Öffentlicher Katalog

Voller NFG-Feature-Katalog

Transparenter Blick auf alles, was heute implementiert ist, geordnet nach Domäne und mit dokumentierten Referenzen.

Kontakt für DiligenceZum Katalog

Domänen

64

Sektionen

298

Features

1748

Einträge mit unklaren Referenzen wurden aus Genauigkeitsgründen entfernt.

Feature-Index

Alle implementierten Fähigkeiten

Suche oder springe direkt in eine Domäne, um den vollen Katalog zu sehen.

Zeige 1748 von 1748 Features

Domänen

64

Sektionen

298

Features

1748

Suche

Zur Domäne springen

Neural Architecture61GPU Backend33Neuromodulation System19Attention & Memory25STDP & Neurodynamics30World Prediction15Audio Communication33Meta-Learning18Multi-Timescale Delay Prediction8Fitness & Evolution29Configuration & Serialization35Weather Simulation24Input/Output Processing26Telemetry & Debugging43Evolution Orchestration22GPU Raycasting34TorqueStorm Vehicle Physics17Track Generator13Quantization System11Hall of Fittest Persistence12GPU Vehicle Simulation17GPU RNN - LSTM/GRU/Elman18GPU Physics Engine14GPU Analytics & Statistics17AgentData System17Camera Systems20Mutation System20Visualization System9NFGManager Subsystems28Backend & Provider Interfaces16Profile & Version System20Sleep Replay System23Kernel Components19Statistics & Monitoring19Weather Subsystems Extended27WorldPredictor GPU Shaders17Model Component Helpers24Orchestration Components17Debug & Vehicle Systems21GPU Shader Architecture46Unit Test Suite22Configuration System22Provider & Interface Patterns17Blueprint Integration16Console Variables & Commands19Logging & Telemetry Infrastructure23Serialization & Persistence23Build System & Module Dependencies27FeedForward Shader Implementation27CMA-ES GPU Kernels36Crossover & Mutation Shaders2Quantization Codec Details11Widget & UI System5Remote IO & Authentication8Input Processing Pipeline10CPU Crossover Implementation47RDG Shader Dispatch Pipeline45NFGLM Tokenizer & Decoder3AudioComm Synthesis Internals37Weather State Machine Internals41GPU Structured Buffer Layouts9Test Coverage Analysis253NFGVision Retina System (87 features)119NFGServer & Distributed Learning9

Neural Architecture

61 Features

1.1 MMS Neurons (Multi-Mode State) - UNIQUE INNOVATION

  • MMSMode

    N - Current active mode (0-7)

    CPUComponent_FeedForward.cpp:1133

  • MMSStateInfluence

    N - State → mode selection weight

    CPUComponent_FeedForward.cpp:1150

  • MMSInputInfluence

    N - Input → mode selection weight

    CPUComponent_FeedForward.cpp:1150

  • MMSDecay

    N - Per-neuron state decay rate

    CPUComponent_FeedForward.cpp:1234

  • MMSEta

    N - Learning rate multiplier

    CPUComponent_FeedForward.cpp:1236

1.3 Recurrent Architectures

  • LSTM (4-gate)

    Input, Forget, Output, Cell gates with peephole connections

    CPUComponent_FeedForward.cpp:300-517

  • GRU (2-gate)

    Update, Reset gates (~25% faster than LSTM)

    CPUComponent_FeedForward.cpp:529-790

  • Gated Elman

    Classic RNN with learnable gating factors

    CPUComponent.cpp:4789-5517

  • Classic Elman

    Simple recurrent with pre-activation buffer

    CPUComponent.cpp:5518-5810

  • Per-Gate RMGA

    Neuromodulator integration for each gate independently

    CPUComponent_FeedForward.cpp:413-432

  • Gate-Specific Gains

    Separate DA/NE/ACh/5-HT coefficients per gate

    CPUComponent_FeedForward.cpp:420-428

  • Peephole Connections

    Cell state directly influences gates

    CPUComponent_FeedForward.cpp:380-395

1.4 Activation Functions (15+ variants)

  • TS-ReLU DZ-Asym

    Per-neuron asymmetric with deadzones, hysteresis, caps

    CPUComponent.cpp:4504-4630

  • PReLU

    Per-neuron learnable negative slope (alpha evolved)

    CPUComponent.cpp:4660-4688

  • Swish

    x × sigmoid(x)

    CPUComponent.cpp:4697-4702

  • GELU

    Gaussian Error Linear Unit

    CPUComponent.cpp:4703-4710

  • Mish

    x × tanh(softplus(x))

    CPUComponent.cpp:4711-4718

  • ELU

    Exponential Linear Unit

    CPUComponent.cpp:4683-4690

  • SELU

    Scaled ELU with self-normalization

    CPUComponent.cpp:4691-4695

  • Softplus

    log(1 + exp(x))

    CPUComponent.cpp:4720

  • Softsign

    x / (1 + - x - )

    CPUComponent.cpp:4722

  • Hard Sigmoid

    Piecewise linear sigmoid approximation

    CPUComponent.cpp:4725

  • Hard Swish

    x × hard_sigmoid(x)

    CPUComponent.cpp:4728

1.5 Mixed Precision & Quantization System

  • 1-bit Binary

    Sign-only weights with threshold logic

    CPUComponent_FeedForward.cpp:1079-1095

  • 2-bit Ternary

    {-1, 0, +1} with evolved thresholds

    CPUComponent_FeedForward.cpp:1097-1110

  • 4-bit Quantized

    16 levels with per-neuron scale

    CPUComponent_FeedForward.cpp:126-140

  • 8-bit Quantized

    256 levels for high-precision

    CPUComponent_FeedForward.cpp:141-155

  • Per-Neuron Scales

    Evolved quantization parameters

    CPUComponent_FeedForward.cpp:942-944

  • STDP Scale Evolution

    Scales adapt via spike-timing rules

    NFGCSQuantGuardTests.cpp:48-92

  • Bitplane Encoding

    Efficient binary input representation

    CPUComponent_FeedForward.cpp:128

  • GPU Bit-Level Decode

    Shader unpacks quantized weights

    FeedForwardBatch2D.usf:245-280

  • Dynamic Precision

    Runtime bit-width selection per layer

    CPUComponent_FeedForward.cpp:156-180

1.6 Crossover Operators (11 methods)

  • Block Crossover

    Fixed-size blocks from each parent

    CPUComponent_Crossover.cpp:60-92

  • Uniform Per-Neuron

    50/50 per neuron random selection

    CPUComponent_Crossover.cpp:94-100

  • Multi-Point

    N random cut points (N=2-5)

    CPUComponent_Crossover.cpp:101-130

  • Weight Averaging

    Arithmetic mean of parent weights

    CPUComponent_Crossover.cpp:152-168

  • Blended

    α×P1 + (1-α)×P2 with random α

    CPUComponent_Crossover.cpp:170-195

  • LSTM-Specific

    Separate handling per gate

    CPUComponent.cpp:1800-1950

  • PReLU Alpha

    Per-neuron slope inheritance

    CPUComponent.cpp:1120-1225

  • Coherent Neuron

    All weights for a neuron from same parent

    NFGCrossoverUtils.cpp:89-145

  • Layer-Wise

    Entire layers from alternating parents

    CPUComponent_Crossover.cpp:200-230

  • Fitness-Weighted

    Interpolation biased by fitness ratio

    CPUComponent_Crossover.cpp:232-260

  • Elite Thawing

    Frozen elite weights gradually unfrozen

    CPUComponent_Crossover.cpp:262-290

1.7 Mutation Strategies (10+ methods)

  • Strategy

    Description - Source

  • Gaussian Additive

    value += N(0,σ) × rate

    CPUComponent.cpp:6378-6388

  • Adaptive Per-Layer Sigma

    Learnable step-size per layer

    CPUComponent.cpp:6389-6422

  • Success-Based Adaptation

    1/5th rule sigma adjustment

    MutationComponent.cpp:145-180

  • Sinusoidal Control

    Periodic exploration cycles

    MutationComponent.cpp:182-210

  • Cauchy Mutation

    Heavy-tailed for large jumps

    CPUComponent.cpp:6424-6445

  • Polynomial Mutation

    Bounded with controllable distribution

    CPUComponent.cpp:6447-6470

  • Uniform Perturbation

    ±δ uniform random

    CPUComponent.cpp:6472-6485

  • Sparse Mutation

    Only K% of weights mutated

    CPUComponent.cpp:6487-6510

  • Layer-Specific Rates

    Different σ per layer depth

    CPUComponent.cpp:6512-6540

  • Momentum Mutation

    Velocity-based directed mutation

    CPUComponent.cpp:6542-6580

1.8 Population Management

  • Hall of Fittest

    Elite archive (64-256 individuals)

