pack	10 to 20	Direction sampling, with avoidance climbing as they close on you.
Horde	50 to 100	Direction sampling across the pack. Cutting `NumSamples` to 8 is the largest single lever.
Stress test	150+	Direction sampling and nav probes. Raise `NavStaggerStride`, and slow action polling for enemies out of the fight.

Cost scales roughly linearly with enemies running movement every frame. An enemy standing at a post, waiting out a dwell, or dead costs a fraction of one in a fight.