Skip to content


Detailed breakdown of networking character movement:


Code Comments GameplayPrediction.h#

Overview of Gameplay Ability Prediction#

High Level Goals#

At the GameplayAbility level (implementing an ability) prediction is transparent. An ability says "Do X->Y->Z", and we will automatically predict the parts of that that we can.
We wish to avoid having logic such as "If Authority: Do X. Else: Do predictive version of X" in the ability itself.

At this point, not all cases are solved, but we have a very solid framework for working with client side prediction.

When we say "client side prediction" we really mean client predicting game simulation state. Things can still be 'completely client side' without having to work within a prediction system.
For example, footsteps are completely client side and never interact with this system. But clients predicting their mana going from 100 to 90 when they cast a spell is 'client side prediction'.

What do we currently predict?

  • Ability activation

  • Triggered Events

  • GameplayEffect application:

  • Attribute modification (EXCEPTIONS: Executions do not currently predict, only attribute modifiers)
  • GameplayTag modification

  • Gameplay Cue events (both from within predictive gameplay effect and on their own)

  • Montages

  • Movement (built into UE4 UCharacterMovement)

Some things we don't predict (most of these we potentially could, but currently dont):

  • GameplayEffect removal
  • GameplayEffect periodic effects (dots ticking)

Problems we attempt to solve:

  1. "Can I do this?" Basic protocol for prediction.
  2. "Undo" How to undo side effects when a prediction fails.
  3. "Redo" How to avoid replaying side effects that we predicted locally but that also get replicated from the server.
  4. "Completeness" How to be sure we /really/ predicted all side effects.
  5. "Dependencies" How to manage dependent prediction and chains of predicted events.
  6. "Override" How to override state predictively that is otherwise replicated/owned by the server.

DaveRatti Networking Explanation#


  • initiated by client
  • rpc
  • create side effeccts => replicate back as properties
  • relies on replicated properties
  • Dependent predictionkey
  • Fortnight/paragon doesnt really use it


  • => Can I do it?
  • <= server says YUP! you did
  • Predictionkey hasn't come back yet, so you still have to come wait for state

Gameeffect applied

  • => stores delegates
  • =>


  • How to turn stuff into gampelaycueparameters
    virtual void InitGameplayCueParameters(FGameplayCueParameters& CueParameters, const FGameplayEffectSpecForRPC &Spec);
    virtual void InitGameplayCueParameters_GESpec(FGameplayCueParameters& CueParameters, const FGameplayEffectSpec &Spec);
    virtual void InitGameplayCueParameters(FGameplayCueParameters& CueParameters, const FGameplayEffectContextHandle& EffectContext);

FGameplayAbilityRepAnimMontage RepAnimMontageInfo

Gameplay Ability System

  • Very lightweight/stellar
  • Higher level tasks is super unpredictable
  • Sometimes flexible but lots of gremlins
  • Replication mechanism:
    • Able to predict very well but also very lightweight
  • Works very well for melee combat
    • At carbine, very twitchy melee combat system
    • Also at nexon
    • For hitscan + melee weapns
    • Thrown weapons
  • Pretty robust features but with cryptic names
    • Some nodes are super heavy (surprise!)
  • Fundemntal issues
    • GAS is a giant iceberg/little exposed
    • Little docs
  • Test:

    • 3 hit combo
    • Can be separate ability.

      • Start|Active|Windown
      • Sword Stab => Sword Swing => Smash
      • GAS
      • Bind On Input => Input == punch => Start Ability
      • SowrdStab Ability
        • TryActivate (do validation; runs on both)
        • Commit Ability (ability now running; means begin)
        • Point of no rollback
        • Play Montage and wait
        • OnNotifyRecieved
          • Start Wait for input for 10 frames
          • If no input => end ability
          • If there was input =>
          • PlayMontage And Wait: Sword Swing
        • Wait Target Node: hjit detection
        • if hits, do hit stuff/effect/cue
      • When this ability ends, schedule hit combo2
      • Can schedule cooldown
    • Charged attack

    • Input cancel
    • Dash/Movement
    • Misprediction?


  • What were issues
  • Character movement
  • Network
  • Debug tooling?

GameAbility Prototype:

  • 3 hit combo ()
  • Charged Attack
  • Dash (moving)
  • Charged (moving)

NetLag ~200-300ms

Implementation Details#


A fundamental concept in this system is the Prediction Key (FPredictionKey). A prediction key on its own is simply a unique ID that is generated in a central place on the client. The client will send his prediction key to the server, and associate predictive actions and side effects with this key. The server may respond with an accept/reject for the prediction key, and will also associate the server-side created side effects with this prediction key.

(IMPORTANT) FPredictionKey always replicate client -> server, but when replicating server -> clients they *only replicate to the client that sent the prediction key to the server in the first place.
This happens in FPredictionKey::NetSerialize. All other clients will receive an invalid (0) prediction key when a prediction key sent from a client is replicated back down through a replicated property.