    CPUComponent.cpp:5955-6290

  • Async Child Pool

    Background thread child generation

    CPUComponent.cpp:404-518, 2202-3370

  • CMA-ES Integration

    Covariance matrix adaptation

    CPUComponent.cpp:5924-6000

  • Fitness-Ranked Selection

    Tournament from elite archive

    CPUComponent.cpp:6501-6533

  • Age-Based Culling

    Remove stale individuals

    HallOfFittestComponent.cpp:280-320

  • Diversity Maintenance

    Novelty bonus in selection

    CPUComponent.cpp:6535-6580

  • Speciation

    Cluster-based niche protection

    CPUComponent.cpp:6582-6640

GPU Backend

33 Features

2.1 Neural Network Compute Shaders

  • Shared Memory Tiling

    16×16 tile matrix multiply

    FeedForwardBatch2D.usf:180-240

  • Wave Intrinsics

    WaveActiveSum for reductions

    FeedForwardBatch2D.usf:300-340

  • Half-Precision Path

    FP16 accumulation option

    FeedForwardBatch2D.usf:85-120

  • Quantized Weight Decode

    On-the-fly 1/2/4/8-bit unpack

    FeedForwardBatch2D.usf:245-280

2.2 GPU CMA-ES (26 kernels) - Complete Implementation

  • Kernel

    Description - Source

  • CMAESInitCS

    Initialize mean, covariance, paths

    CMAES.usf:39-88

  • CMAESGenerateSamplesCS

    Box-Muller sampling

    CMAES.usf:134-251

  • CMAESEvaluateFitnessCS

    Parallel fitness computation

    CMAES.usf:253-320

  • CMAESSortCS

    Bitonic sort by fitness

    CMAES.usf:322-398

  • CMAESCovUpdate2DCS

    Full covariance update

    CMAES.usf:401-442

  • CMAESUpdateMeanCS

    Weighted recombination

    CMAES.usf:444-490

  • CMAESUpdatePathsCS

    Evolution path update

    CMAES.usf:492-542

  • CMAESEigenDecompositionCS

    B×D×B^T decomposition

    CMAES.usf:545-606

  • CMAESBuildInvSqrtCCS

    C^(-1/2) computation

    CMAES.usf:634-661

  • CMAESLowRankUpdateCS

    Rank-μ approximation

    CMAES.usf:663-720

  • CMAESDeflationCS

    Multi-modal niching

    CMAES.usf:722-780

  • CMAESQueueNudgeCS

    Diversity maintenance

    CMAES.usf:782-840

  • CMAESBoundConstraintCS

    Box constraint handling

    CMAES.usf:842-890

  • CMAESStepSizeCS

    σ adaptation (CSA)

    CMAES.usf:892-940

  • CMAESRestartCS

    IPOP/BIPOP restart

    CMAES.usf:942-960

2.3 GPU STDP & Plasticity

  • UpdateSTDPCS

    Spike-timing trace update

    STDP.usf:18-42

  • ApplySTDPCS

    Weight/bias updates

    STDP.usf:75-150+

  • PreSynapticTrace

    Exponential decay + spike

    STDP.usf:25-32

  • PostSynapticTrace

    Activity-dependent trace

    STDP.usf:33-40

  • NeuromodulatorGating

    DA/NE/ACh/5-HT scaling

    STDP.usf:85-110

  • EligibilityTrace

    Temporal credit assignment

    STDP.usf:112-135

  • WeightBounding

    Min/max constraint

    STDP.usf:140-148

2.4 GPU Vehicle Physics (10 subsystems)

  • Subsystem

    Description - Source

2.6 GPU Memory Management

  • Triple Buffering

    Async GPU/CPU overlap

    ComputeShadersComponent.cpp:1200-1280

  • Ring Buffer Histories

    Temporal data storage

    ComputeShadersComponent.cpp:1282-1350

  • Structured Buffers

    Type-safe GPU data

    ComputeShadersComponent.cpp:1352-1420

  • RDG Integration

    Render Dependency Graph

    ComputeShadersComponent.cpp:1422-1500

  • Async Readback

    Non-blocking GPU→CPU

    ComputeShadersComponent.cpp:1502-1580

Neuromodulation System

19 Features

3.1 Four-Channel Neurotransmitter Simulation

  • Signal

    Computation - Effect - Source

  • Dopamine (DA)

    Reward - RewardEWMA - Reinforces positive outcomes

    NeuromodulatorComputer.cpp:102-107

  • Serotonin (5-HT)

    0.5×Volatility + 0.3×NegativeRatio + 0.2×Drawdown - Dampens during risk

    NeuromodulatorComputer.cpp:109-138

  • Noradrenaline (NE)

    max(RewardSurprise, HiddenSurprise_L2) - Enhances on surprise

    NeuromodulatorComputer.cpp:140-157

  • Acetylcholine (ACh)

