neat/mutation/flow

Flow helpers for the mutation root orchestration.

This chapter owns the per-genome mutation loop: initialize adaptive state, resolve effective rates and counts, dispatch operators, and keep operator statistics in sync with the actual structural outcome.

The root mutation/ chapter explains the whole-generation view. This file explains what happens once one genome reaches the front of that queue. The lifecycle is intentionally split into four small stages:

1. bootstrap per-genome adaptive state,
2. decide whether this genome should mutate at all and how many attempts it receives,
3. ask select/ for one concrete operator at a time and apply it,
4. record the structural outcome so later adaptive policy can learn which operators are actually producing growth.

That separation matters because mutation is not just "roll randomness and edit topology." The controller is trying to preserve several invariants at once: adaptive per-genome settings should remain stable across generations, structural operators should reuse innovation-aware helpers, caches should be invalidated only when topology really changed, and operator statistics should reflect actual outcomes rather than mere selection attempts.

Read this chapter in execution order:

• mutateGenome() for the complete per-genome story,
• initializeAdaptiveMutation(), resolveEffectiveRate(), and resolveEffectiveAmount() for the gating setup,
• selectConcreteMutationMethod() plus applyMutationOperator() for operator dispatch,
• captureStructuralSizes() and updateOperatorStatsIfNeeded() for the feedback loop that supports later operator adaptation and bandit policy.

flowchart TD
  Genome[One genome enters mutation flow] --> Init[Bootstrap adaptive state]
  Init --> Gate{Mutate this genome now?}
  Gate -->|no| Skip[Leave genome unchanged]
  Gate -->|yes| Amount[Resolve mutation attempts]
  Amount --> Select[Choose one concrete operator]
  Select --> Apply[Apply structural or local mutation]
  Apply --> Repair[Optional extra connection and cache upkeep]
  Repair --> Stats[Record structural outcome for operator stats]

neat/mutation/flow/mutation.flow.ts

applyAddConnMutation

applyAddConnMutation(
  genome: GenomeWithMetadata,
  internal: NeatControllerForMutation,
  methods: { mutation: unknown; },
): void

Apply an ADD_CONN mutation with reuse and weight nudging.

This is the connection-growth companion to applyAddNodeMutation(). The structural edit itself is delegated to the reuse-aware connection helper so identical edge discoveries can still share innovation identity across the population.

Parameters:

• genome - Genome to mutate.
• internal - NEAT controller context.
• methods - Mutation methods module.

Returns: Nothing.

applyAddNodeMutation

applyAddNodeMutation(
  genome: GenomeWithMetadata,
  internal: NeatControllerForMutation,
  methods: { mutation: unknown; },
): Promise<void>

Apply an ADD_NODE mutation with reuse and weight nudging.

The add-node path is special because it needs both innovation-aware structural growth and a post-split weight nudge that makes the topology change observable immediately in downstream behavior and tests. The helper intentionally treats cache invalidation as part of the operation rather than leaving it to callers.

Parameters:

• genome - Genome to mutate.
• internal - NEAT controller context.
• methods - Mutation methods module.

Returns: Promise resolving after the add-node operation completes.

applyDeterministicWeightNudge

applyDeterministicWeightNudge(
  genome: GenomeWithMetadata,
  internal: NeatControllerForMutation,
  mutationMethod: MutationMethod,
): void

Apply the post-structural weight nudge using the controller RNG.

The standalone network mutation helpers are allowed to own their own random streams, but NEAT replay needs these follow-up weight changes to come from the controller-owned RNG so the same checkpoint resumes identically.

Parameters:

• genome - Genome whose connection weight should be nudged.
• internal - NEAT controller owning the deterministic RNG.
• mutationMethod - MOD_WEIGHT descriptor providing the delta range.

Returns: Nothing.

applyMutationOperator

applyMutationOperator(
  genome: GenomeWithMetadata,
  mutationMethod: MutationMethod,
  internal: NeatControllerForMutation,
  methods: { mutation: unknown; },
): Promise<void>

Apply a mutation operator to a genome and invalidate caches as needed.

This helper is the dispatch hinge between policy and topology. Structural growth operators such as ADD_NODE and ADD_CONN are routed through the innovation-reuse helpers because the controller cares about more than local graph edits; it also needs stable innovation history for later crossover and speciation.

Non-structural operators stay delegated to the genome's own mutate() implementation. Cache invalidation is then handled separately so the flow can keep the expensive cleanup targeted to methods that plausibly changed the structural view of the genome.

Parameters:

• genome - Genome to mutate.
• mutationMethod - Mutation operator to apply.
• internal - NEAT controller context.
• methods - Mutation methods module.

Returns: Promise resolving after the operator has been applied.

captureStructuralSizes

captureStructuralSizes(
  genome: GenomeWithMetadata,
): { beforeNodes: number; beforeConns: number; }

Capture structural sizes used to evaluate operator success.

Operator adaptation in this subtree is intentionally coarse-grained: it asks whether an attempted mutation increased structural size, not whether the resulting genome later scored better. This helper records the pre-mutation node and connection counts that make that local success signal possible.

Parameters:

• genome - Genome to inspect.

Returns: Structural size snapshot.

initializeAdaptiveMutation

initializeAdaptiveMutation(
  genome: GenomeWithMetadata,
  internal: NeatControllerForMutation,
): void

Initialize per-genome adaptive mutation parameters if configured.

Adaptive mutation is stateful at the genome level, not just a controller default. This helper assigns the first persistent rate and optional amount so later generations can treat the genome as carrying its own mutation budget.