Ability Activation#

Ability Activation is a first class predictive action. Whenever a client predictively activates an ability, he explicitly asks the server and the server explicitly responds. Once an ability has been predictively activated, the client has a valid 'prediction window' where predictive side effects can happen which are not explicitly 'asked about'. (E.g., we do not explicitly ask 'Can I decrement mana, Can I put this ability on cooldown. Those actions are considered logically atomic with activating an ability).

AbilitySystemComponent provides a set of functions for communicating ability activation between clients and server: TryActivateAbility -> ServerTryActivateAbility -> ClientActivateAbility(Failed/Succeed).

  1. Client calls TryActivateAbility which generates a new FPredictionKey and calls ServerTryActivateAbility.

  2. Client continues (before hearing back from server) and calls ActivateAbility with the generated PredictionKey associated with the Ability's ActivationInfo.

  3. Any side effects that happen /before the call to ActivatAbility finish/ have the generated FPredictionKey associated with them.

  4. Server decides if the ability really happened in ServerTryActivateAbility, calls ClientActivateAbility(Failed/Succeed) and sets UAbilitySystemComponent::ReplicatedPredictionKey to the generated key that was sent.

  5. If client receives ClientAbilityFailed, he immediately kills the ability and rolls back side effects that were associated with the prediction key.

  • 'Rolling back' is accomplished via FPredictionKeyDelegates and FPredictionKey::NewRejectedDelegate/NewCaughtUpDelegate/NewRejectOrCaughtUpDelegate.

  • Registering the callback in TryActivateAbility:
    // If this PredictionKey is rejected, we will call OnClientActivateAbilityFailed.
    ThisPredictionKey.NewRejectedDelegate().BindUObject(this, &UAbilitySystemComponent::OnClientActivateAbilityFailed, Handle, ThisPredictionKey.Current);

  • Invoking the callback in ClientActivateAbilityFailed_Implementation:

  1. If accepted, client must wait until property replication catches up (the Succeed RPC will be sent immediately, property replication will happen on its own). Once the ReplicatedPredictionKey catches up to the
    key used previous steps, the client can undo his predictive side effects. See UAbilitySystemComponent::OnRep_PredictionKey.

GameplayEffect Prediction#

GameplayEffects are considered side effects of prediction and are not explicitly asked about.

  1. GameplayEffects are only applied on clients if there is a valid prediction key. (If no prediction key, it simply skips the application on client).
  2. Attributes, GameplayCues, and GameplayTags are all predicted if the GameplayEffect is predicted.
  3. When the FActiveGameplayEffect is created, it stores the prediction key (FActiveGameplayEffect::PredictionKey)
    - Instant effects are explained below in "Attribute Prediction".
  4. On the server, the same prediction key is also set on the server's FActiveGameplayEffect that will be replicated down.
  5. As a client, if you get a replicated FActiveGameplayEffect with a valid prediction key on it, you check to see if you have an ActiveGameplayEffect with that same key, if there is match, we do not apply the 'on applied' type of logic, e.g., GameplayCues. The solves the "Redo" problem. However we will have 2 of the 'same' GameplayEffects in our ActiveGameplayEffects container, temporarily:
  6. At the same time, UAbilitySystemComponent::ReplicatedPredictionKey will catch up and the predictive effects will be removed. When they are removed in this case, we again check PredicitonKey and decide
    if we should not do the 'On Remove' logic / GameplayCue.

At this point, we have effectively predicted a gameplay effect as a side effect and handled the 'Undo' and 'Redo' problems.

Attribute Prediction#

Since attributes are replicated as standard uproperties, predicting modification to them can be tricky ("Override" problem). Instantaneous modification can be even harder since these are non stateful by nature.
(E.g., rolling back an attribute mod is difficult if there is no book keeping past the modification). This makes the "Undo" and "Redo" problem also hard in this case.

The basic plan of attack is to treat attribute prediction as delta prediction rather than absolute value prediction. We do not predict that we have 90 mana, we predict that we have -10 mana from the server value, until
the server confirms our prediction key. Basically, treat instant modifications as /infinite duration modifications/ to attributes while they are done predictively. The solves "Undo" and "Redo".

For the "override" problem, we can handle this in the properties OnRep by treating the replicated (server) value as the 'base value' instead of 'final value' of the attribute, and to
reaggregate our 'final value' after a replication happens.