    ContextShift + 0.5×InputActivity - Focus on context change

    NeuromodulatorComputer.cpp:159-192

3.2 Signal Processing Pipeline

  • Z-Score Normalization

    Welford's online algorithm, ±3σ clipping

    NeuromodulatorComputer.cpp:194-203

  • DA-5HT Decorrelation

    Linear regression removes correlation

    NeuromodulatorComputer.cpp:217-225

  • Spiking Dynamics

    Spike threshold, burst amplitude, refractory

    NeuromodulatorComputer.cpp:227-244

  • Hysteresis Gating

    Prevents oscillation, ε=0.02

    NeuromodulatorComputer.cpp:246-253

  • Temporal Smoothing

    EMA with configurable tau

    NeuromodulatorComputer.cpp:255-270

3.3 Gate Modulation Formulas

  • Feature

    Formula - Source

  • Multi-Modulator RMGA

    Gain = 1.0 + g_DA×DA + g_NE×NE + g_ACh×ACh + g_5HT×5HT

    NeuromodulatorComputer.cpp:90-94

  • Per-Gate Coefficients

    Each LSTM/GRU gate has separate g_* values

    CPUComponent_FeedForward.cpp:420-428

  • Eligibility-Weighted

    Temporal credit via eligibility traces

    MultiModulatorRMGA.ush:126-179

  • Hybrid Spiking

    Membrane potential + spike generation

    rmga_modulation.ush:104-156

3.4 GPU Neuromodulation

  • GPU Multi-Modulator

    Shader implementation for batch

    MultiModulatorRMGA.ush:31-58

  • Per-Agent Modulators

    Independent state per agent

    MultiModulatorRMGA.ush:60-85

  • Async Update

    Non-blocking modulator compute

    MultiModulatorRMGA.ush:87-102

  • Unified Interface

    Legacy + new modulator support

    UnifiedModulatorComputer.cpp:1-200

Attention & Memory

25 Features

4.1 Attention Mechanisms (4 Modes)

  • Mode

    Description - Formula - Source

  • Disabled

    No attention - Pass-through

    NFGAttentionProcessor.cpp:85

  • Layer

    Per-neuron gate from magnitudes - Gate = \ - Act\ - / Sum

    NFGAttentionProcessor.cpp:87-112

  • Temporal

    Multi-head Q/K/V multiplicative - Softmax(Q·K/√d) × V

    NFGAttentionProcessor.cpp:115-336

  • TemporalAdditive

    Residual attention (RECOMMENDED) - 1.0 + (Gate-0.5)×2×Scale

    NFGAttentionProcessor.cpp:339-354

4.2 GPU Fused Attention

  • Half-Precision

    FP16 accumulation for speed

    AttentionFusedCS.usf:84-120

  • Wave Intrinsics

    WaveActiveSum for softmax

    AttentionFusedCS.usf:180-220

  • Triple Buffering

    Async Q/K/V updates

    AttentionFusedCS.usf:250-300

  • Scaled Dot-Product

    Efficient attention kernel

    AttentionFusedCS.usf:302-350

  • Multi-Head Support

    Up to 8 attention heads

    NFGAttentionGPU.cpp:120-180

4.3 Multi-Timescale Memory (8 timescales)

  • Timescale

    Alpha (EMA) - Effective Range - Use Case

  • Instant

    0.0 - Current frame - Immediate response

  • Fast

    0.5 - ~16ms - Reflexes

  • 1 Second

    0.9 - ~1s - Short-term

  • 5 Seconds

    0.95 - ~5s - Working memory

  • 10 Seconds

    0.99 - ~10s - Recent context

  • 30 Seconds

    0.995 - ~30s - Episode context

  • 1 Minute

    0.999 - ~60s - Extended memory

  • 3 Minutes

    0.9999 - ~180s - Long-term patterns

4.4 Composite Memory Neurons (12-14 injected)

  • Category

    Neurons - Purpose - Source

  • Activation Memory

    Level, Trend, Stability - Activity patterns

    TemporalMemoryCS.usf:174-200

  • Network Uncertainty

    Confidence, Saturation, Exploration - Learning state

    TemporalMemoryCS.usf:201-220

  • Reward Dynamics

    Cascade, Advantage, Momentum - Reward trends

    TemporalMemoryCS.usf:221-240

  • Episode Context

    Phase, Performance, Crisis - Episode state

    TemporalMemoryCS.usf:241-260

  • WorldPredictor

    ErrorBlend, SurpriseTrend - Prediction quality

    TemporalMemoryCS.usf:252-264

STDP & Neurodynamics

30 Features

5.1 Core STDP Formulas

  • Feature

    Formula - Source

  • Pre-Synaptic Trace

    τ_pre × dTrace/dt = -Trace + δ(spike)

    NFGNeuroDynamicsProvider.cpp:567-569

  • Post-Synaptic Trace

    τ_post × dTrace/dt = -Trace + δ(spike)

    NFGNeuroDynamicsProvider.cpp:598-599

  • LTP (Potentiation)

    Δw = η_LTP × Pre_trace × δ(post_spike)

    NFGNeuroDynamicsProvider.cpp:606

  • LTD (Depression)

    Δw = -η_LTD × Post_trace × δ(pre_spike)

    NFGNeuroDynamicsProvider.cpp:606

  • Weight Update

    w_new = w_old + DA × (LTP - LTD)

    NFGNeuroDynamicsProvider.cpp:610-620

5.2 Reward-Modulated STDP

  • Factor

    Computation - Source

  • RewardPotFactor

    1.0 + Scale × max(0, Reward)

    NFGNeuroDynamicsProvider.cpp:510

  • RewardDepFactor

    1.0 + Scale × max(0, -Reward)

    NFGNeuroDynamicsProvider.cpp:511

  • DopamineFactor

    1.0 + Scale × Dopamine

    NFGNeuroDynamicsProvider.cpp:533

  • SerotoninFactor

    1.0 - Scale × Serotonin

    NFGNeuroDynamicsProvider.cpp:534

  • EffectiveLTP

    LTPRate × RewardPotFactor × DopamineFactor

    NFGNeuroDynamicsProvider.cpp:540

  • EffectiveLTD

    LTDRate × RewardDepFactor × SerotoninFactor

    NFGNeuroDynamicsProvider.cpp:541

5.3 Predictive Coding

  • Prediction Error

    Observation - Prediction from top-down

    NFGNeuroDynamicsProvider.cpp:936-1018

  • Running Error Mean

    EMA of - error - for baseline

    NFGNeuroDynamicsProvider.cpp:620-624

  • Surprise Detection

    Boost when - error - > threshold × mean

    NFGNeuroDynamicsProvider.cpp:626-629

  • Top-Down Weights

    Auto-transposed or dedicated

    NFGNeuroDynamicsProvider.cpp:840-889

  • Error-Driven Learning

    Prediction error modulates STDP

    NFGNeuroDynamicsProvider.cpp:630-660

  • Hierarchical Prediction

    Each layer predicts layer below

    NFGNeuroDynamicsProvider.cpp:890-935

5.4 Sleep Replay (Memory Consolidation)

  • Experience Buffer

    Circular buffer (activations, rewards, frame IDs)

    NFGNeuroDynamicsProvider.h:42-59

  • Priority Sampling

    High- - reward - frames prioritized

    NFGNeuroDynamicsProvider.cpp:1209-1244

  • Sync Replay

    Blocking replay with full STDP pass

    NFGNeuroDynamicsProvider.cpp:1283-1350

  • Async Replay

    Incremental replay over frames

    NFGNeuroDynamicsProvider.cpp:1352-1423

  • Replay Pass Signal

    Global flag notifies providers

    NFGNeuroDynamicsProvider.cpp:121-126

  • Replay Batch Size

    Configurable experiences per cycle

    NFGNeuroDynamicsProvider.cpp:1246-1280

  • Recency Weighting

    Recent experiences prioritized

    NFGNeuroDynamicsProvider.cpp:1200-1208

5.5 Spike Generation

  • Threshold Crossing

    Spike when activation > threshold

    NFGNeuroDynamicsProvider.cpp:480-495

  • Refractory Period

    Minimum inter-spike interval

    NFGNeuroDynamicsProvider.cpp:497-510

  • Adaptive Threshold

    Threshold increases after spike

    NFGNeuroDynamicsProvider.cpp:512-525

  • Stochastic Spiking

    Probabilistic based on activation

    NFGNeuroDynamicsProvider.cpp:527-545

World Prediction

15 Features

6.2 What-If Rollouts (Counterfactual Reasoning)

  • Hidden State Rollout

    H' = tanh(W_h × H + W_a × A + b)

    WorldPredictorWhatIf.usf:1-262

  • Confidence Scoring

    1/(1+RMSE) for stability

    WorldPredictorWhatIf.usf:243-247

  • Danger Scoring

    Designated danger neurons + threshold

    WorldPredictorWhatIf.usf:188-238

  • Multi-Step Rollout

    Ping-pong buffers for iteration

    WorldPredictorWhatIf.usf:117-186

  • Best Action Selection

    Score = Confidence - DangerScore

    NFGWorldPredictorComponent.h:107-108

  • Parallel Action Queries

    Multiple actions evaluated simultaneously

    WorldPredictorWhatIf.usf:45-80

  • State Caching

    Reuse common prefixes

    WorldPredictorWhatIf.usf:82-115

6.3 Action Variants System

  • Gradient Estimation

    Finite difference gradients

    WorldPredictorActionVariants.usf:80-120

6.4 Action Effects & Correction

  • Action-Effect Error

    Delta prediction: - actual - predicted

    WorldPredictorActionEffectError.usf:1-83

  • What-If Blending

    Softmax-weighted correction

    WorldPredictorActionCorrection.usf:1-100+

  • Action Context

    N-frame action history as input

    WorldPredictorActionContext.usf:1-108

6.5 Multi-Timescale Prediction

  • Horizon Weighting

    Exponential decay by horizon

    WorldPredictorMultiTimescale.usf:45-80

6.6 Token-Level Prediction (for NFGLM)

  • Cross-Entropy Loss

    Token prediction loss

    WorldPredictorTokenCE.usf:1-120

  • Sequence Masking

    Causal masking for AR

    WorldPredictorTokenCE.usf:40-70

  • Position Encoding

    Learned positional embeddings

    WorldPredictorTokenCE.usf:72-100

Audio Communication

33 Features

8.1 Speech Synthesis (20 Control Parameters)

  • F0 (Pitch)