The helper only writes when the genome has not yet been initialized. That is why it runs at the top of mutateGenome() on every pass: it behaves like a cheap bootstrap check rather than a repeated reset of evolved mutation behavior.

Parameters:

• genome - Genome to initialize.
• internal - NEAT controller context.

Returns: Nothing.

maybeAddExtraConnection

maybeAddExtraConnection(
  genome: GenomeWithMetadata,
  internal: NeatControllerForMutation,
): void

Optionally add an extra connection to increase exploration.

This small post-operator hook gives the controller one extra chance to add connectivity after the main mutation has landed. It is intentionally probabilistic and lightweight: the flow uses it as a gentle exploration bump, not as a second full operator-selection phase.

Parameters:

• genome - Genome to mutate.
• internal - NEAT controller context.

Returns: Nothing.

mutateGenome

mutateGenome(
  genome: GenomeWithMetadata,
  internal: NeatControllerForMutation,
  methods: { mutation: unknown; },
): Promise<void>

Mutate a single genome based on configured mutation policies.

This is the orchestration entry point for the flow/ chapter. It keeps the per-genome lifecycle linear: adaptive settings are prepared first, then one probability gate decides whether work happens at all, and only then does the helper loop ask select/ for concrete operators.

The important design choice is that mutation amount is resolved after the mutate-or-skip gate passes. That keeps genomes with low effective rates from paying the full operator-selection cost every generation while still letting successful genomes perform multiple edits once they have been admitted into the flow.

Each loop iteration captures a before-snapshot, applies one operator, performs an extra exploration edge when configured, and finally records whether the genome actually grew. That final feedback is what later allows operator adaptation and bandit-style policies to reward operators that change structure instead of merely consuming attempts.

Parameters:

• genome - Genome to mutate.
• internal - NEAT controller context.
• methods - Mutation methods module.

Returns: Promise resolving after mutation attempts complete.

resolveEffectiveAmount

resolveEffectiveAmount(
  genome: GenomeWithMetadata,
  internal: NeatControllerForMutation,
): number

Resolve the effective mutation amount for a genome.

Mutation amount answers a different question from mutation rate. Rate decides whether the genome enters the flow. Amount decides how many operator draws it receives once admitted. When adaptive mutation amount is enabled, the genome may carry its own evolving attempt budget; otherwise the controller-wide amount stays authoritative.

Parameters:

• genome - Genome to resolve for.
• internal - NEAT controller context.

Returns: Effective mutation amount.

resolveEffectiveRate

resolveEffectiveRate(
  genome: GenomeWithMetadata,
  internal: NeatControllerForMutation,
): number

Resolve the effective mutation rate for a genome.

This helper explains the precedence order for rate policy:

1. an explicit controller-level mutationRate wins,
2. otherwise an adaptive per-genome _mutRate is used when adaptive mutation is enabled,
3. otherwise the flow falls back to the local default.

Keeping that precedence isolated here makes the rest of the mutation flow read as orchestration instead of configuration branching.

Parameters:

• genome - Genome to resolve for.
• internal - NEAT controller context.

Returns: Effective mutation rate.

selectConcreteMutationMethod

selectConcreteMutationMethod(
  genome: GenomeWithMetadata,
  internal: NeatControllerForMutation,
): Promise<MutationMethod | null>

Select a concrete mutation method, resolving legacy arrays when present.

The selection boundary may already return one final operator, or it may return a legacy array-like pool for backward-compatible paths. This helper is the bridge between that policy layer and the execution layer in flow/: it guarantees the rest of the loop sees either one concrete operator or null.

That normalization keeps mutateGenome() focused on lifecycle sequencing rather than on legacy selection-shape quirks.

Parameters:

• genome - Genome to select for.
• internal - NEAT controller context.

Returns: Resolved mutation method or null.

shouldInvalidateCaches

shouldInvalidateCaches(
  mutationMethod: MutationMethod,
  methods: { mutation: unknown; },
): boolean

Determine whether a mutation method invalidates cached structures.

Not every mutation should force expensive cache rebuilds. This helper keeps the invalidation policy explicit by listing the operators that can change the graph structure or traversal semantics enough to make cached topology views unsafe.

Parameters:

• mutationMethod - Mutation operator to inspect.
• methods - Mutation methods module.

Returns: True when caches should be invalidated.

shouldMutateGenome

shouldMutateGenome(
  effectiveRate: number,
  internal: NeatControllerForMutation,
): boolean

Decide whether a genome should be mutated based on probability.

This is the narrow admission gate for the per-genome flow. Keeping the RNG comparison in one helper makes the orchestration read clearly and gives tests one stable seam for deterministic gating behavior.

Parameters:

• effectiveRate - Effective mutation probability.
• internal - NEAT controller context.

Returns: True when the genome should be mutated.

updateOperatorStatsIfNeeded

updateOperatorStatsIfNeeded(
  genome: GenomeWithMetadata,
  mutationMethod: MutationMethod,
  beforeSizes: { beforeNodes: number; beforeConns: number; },
  internal: NeatControllerForMutation,
): void

Update operator statistics when adaptation is enabled.

The mutation subtree measures operator success using a local structural proxy: did the attempted operator increase nodes or connections compared with the pre-mutation snapshot? That signal is imperfect, but it is cheap enough to collect every generation and concrete enough for later adaptation logic to bias toward operators that are actually creating new structure.

Parameters:

• genome - Genome used to compute after-sizes.
• mutationMethod - Operator being recorded.
• beforeSizes - Structural sizes captured before mutation.
• internal - NEAT controller context.

Returns: Nothing.