  1. We treat predictive instant gameplay effects as infinite duration gamepaly effects. See UAbilitySystemComponent::ApplyGameplayEffectSpecToSelf.
  2. We have to always receive RepNotify calls on our attributes (not just when there is a change from last local value, since we will predict the change ahead of time). Done with REPNOTIFY_Always.
  3. In the attribute RepNotify, we call into the AbilitySystemComponent::ActiveGameplayEffects to update our 'final value' give the new 'base value'. the GAMEPLAYATTRIBUTE_REPNOTIFY can do this.
  4. Everything else will work like above (GameplayEffect prediction) : when the prediction key is caught up, the predictive GameplayEffect is removed and we will return to the server given value.


void UMyHealthSet::GetLifetimeReplicatedProps(TArray< FLifetimeProperty > & OutLifetimeProps) const


void UMyHealthSet::OnRep_Health()

Gameplay Cue Events#

Outside of GameplayEffects which are already explained, Gameplay Cues can be activated on their own. These functions (UAbilitySystemComponent::ExecuteGameplayCue etc) take network role and prediction keys into account.

  1. In UAbilitySystemComponent::ExecuteGameplayCue, if authority then do the multicast event (with replication key). If non authority but w/ a valid prediction key, predict the GameplayCue.
  2. On the receiving end (NetMulticast_InvokeGameplayCueExecuted etc), if there is a replication key, then don't do the event (assume you predicted it).

Remember that FPredictionKeys only replicate to the originating owner. This is an intrinsic property of FReplicationKey.

Triggered Data Prediction#

Triggered Data is currently used to activate abilities. Essentially this all goes through the same code path as ActivateAbility. Rather than the ability being activated from input press, it is activated from
another game code driven event. Clients are able to predictively execute these events which predictively activate abilities.

There are some nuances to however, since the server will also run the code that triggers events. The server won't just wait to hear from the client. The server will keep a list of triggered abilities that have been
activated from a predictive ability. When receiving a TryActivate from a triggered ability, the server will look to see if /he/ has already run this ability, and respond with that information.

There is work left to do on Triggered Events and replication. (explained at the end).

Advanced topic#


We can have situations such as "Ability X activates and immediately triggers an event which activates Ability Y which triggers another Ability Z". The dependency chain is X->Y->Z.
Each of those abilities could be rejected by the server. If Y is rejected, then Z also never happened, but the server never tries to run Z, so the server doesn't explicitly decide 'no Z can't run'.

To handle this, we have a concept of a Base PredictionKey, which is a member of FPredictionKey. When calling TryActivateAbility, we pass in the current PredictionKey (if applicable).
That prediction key is used as the base for any new prediction keys generated. We build a chain of keys this way, and can then invalidate Z if Y is rejected.

This is slightly more nuanced though. In the X->Y->Z case, the server will only recieve the PredictionKey for X before trying to run the chain himself
E.g., he will TryActivate Y and Z with the original prediction key sent to him from the client.
Where as the client will generate a new PredictionKey each time he calls TryActivateAbility. The client *has to generate a new PRedictionKey for each ability activate, since each activate is not logically atomic. Each side effect produced in the chain of events has to have a unique PredictionKey. We cannot have GameplayEffects produced in X have the same PredictionKey produced in Z.

To get around this, The prediction key of X is considered the Base key for Y and Z. The dependancy from Y to Z is kept completely client side, which is done in by FPredictionKeyDelegates::AddDependancy. We add delegates
to reject/catchup Z if Y rejected/confirmed.

This dependency system allows us to have multiple predictive actions that are not logically atomic within a single prediction window/scope.

Additional Prediction Windows (within an Ability)#

As stated, A prediction key is only usable during a single logical scope. Once ActivateAbility returns, we are essentially done with that key. If the ability is waiting on an external event or timer, by the time it returns, we will have gotten a confirm/reject from the server. Any side effects produced after this will no longer be tied to the lifespan of the original key.

This isn't that bad, except that abilities will sometimes want to react to player input. For example, 'a hold down and charge' ability wants to instantly predict some stuff when the button is released. It is possible to create a new prediction window within an ability with FScopedPredictionWindow.

FScopedPredictionWindows provides a way to send the server a new prediction key and have the server pick up and use that key within the same logical scope.

UAbilityTask_WaitInputRelease::OnReleaseCallback is a good example. The flow of events is as followed:

  1. Client enters UAbilityTask_WaitInputRelease::OnReleaseCallback and starts a new FScopedPredictionWindow. This creates a new prediction key for this scope (FScopedPredictionWindow::ScopedPredictionKey).
  2. Client calls AbilitySystemComponent->ServerInputRelease which passes ScopedPrediction.ScopedPredictionKey as a parameter.
  3. Server runs ServerInputRelease_Implementation which takes the passed in PredictionKey and sets it as UAbilitySystemComponent::ScopedPredictionKey with an FScopedPredictionWindow.
  4. Server runs UAbilityTask_WaitInputRelease::OnReleaseCallback /within the same scope/
  5. When the server hits the FScopedPredictionWindow in ::OnReleaseCallback, it gets the prediction key from UAbilitySystemComponent::ScopedPredictionKey. That is now used for all side effects within this logical scope.
  6. Once the server ends this scoped prediction window, the prediction key used is finished and set to ReplicatedPredictionKey.
  7. All side effects created in this scope now share a key between client and server.