    80-400 Hz - Fundamental frequency

    NFGAudioCommSynthComponent.cpp:486-495

  • Voicing

    0.0-1.0 - Voiced/unvoiced ratio

    NFGAudioCommSynthComponent.cpp:497-505

  • F1, F2, F3

    200-3500 Hz - Formant frequencies

    NFGAudioCommSynthComponent.cpp:416-440

  • B1, B2, B3

    50-400 Hz - Formant bandwidths

    NFGAudioCommSynthComponent.cpp:442-458

  • Aspiration

    0.0-1.0 - Breathiness

    NFGAudioCommSynthComponent.cpp:545-550

  • Tremolo Rate

    0-10 Hz - Vibrato frequency

    NFGAudioCommSynthComponent.cpp:507-515

  • Tremolo Depth

    0.0-0.3 - Vibrato amplitude

    NFGAudioCommSynthComponent.cpp:517-525

  • Jitter

    0.0-0.05 - Pitch irregularity

    NFGAudioCommSynthComponent.cpp:527-535

  • Shimmer

    0.0-0.05 - Amplitude irregularity

    NFGAudioCommSynthComponent.cpp:537-545

  • Nasality

    0.0-1.0 - Nasal resonance

    NFGAudioCommSynthComponent.cpp:547-555

8.2 Glottal Source Model

  • LF Model

    Liljencrants-Fant glottal pulse

    NFGAudioCommSynthComponent.cpp:486-500

  • Open Quotient

    Glottal open phase ratio

    NFGAudioCommSynthComponent.cpp:502-510

  • Speed Quotient

    Opening/closing asymmetry

    NFGAudioCommSynthComponent.cpp:512-520

  • Return Phase

    Pulse return coefficient

    NFGAudioCommSynthComponent.cpp:522-530

8.3 Formant Filter Bank

  • 3-Band IIR

    F1, F2, F3 resonators

    NFGAudioCommSynthComponent.cpp:416-458

  • Parallel Config

    Summed formant outputs

    NFGAudioCommSynthComponent.cpp:460-475

  • Cascade Config

    Serial formant chain

    NFGAudioCommSynthComponent.cpp:477-485

  • Anti-Formants

    Nasal zero modeling

    NFGAudioCommSynthComponent.cpp:380-400

8.4 Phoneme System (9 Phonemes)

  • Phoneme

    Formants (F1/F2/F3) - Source

  • /a/

    730/1090/2440

    NFGAudioCommSynthComponent.cpp:357-360

  • /e/

    530/1840/2480

    NFGAudioCommSynthComponent.cpp:361-364

  • /i/

    270/2290/3010

    NFGAudioCommSynthComponent.cpp:365-368

  • /o/

    570/840/2410

    NFGAudioCommSynthComponent.cpp:369-372

  • /u/

    300/870/2240

    NFGAudioCommSynthComponent.cpp:373-376

  • /ə/ (schwa)

    500/1500/2500

    NFGAudioCommSynthComponent.cpp:377-379

  • /æ/

    660/1720/2410

    NFGAudioCommSynthComponent.cpp:380-382

  • /ʌ/

    600/1190/2390

    NFGAudioCommSynthComponent.cpp:383-385

  • Interpolated

    Lerped F1/F2/F3

    NFGAudioCommSynthComponent.cpp:387-400

8.5 Agent-to-Agent Communication

  • Global Mean

    Distance-weighted parameter averaging

    NFGAudioCommWorldSubsystem.cpp:361-407

  • Nearest-K

    K nearest neighbors aggregation

    NFGAudioCommWorldSubsystem.cpp:441-502

  • Spatial Attenuation

    Exponential distance falloff

    NFGAudioCommSpeakerComponent.cpp:160-174

  • Voice Profiles

    Per-agent 7-bias customization

    NFGAudioCommSpeakerComponent.cpp:37-110

  • Emergent Protocol

    Communication evolves through fitness

    NFGAudioCommWorldSubsystem.cpp:504-560

Meta-Learning

18 Features

10.1 UCB Selection System

  • Feature

    Formula - Source

  • UCB Score

    Exploit + C × √(ln(N)/n)

    NFGMetaRegulatorComponent.cpp:72-113

  • Exploit Term

    Mean contribution estimate

    NFGMetaRegulatorComponent.cpp:80-90

  • Explore Term

    Uncertainty bonus

    NFGMetaRegulatorComponent.cpp:92-110

  • C Parameter

    Exploration coefficient (√2)

    NFGMetaRegulatorComponent.cpp:75

10.2 Output Family Grouping

  • Semantic Clustering

    Group outputs by name prefix

    NFGMetaRegulatorComponent.cpp:406-567

  • Wildcard Matching

    "Steering_*" matches all steering

    NFGMetaRegulatorComponent.cpp:420-450

  • Auto-Discovery

    Detect families from output names

    NFGMetaRegulatorComponent.cpp:452-500

  • Family Budget

    Per-family resource allocation

    NFGMetaRegulatorComponent.cpp:502-567

10.3 Intervention System

  • Gain Intervention

    Temporary gain modification

    NFGMetaRegulatorComponent.cpp:815-900

  • Impact Measurement

    Δfitness during intervention

    NFGMetaRegulatorComponent.cpp:902-950

  • Contribution Tracking

    Normalized delta measurement

    NFGMetaRegulatorComponent.cpp:952-984

  • Rollback Support

    Revert unsuccessful interventions

    NFGMetaRegulatorComponent.cpp:986-1020

10.4 Adaptive Control

  • Softmax Budget

    Temperature-scaled distribution

    NFGMetaRegulatorComponent.cpp:678-759

  • Mutation Scale

    Per-family mutation adaptation

    NFGMetaRegulatorComponent.cpp:761-813

  • Input Gain

    Per-family input sensitivity

    NFGMetaRegulatorComponent.cpp:1056-1141

  • Curiosity Signal

    Exploration boost for surprising

    NFGMetaRegulatorComponent.cpp:954, 789-791

  • Progress Tracking

    Per-family improvement history

    NFGMetaRegulatorComponent.cpp:1143-1200

Multi-Timescale Delay Prediction

8 Features

11.1 Temporal Layer Projection

  • Layer Selection

    1-2 hidden layers for projection

    NFGMTDPComponent.cpp:728-777

  • Early Timescales

    Layer 0 = fast (5-10%)

    NFGMTDPComponent.cpp:214-220

  • Mid Timescales

    Layer 1 = moderate (10-25%)

    NFGMTDPComponent.cpp:221-223

  • GPU Contribution

    Shader-based dot product

    MTDP.ush:8-19

11.2 Genetic Operations

  • 9 Crossover Methods

    Block, Uniform, PerLayer, Average, etc.

    NFGMTDPComponent.cpp:445-578

  • Output Coherence

    All weights per output from same parent

    NFGMTDPComponent.cpp:486-496

  • Weight Clamping

    Bounded evolution (0.0001 to 5.0)

    NFGMTDPComponent.cpp:854-867

  • Adaptive Mixing

    Fitness-proportional parent blend

    NFGMTDPComponent.cpp:580-620

Fitness & Evolution

29 Features

12.1 Internal Meta-Fitness Metrics (18 metrics)

  • Metric

    Description - Source

  • NetworkActivityBalance

    Activation distribution evenness

    NFGFitnessComponent.cpp:180-200

  • PredictionErrorLoss

    World prediction MSE

    NFGFitnessComponent.cpp:202-220

  • TemporalCoherence

    Output stability over time

    NFGFitnessComponent.cpp:222-240

  • ExplorationBonus

    Novel state visitation reward

    NFGFitnessComponent.cpp:242-260

  • CuriosityReward

    Prediction error as intrinsic reward

    NFGFitnessComponent.cpp:262-280

  • WeightMagnitudePenalty

    L2 regularization

    NFGFitnessComponent.cpp:282-300

  • SpikeEfficiency

    Information per spike

    NFGFitnessComponent.cpp:302-320

  • MemoryUtilization

    Temporal retention usage

    NFGFitnessComponent.cpp:322-340

  • AttentionEfficiency

    Attention gate effectiveness

    NFGFitnessComponent.cpp:342-360

  • ModulatorBalance

    DA/NE/ACh/5-HT equilibrium

    NFGFitnessComponent.cpp:362-380

  • GradientHealth

    Weight change stability

    NFGFitnessComponent.cpp:382-400

  • SpecializationIndex

    Neuron role differentiation

    NFGFitnessComponent.cpp:402-420

  • InformationFlow

    Layer-to-layer transfer

    NFGFitnessComponent.cpp:422-440

  • AdaptiveComplexity

    Appropriate model complexity

    NFGFitnessComponent.cpp:442-460

  • RobustnessScore

    Performance under noise

    NFGFitnessComponent.cpp:462-480

  • GeneralizationProxy

    Train/test divergence

    NFGFitnessComponent.cpp:482-500

  • ConvergenceRate

    Learning speed

    NFGFitnessComponent.cpp:502-520

  • StabilityScore

    Fitness variance

    NFGFitnessComponent.cpp:522-540

12.2 Hall of Fittest Management

  • Elite Archive

    64-256 top individuals

    HallOfFittestComponent.cpp:50-100

  • Fitness Ranking

    Sorted by composite fitness

    HallOfFittestComponent.cpp:102-150

  • Age Tracking

    Generation stamps

    HallOfFittestComponent.cpp:152-180

  • Diversity Bonus

    Novelty-based selection

    HallOfFittestComponent.cpp:182-220

  • Periodic Culling

    Remove stale individuals

    HallOfFittestComponent.cpp:222-260

  • Serialization

    Checkpoint save/load

    HallOfFittestComponent.cpp:262-320

12.3 Multi-Objective Aggregation

  • Weighted Sum

    Σ w_i × f_i

    NFGFitnessComponent.cpp:550-580

  • Pareto Ranking

    Non-dominated sorting

    NFGFitnessComponent.cpp:582-620

  • Hypervolume

    Coverage metric

    NFGFitnessComponent.cpp:622-660

  • Scalarization

    Configurable aggregation

    NFGFitnessComponent.cpp:662-700

Configuration & Serialization

35 Features

13.1 JSON Configuration System (17 Config Files)

  • Config File

    Purpose - Source

  • ArchitectureConfig.json

    Network topology, MMS neurons, per-neuron modifiers

    Config/ArchitectureConfig.json

  • EvolutionConfig.json

    Mutation rates, crossover methods, parent selection

    Config/EvolutionConfig.json

  • FitnessConfig.json

    Fitness evaluation, reward weights, meta-regulation

    Config/FitnessConfig.json

  • AttentionConfig.json

    Temporal attention modes, heads, temporal windows

    Config/AttentionConfig.json

  • BackendConfig.json

    GPU/CPU selection, precision, quantization

    Config/BackendConfig.json

  • NeuroDynamicsConfig.json

    STDP, predictive coding, sleep consolidation

    Config/NeuroDynamicsConfig.json

  • RMGAConfig.json

    Reward-modulated gain, neuromodulators

    Config/RMGAConfig.json

  • MTDPConfig.json

    Multi-timescale decision projection

    Config/MTDPConfig.json

  • CMAESConfig.json

    CMA-ES evolution strategy parameters (172+)

    Config/CMAESConfig.json

  • RNNConfig.json

    RNN/LSTM/GRU configuration

    Config/RNNConfig.json

  • OutputManagerConfig.json

    Output layer configuration

    Config/OutputManagerConfig.json

  • SensorsConfig.json

    Sensor/input configuration

    Config/SensorsConfig.json

  • WorldPredictorConfig.json

    World model prediction settings

    Config/WorldPredictorConfig.json

  • GameConfig.json

    Game-specific parameters

    Config/GameConfig.json

13.2 Model Serialization

  • SerializeJSON

    Complete model to JSON with weights/biases

    NFGModelComponent_Serialization.cpp:17-200

  • Quantization Support

    2-64 bit precision with emulated quantize-dequantize

    NFGModelComponent_Serialization.cpp:42-157

  • Implicit Bias

    Biases as +1 augmented weight column

    NFGModelComponent_Serialization.cpp:160-162

  • Per-Neuron Scales

    Individual scale factor per output (QuantPerNeuronScales)

    NFGModelComponent_Serialization.cpp:100-105

  • Base64 Weights

    QWeightsPacked_b64 for efficient storage

    NFGModelComponent_Serialization.cpp:120-140

  • NeuroDynamics Params

    STDP, predictive coding, sleep settings persistence

    NFGModelComponent_Serialization.cpp:164-186

13.3 Checkpoint & Persistence

  • Hall of Fittest Persistence

    Elite model archive with sorted ranking

    HallOfFittestComponent.cpp:165-206

  • Architecture Hash

    MD5 hash for model versioning/compatibility

    IOManager.cpp:70-88

  • Pre-allocated Model Pool

    Memory reserved before async load

    FileIOManager.cpp:120-131

  • Async File I/O

    Worker thread loads, game thread creates

    FileIOManager.cpp:106-200+

  • Remote I/O

    Cloud/server model upload/download

    IOManager.cpp:30-60

13.4 Runtime CVars

  • CVar

    Type - Purpose - Source

  • nfg.crossover.cpu_trace

    int32 - Trace CPU child production

    CPUComponentCVars.h:10

  • nfg.cpu.quant_dense

    int32 - CPU quantized dense fast path

    CPUComponentCVars.h:18

  • nfg.cpu.quant_lstm

    int32 - CPU quantized LSTM fast path

    CPUComponentCVars.h:20

  • nfg.attention.enable

    int32 - -1=inherit, 0=off, 1=on

    NFGAttentionCVars.cpp:7

  • nfg.attention.mode

    int32 - 0-3 (Disabled/Layer/Temporal/TemporalAdditive)

    NFGAttentionCVars.cpp:12

  • nfg.attention.heads

    int32 - Override head count

    NFGAttentionCVars.cpp:18

  • nfg.attention.use_gpu

    int32 - -1=inherit, 0=CPU, 1=GPU

    NFGAttentionCVars.cpp:35

  • nfg.hof.use_reward_for_ranking

    int32 - Fitness vs reward ranking

    HallOfFittestComponent.cpp:180

Weather Simulation

24 Features

14.1 Weather State Machine

  • 8 Weather States

    Clear, Cloudy, Overcast, LightRain, HeavyRain, Thunderstorm, Snow, Fog

    NFGWeatherStateMachine.cpp:40-80

  • Smooth Transitions

    Interpolation over configurable duration (default 300s)

    NFGWeatherStateMachine.cpp:120-180

  • Automatic Rules

    Condition-based state changes with probability

    NFGWeatherStateMachine.cpp:200-280

  • Weather Front Integration

    Dynamic front-based state modifications

    NFGWeatherStateMachine.cpp:300-350

14.2 Atmospheric Parameters

  • Temperature

    -50°C to +50°C - With lapse rate 0.0065°C/meter

    NFGWeatherSimulationComponent.cpp:150-180

  • Humidity

    0.0-1.0 - Location-dependent

    NFGWeatherSimulationComponent.cpp:200-230

  • Pressure

    95000-105000 Pa - Pressure cell simulation

    NFGWeatherSimulationComponent.cpp:250-300

  • Wind

    0-50 m/s - With turbulence and gusts

    NFGWeatherSimulationComponent.cpp:320-380

  • Cloud Coverage

    0.0-1.0 - Affects visibility

    NFGWeatherStateMachine.cpp:90-110

14.3 Weather Front Physics

  • 4 Front Types

    Cold, Warm, Occluded, Stationary

    NFGWeatherFrontManager.cpp:50-100

  • Frontal Dynamics

    Slope, vertical velocity, vorticity

    NFGWeatherFrontManager.cpp:150-200

  • CAPE Calculation

    Convective Available Potential Energy (0-5000+ J/kg)

    NFGWeatherFrontManager.cpp:220-260

  • Precipitation

    Based on lifting rate and moisture

    NFGWeatherFrontManager.cpp:280-340

  • Front Lifespan

    Decay over time (default 48 hours)

    NFGWeatherFrontManager.cpp:400-450

14.4 Weather Friction Effects

  • Condition

    Friction Multiplier - Source

  • Dry

    1.0

    NFGWeatherFrictionComponent.cpp:80

  • Wet

    0.7

    NFGWeatherFrictionComponent.cpp:82

  • Snow

    0.35

    NFGWeatherFrictionComponent.cpp:84

  • Ice

    0.15

    NFGWeatherFrictionComponent.cpp:86

14.5 Weather Integration

  • UDS Integration

    Ultra Dynamic Sky property control via reflection

    NFGWeatherControlComponent.cpp:100-200

  • 13 Weather Presets

    ClearSkies, Cloudy, Foggy, Rain, Snow, etc.