The key to this working is that ::OnReleaseCallback calls ::ServerInputRelease which calls ::OnReleaseCallback on the server. There is no room for anything else to happen and use the given prediction key.

While there is no "Try/Failed/Succeed" calls in this example, all side effects are procedurally grouped/atomic. This solves the "Undo" and "Redo" problems for any arbitrary function calls that run on the server and client.

Unsupported / Issues/ Todo#

Triggered events do not explicitly replicate. E.g., if a triggered event only runs on the server, the client will never hear about it. This also prevents us from doing cross player/AI etc events.
Support for this should eventually be added and it should follow the same pattern that GameplayEffect and GameplayCues follow (predict triggered event with a prediction key, ignore the RPC event if it has a prediction key).

Predicting "Meta" Attributes such as Damage/Healing vs "real" attributes such as Health#

We are unable to apply meta attributes predictively. Meta attributes only work on instant effects, in the back end of GameplayEffect (Pre/Post Modify Attribute on the UAttributeSet). These events are not called when applying duration-based gameplay effects. E.g., a GameplayEffect that modifies damage for 5 seconds doesn't make sense.

In order to support this, we would probably add some limited support for duration based meta attributes, and move the transform of the instant gameplay effect from the front end (UAbilitySystemComponent::ApplyGameplayEffectSpecToSelf)
to the backend (UAttributeSet::PostModifyAttribute).

Predicting ongoing multiplicitive GameplayEffects#

There are also limitations when predicting % based gameplay effects. Since the server replicates down the 'final value' of an attribute, but not the entire aggregator chain of what is modifying it, we may run into cases where
the client cannot accurately predict new gameplay effects.

For example:

  • Client has a perm +10% movement speed buff with base movement speed of 500 -> 550 is the final movement speed for this client.
  • Client has an ability which grants an additional 10% movement speed buff. It is expected to *sum the % based multipliers for a final 20% bonus to 500 -> 600 movement speed.
  • However on the client, we just apply a 10% buff to 550 -> 605.

This will need to be fixed by replicating down the aggregator chain for attributes. We already replicate some of this data, but not the full modifier list. We will need to look into supporting this eventually.

"Weak Prediction"#

We will probably still have cases that do not fit well into this system. Some situations will exist where a prediction key exchange is not feasible. For example, an ability where any one that player collides with/touches
receives a GameplayEffect that slows them and their material blue. Since we can't send Server RPCs every time this happens (and the server couldn't necessarily handle the message at his point in the simulation), there is no
way to correlate the gameplay effect side effects between client and server.

One approach here may be to think about a weaker form of prediction. One where there is not a fresh prediction key used and instead the server assumes the client will predict all side effects from an entire ability. This would
at least solve the "redo" problem but would not solve the "completeness" problem. If the client side prediction could be made as minimal as possible - for example only predicting an initial particle effect rather than
predicting the state and attribute change - then the problems get less severe.

I can envision a weak prediction mode which is what (certain abilities? All abilities?) fall back to when there is no fresh prediction key that can accurately correlate side effects. When in weak prediction mode, perhaps
only certain actions can be predicted - for example GameplayCue execute events, but not OnAdded/OnRemove events.

FPredictionKey is a generic way of supporting Clientside Prediction in the GameplayAbility system.
A FPredictionKey is essentially an ID for identifying predictive actions and side effects that are
done on a client. UAbilitySystemComponent supports synchronization of the prediction key and its side effects
between client and server.

Essentially, anything can be associated with a PredictionKey, for example activating an Ability.
The client can generates a fresh PredictionKey and sends it to the server in his ServerTryActivateAbility call.
The server can confirm or reject this call (ClientActivateAbilitySucceed/Failed).

While the client is predictively his ability, he is creating side effects (GameplayEffects, TriggeredEvents, Animations, etc).
As the client predicts these side effects, he associates each one with the prediction key generated at the start of the ability

If the ability activation is rejected, the client can immediately revert these side effects.
If the ability activation is accepted, the client must wait until the replicated side effects are sent to the server.
(the ClientActivatbleAbilitySucceed RPC will be immediately sent. Property replication may happen a few frames later).
Once replication of the server created side effects is finished, the client can undo his locally predictive side effects.

The main things FPredictionKey itself provides are:

  • Unique ID and a system for having dependant chains of Prediction Keys ("Current" and "Base" integers)
  • A special implementation of ::NetSerialize *** which only serializes the prediction key to the predicting client ***
  • This is important as it allows us to serialize prediction keys in replicated state, knowing that only clients that gave the server the prediction key will actually see them!