    NFGWeatherControlComponent.cpp:50-90

  • 24-Hour Forecasting

    Predict weather changes with configurable interval

    NFGIntegratedWeatherSystem.cpp:200-280

  • Severe Weather Warnings

    Alert when thunderstorms approach

    NFGIntegratedWeatherSystem.cpp:320-380

  • Seasonal Temperature

    Spring +0°C, Summer +5°C, Autumn -5°C, Winter -15°C

    NFGWeatherManagerActor.cpp:80-120

Input/Output Processing

26 Features

15.1 Input Data Structures

  • Structure

    Description - Source

  • FInputWithName

    Named input with float value

    NFGModelComponent.h:72-92

  • InputsWithName Array

    Thread-safe array of current inputs

    NFGModelComponent.h:839

  • PreviousInputsWithName

    Historical inputs for recurrent connections

    NFGModelComponent.h:869

  • InputNameToIndexCache

    O(1) lookup map (lowercase name → index)

    NFGModelComponent.h:841-843

15.2 Input Relay Component

  • JSON-Based Mapping

    Maps actor properties to network inputs

    NFGInputRelayComponent.h:23-38

  • 8 Source Types

    LocationX/Y/Z, RotationPitch/Roll/Yaw, TimeSeconds

    NFGInputRelayComponent.h:10-20

  • Pre-Resolved Enum

    O(1) source type lookup

    NFGInputRelayComponent.cpp:62-74

  • BeginProjectInputs/AddProjectInput

    Dynamic input appending API

    NFGModelComponent_Lifecycle.cpp:1862-1916

15.3 Reserved Input Ranges

  • Range

    Purpose - Source

  • NFGVision

    Vision input slots

    NFGInputLayout.h:85-90

  • GazeAbsolute

    Attention gaze input

    NFGInputLayout.h:92-95

  • AudioAggregated

    Aggregated audio input

    NFGInputLayout.h:97-100

  • MetaRange

    Meta-regulator status (10+ inputs)

    NFGInputLayout.h:102-110

  • EvoRange

    Evolution feature inputs

    NFGInputLayout.h:112-115

15.4 Output Processing

  • FOutputWithName

    Named output with float value

    OutputsWithName.h:12-45

  • OutputHistory Ring Buffer

    Configurable depth (default 10 frames)

    OutputHistoryComponent.h:74-85

  • Weighted History

    Linear decay from 1.0 (newest) to ~0.0 (oldest)

    OutputHistoryComponent.cpp:75-85

  • Reward History

    Circular buffer with CurrentRewardIndex

    NFGModelComponent_History.cpp:37-60

  • 27 Activation Functions

    Full output post-processing

    CPUComponent.cpp:4643-4768

15.5 Sensor Configuration

  • ESensorsDirectionMode

    FullCircle (360°) or HalfCircle (180°)

    SensorsConfig.h:9-13

  • EGridQueryMode

    Circle or Rectangle coverage

    SensorsConfig.h:19-24

  • VectorsNum

    Number of ray-cast directions

    SensorsConfig.h:48-49

  • MaxRadius

    Detection range in world units

    SensorsConfig.h:60-61

  • bNormalize

    Enable [0,1] scaling

    SensorsConfig.h:70-71

  • PrecomputedDirectionVectors

    Cached ray directions

    SensorsConfig.h:73-74

Telemetry & Debugging

43 Features

16.1 Debug Manager

  • AddDebugLine

    Draw debug lines with thickness and color

    DebugManager.cpp:66

  • AddDebugSphere

    Render spheres at world locations

    DebugManager.cpp:101

  • AddDebugPlane

    Place planes with normal orientation

    DebugManager.cpp:130

  • HISM Spheres/Planes

    Hierarchical instanced mesh for performance

    DebugManager.h:20-23

  • Single-Frame Visualization

    Auto-clears every frame

    DebugManager.cpp:62-63

16.2 Kernel Performance Monitor

  • BeginKernelExecution

    Start timing for named kernel

    KernelPerformanceMonitor.cpp:72

  • EndKernelExecution

    Stop timing and record data

    KernelPerformanceMonitor.cpp:78

  • GFLOPS Estimation

    Kernel-type-aware compute estimate

    KernelPerformanceMonitor.cpp:225

  • Memory Bandwidth

    Rough estimate: 3 ops × 2 bytes per work item

    KernelPerformanceMonitor.cpp:230

  • Performance History

    Last 100 snapshots for trend analysis

    KernelPerformanceMonitor.h:156

  • ExportToCSV

    Output kernel stats to file

    KernelPerformanceMonitor.cpp:162

  • CheckPerformanceIssues

    Detect 20%+ degradation vs baseline

    KernelPerformanceMonitor.cpp:262

16.3 Visualization Component (6 Modes)

  • Mode

    Description - Source

  • State1: ShowOutputOnly

    Display only output neurons

    VisualizationComponent.h:75

  • State2: ShowInputAndOutput

    Input + output neurons

    VisualizationComponent.h:77

  • State3: ShowAllNeurons

    Full network visualization

    VisualizationComponent.h:79

  • State4: + InputValueText

    Neurons with input labels

    VisualizationComponent.h:81

  • State5: + InputOutputText

    Input and output labels

    VisualizationComponent.h:83

  • State6: + AllText

    Complete annotation

    VisualizationComponent.h:85

16.4 Console Commands (50+)

  • Category

    Commands - Source

  • Logging

    EnableFileLogging, DisableFileLogging, FlushLog

    NFGConsoleCommands.cpp:52-66

  • CMA-ES

    EnableCMAES, DisableCMAES, CMAESStatus, SetCMAESLearningRate

    NFGConsoleCommands.cpp:108-294

  • Widgets

    ToggleDecisionExplainer, ToggleTrainingProgress, ToggleAIMilestones

    NFGConsoleCommands.cpp:330-403

  • Testing

    RunAllTests, RunComponentTests, GenerateTestReport

    NFGConsoleCommands.cpp:430-457

  • Weather

    SetWeather, SetTimeOfDay, SetRain, SetSnow, SetFog (17+ commands)

    NFGConsoleCommands.cpp:495-1049

16.5 Logging System (9 Categories)

  • Category

    Purpose - Source

  • NFGLog

    Main general-purpose logging

    NFGLogging.h:11

  • NFGWidget

    UI widget system

    NFGLogging.h:12

  • NFGNetwork

    Neural network operations

    NFGLogging.h:13

  • NFGGpu

    GPU backend and shaders

    NFGLogging.h:14

  • NFGTraining

    Evolution and fitness

    NFGLogging.h:15

  • NFGIo

    File I/O operations

    NFGLogging.h:16

  • NFGWeather

    Environmental system

    NFGLogging.h:17

  • NFGDebug

    Debug-only (suppressed)

    NFGLogging.h:18

  • NFGVersion

    Version and build info

    NFGLogging.h:19

16.6 Training Telemetry (20+ Metrics)

  • Metric

    Description - Source

  • Generation/Epoch

    Evolution progress counters

    NFGTrainingTelemetry.h:15-20

  • ValidModelCount

    Models with valid fitness

    NFGTrainingTelemetry.h:22

  • AverageFitness

    Mean population fitness

    NFGTrainingTelemetry.h:28

  • FitnessStdDev

    Population diversity metric

    NFGTrainingTelemetry.h:30

  • BestFitness/WorstFitness

    Champion and weakest

    NFGTrainingTelemetry.h:32-34

  • HallCount/HallAverageFitness

    Elite pool statistics

    NFGTrainingTelemetry.h:38-44

  • CPUChildQueue/GPUChildQueue

    Pipeline statistics

    NFGTrainingTelemetry.h:48-52

Evolution Orchestration

22 Features

17.1 Orchestration Components

  • Component

    Responsibility - Source

  • NFGEvolutionComponent

    Genetic operators, CMA-ES scheduling

    NFGEvolutionComponent.cpp

  • NFGPopulationComponent

    Population initialization, flat arrays

    NFGPopulationComponent.cpp

  • NFGTrainingComponent

    Inference dispatch to backends

    NFGTrainingComponent.cpp

  • NFGPersistenceComponent

    Model saving/loading

    NFGPersistenceComponent.cpp

  • HallOfFittestComponent

    Elite ranking and selection

    HallOfFittestComponent.cpp

17.2 Evolution Component

  • Crossover

    Parent selection and child generation

    NFGEvolutionComponent.cpp:29-149

  • CrossoverCompleteHallOfFittest

    Batch crossover for all elites

    NFGEvolutionComponent.cpp:185-195

  • TickCMAES

    CMA-ES update scheduling with frame counter

    NFGEvolutionComponent.cpp:151-183

  • TryInitializeCMAES

    Graceful degradation with fallback sizes

    NFGEvolutionComponent.cpp:197-236

  • GPU-Exclusive Handling

    Defers operations if GPU busy

    NFGEvolutionComponent.cpp:40

17.3 Population Component

  • CreateModels

    Initial population with unique ModelIndex

    NFGPopulationComponent.cpp:19-151

  • PrepareFlatArrays

    Memory layout optimization for GPU/CPU

    NFGPopulationComponent.cpp:153-244+

  • InitializeLSTMWeights

    LSTM cell and gate weight setup

    NFGPopulationComponent.cpp:505+

  • InitializeGRUWeights

    GRU weight initialization

    NFGPopulationComponent.cpp:691+

  • NFGVision Integration

    Reserve input space for vision

    NFGPopulationComponent.cpp:45-60

17.4 Hall of Fittest Component

  • InitList

    Initialize elite population with sorting

    HallOfFittestComponent.cpp:49-68

  • AddToHallOfFittest

    Fitness threshold and sorted insertion

    HallOfFittestComponent.cpp:165-206

  • Insert

    Binary search O(log n) with memory recycling

    HallOfFittestComponent.cpp:208-410+

  • AcquireEntry/RecycleEntry

    Memory pool management

    HallOfFittestComponent.cpp:70-111

  • ShrinkList

    Probabilistic culling of worst performers

    HallOfFittestComponent.cpp:467-491

  • OnFittestUpdated Delegate

    Event notification for UI/monitoring

    HallOfFittestComponent.cpp:462-465

GPU Raycasting

34 Features

18.1 GPU Raycast Architecture

  • DXR Path (Legacy)

    Disabled in UE 5.5+ due to API changes

    RaycastGPUManager.cpp:30-60

  • Fallback Compute

    Ultra-optimized compute shader active

    RaycastGPUManager.cpp:73-87

  • Render Graph Integration

    FRenderGraphBuilder for GPU resource management

    RaycastGPUManager.cpp:653-750

  • Async Triple-Buffered Readback

    Non-blocking GPU→CPU transfer

    RaycastGPUManager.cpp:500-600

18.2 360-Degree Omnidirectional Sensing

  • NumRays

    4-720 rays (typical: 8-360)

    RaycastGPU360Mode.h:15

  • MaxDistance

    Ray range in world units (default 5000)

    RaycastGPU360Mode.h:18

  • NumSectors

    Angular sectors for aggregation (default 8)

    RaycastGPU360Mode.h:21

  • Uniform Angular Distribution

    AngleStep = 360 / NumRays degrees

    RaycastGPU360Mode.cpp:20-40

  • Height Offset

    Multi-level perception (default 100 UU)

    RaycastGPU360Mode.h:24

18.3 Sector Aggregation (3 Metrics per Sector)

  • Metric

    Formula - Purpose - Source

  • MinDistance

    min(sector rays) - Closest obstacle (safety)

    RaycastGPU360Mode.cpp:100

  • MaxDistance

    max(sector rays) - Farthest clear space (navigation)

    RaycastGPU360Mode.cpp:102

  • AvgDistance

    mean(sector rays) - Overall sector density

    RaycastGPU360Mode.cpp:104

18.4 Batch Processing

  • Vehicle Batch Actor

    Register/unregister vehicle components

    RaycastGPUVehicleBatch.h:30-50

  • ProcessBatch

    Single GPU dispatch for all vehicles

    RaycastGPUVehicleBatch.cpp:70-200

  • Memory Optimization

    Pre-allocation with Reserve(), MoveTemp

    RaycastGPUVehicleBatch.cpp:85-100

  • UpdateFrequency

    0-120 Hz (0 = every frame)

    RaycastGPUVehicleBatch.h:55

18.5 BVH Acceleration

  • SAH Optimization

    Surface Area Heuristic cost function

    RaycastGPUBVHOptimizer.cpp:50-100

  • Configurable Bins

    32 bins, adaptive leaf sizing

    RaycastGPUBVHOptimizer.cpp:25

  • Branchless AABB

    Vectorized min/max intersection

    FallbackComputeUltraOptimized.usf:70-85

  • Shared Memory Stack

    2048 entries for BVH traversal

    FallbackComputeUltraOptimized.usf:80

18.6 Distance Encoding

  • Level

    Format - Range - Purpose - Source

  • Raw

    FRaycastResult - [0, MaxDist] - Precise distance

    RaycastGPUTypes.h:35-45

  • Normalized

    RayDistances[] - [0, 1] - Neural input

    RaycastGPU360Mode.cpp:52-70

  • Sector

    SectorData[] - [0, 1] × 3 - Aggregate perception

    RaycastGPU360Mode.cpp:100-120

18.7 Performance Characteristics

  • Configuration

    Throughput - Latency

  • 32 rays × 100 agents

    3,200 rays/frame - 1-2 ms

  • 360 rays × 100 agents

    36,000 rays/frame - 5-8 ms

  • 360 rays × 1000 agents

    360,000 rays/frame - 20-30 ms

  • Max (320K rays)

    320,000 rays/frame - 15-25 ms

18.8 NFG Integration

  • UNFGRaycastIntegration

    Direct GPU buffer integration

    NFGRaycastIntegration.cpp:1-100

  • Zero-Copy Transfer

    GPU→GPU data flow without CPU roundtrip

    NFGRaycastIntegration.cpp:70-80

  • UpdateRaycastsForModels

    Batch raycast update for all models

    NFGRaycastIntegration.cpp:85-120

  • 1-3 Frame Latency

    From physics update to network input

    NFGRaycastIntegration.cpp:30-50

TorqueStorm Vehicle Physics

17 Features

19.1 Core Architecture

  • Structure-of-Arrays Layout

    GPU-optimized cache-coherent data

    TorqueStormVehicleData.h:9-45

  • 1000+ Vehicle Capacity

    Scalable to 10,000 vehicles

    TorqueStormConfig.h:15-74

  • O(1) Vehicle Lookup

    TMap-based indexing

    TorqueStormVehicleManager.h:108-109

  • Dirty Transform Tracking

    Only update changed vehicles

    TorqueStormVehicleData.h:40

19.2 Physics Algorithms

  • Ackermann Steering

    Realistic steering geometry

    TorqueStormPhysicsSystem.cpp:200-209

  • Slip Angle Tire Model

    Dynamic grip based on slip

    TorqueStormPhysicsSystem.cpp:244-287

  • Aerodynamic Drag

    Speed² drag formula

    TorqueStormPhysicsSystem.cpp:223-232

  • Dynamic Downforce

    Speed-dependent grip boost

    TorqueStormPhysicsSystem.cpp:234-239

  • Multi-Wheel Suspension

    Per-wheel spring/damper

    TorqueStormPhysicsSystem.cpp:298-438

  • Wheel Raycast Ground Detection

    Physics-accurate ground contact

    TorqueStormPhysicsSystem.cpp:441-470

  • Traction-Limited Forces

    Friction-clamped acceleration

    TorqueStormPhysicsSystem.cpp:156-177

  • Rolling Resistance

    Realistic energy dissipation

    TorqueStormPhysicsSystem.cpp:289-294

  • Physics Substeps

    Configurable 1-4 substeps

    TorqueStormPhysicsSystem.cpp:58-72

19.3 Vehicle Configurations

  • Preset

    Description

  • 4-Wheel Car

    Standard car with driven rear wheels

  • 6-Wheel Truck

    Enhanced stability configuration

  • Formula 1

    Wide front track, narrow rear

Track Generator

13 Features

20.1 Track Generation Algorithms

  • Perlin Noise Layout

    Seed-based deterministic generation

    ProceduralTrack.cpp:139-203

  • Spline Optimization

    Hill-climbing optimization (50 iterations)

    ProceduralTrack.cpp:226-279

  • Self-Intersection Detection

    Prevents overlapping track segments

    ProceduralTrack.cpp:281-308

  • Catmull-Rom Smoothing

    Natural spline curves

    ProceduralTrack.cpp:310-333

  • Parallel Wall Splines

    Left/right boundary generation

    ProceduralTrack.cpp:205-224

20.2 Terrain Generation

  • Multi-Octave Perlin

    6 octaves, 0.35 persistence

    TrackProceduralFloorCPU.h:99-220

  • Domain Warping

    Natural terrain features

    TrackProceduralFloorCPU.cpp

  • GPU Compute Shader

    RDG-based terrain generation

    TerrainCS.cpp

  • Roadside High Detail

    2x density near track

    TrackProceduralFloorCPU.h

20.3 Racing Line & Sectors

  • KD-Tree Spline Cache

    O(log n) nearest-point queries

    TrackSplineCacheComponent.h:19-68

  • Ideal Racing Line

    Speed-optimized path calculation

    TrackIdealLineComponent.h:7-48

  • Sector Division

    Vehicle progress tracking

    TrackGenSectorComponent.h:11-66

  • OnSectorEntered/Exited

    Event-driven sector tracking

    TrackGenSectorComponent.cpp

Quantization System

11 Features

21.1 Bit Depths

  • Bits

    Range - Memory Savings - Source

  • 1-bit

    ±1 (binary) - 97%

    NFGQuantBinaryHelpers.h:6-29

21.2 Quantization Strategies

  • Per-Output Scale

    Per-neuron adaptive scaling

    NFGQuantizationHelper.cpp:219-238

  • Per-Tensor Mode

    Global layer-wide scale

    Hyperparameters.h

  • Symmetric Quantization

    Zero-centered range

    NFGQuantizationHelper.cpp:240-255

  • LSTM Gate Control

    Stability-aware gating

    NFGQuantizationBackend.cpp:38-44

21.3 Dequantization Methods

  • Method

    Use Case - Source

  • Single Value

    Point dequantization

    NFGQuantizationHelper.cpp:257-270

  • Packed Unpacking

    Batch extraction from uint32

    NFGQuantCodec.h:82-114

  • Column Decoding

    GPU matrix operations

    CPUQuantizationHelpers.h:27-66

  • Input Bitplanes

    Multi-channel input encoding

    NFGQuantCodec.h:22-41

Hall of Fittest Persistence

12 Features

22.1 Elite Preservation

  • Binary Search Insertion

    O(log N) ranking

    HallOfFittestComponent.cpp:208-432

  • Fitness/Reward Ranking

    Dual ranking modes

    HallOfFittestComponent.cpp:34-46

  • Elite Replacement

    Lowest-performer eviction

    HallOfFittestComponent.cpp:165-206

  • Entry Recycling

    UObject pool management

    HallOfFittestComponent.cpp:70-111

22.2 Model Serialization

  • Deep Model Copy

    50+ fields per model

    NFGModelComponent.cpp:1180-1299

  • MMS State Arrays

    Full neuron state preservation

    HallOfFittestComponent.cpp:314-413

  • JSON Serialization

    Human-readable format

    NFGModelComponent_Serialization.cpp:17-162

  • Base64 Weight Encoding

    Compact binary storage

    NFGModelComponent_Serialization.cpp

22.3 Validation Guards

  • Guard

    Protection - Source

  • NaN Fitness Detection

    Rejects corrupted models

    HallOfFittestComponent.cpp:185-190

  • TS-ReLU Capacity Validation

    Array size matching

    HallOfFittestComponent.cpp:306-310

  • MMS Array Validation

    Dimensional consistency

    HallOfFittestComponent.cpp:315-413

GPU Vehicle Simulation

17 Features

23.1 Tire Model (Pacejka Magic Formula)

  • Parameter

    Value - Purpose

  • B_long

    10.0 - Longitudinal stiffness

  • C_long

    1.9 - Longitudinal shape factor

  • B_lat

    10.0 - Lateral cornering stiffness

  • C_lat

    1.3 - Lateral shape factor

  • Fz0

    3750N - Reference load

23.2 Suspension System

  • Progressive Spring Rate

    Quadratic stiffening

    VehiclePhysics.usf

  • Asymmetric Damping

    100% compression, 70% extension

    VehiclePhysics.usf

  • Bump Stop Simulation

    95% max compression trigger

    VehiclePhysics.usf

  • Anti-Roll Bar

    Front/rear independent ARB

    VehiclePhysics.usf

23.3 Weight Transfer

  • Longitudinal Transfer

    Acceleration → rear weight gain

  • Lateral Transfer

    Cornering → outside wheel loading

  • CG Height Effect

    Higher CG = more transfer

23.4 Drivetrain

  • Engine Torque Curve

    Texture-based lookup

    VehiclePhysics.usf

  • Gear Ratios

    Per-gear configuration

    VehiclePhysics.usf

  • Clutch Engagement

    Smooth slip simulation

    VehiclePhysics.usf

  • Engine Braking

    Throttle-off deceleration

    VehiclePhysics.usf

GPU RNN - LSTM/GRU/Elman

18 Features

24.1 LSTM Implementation

  • Gate

    Activation - Purpose - Source

  • Input Gate

    Sigmoid - Controls input flow

    LSTM.usf:335-391

  • Forget Gate

    Sigmoid - Controls memory retention

    LSTM.usf:392-441

  • Cell Gate

    Tanh - Candidate cell state

    LSTM.usf:443-492

  • Output Gate

    Sigmoid - Controls output

    LSTM.usf:494-543

24.2 GRU Implementation

  • Gate

    Activation - Purpose - Source

  • Update Gate

    Sigmoid - New vs old state blend

    GRU.usf:283-338

  • Reset Gate

    Sigmoid - Selective memory reset

    GRU.usf:340-393

  • Candidate

    Tanh - New hidden state

    GRU.usf:412-465

24.3 Elman Features

  • Self-Recurrence Mode

    Diagonal weights

    FeedForwardBatch2D.usf:113-115

  • Full Recurrence Mode

    Dense weight matrix

    FeedForwardBatch2D.usf

  • Working Memory

    Short-term latching

    FeedForwardBatch2D.usf:152-154

  • MMS Integration

    Multi-mode neurons

    FeedForwardBatch2D.usf:180-192

24.4 RMGA Per-Gate Modulation

  • Parameter

    Per Gate - Purpose

  • Reward Modulation

    4 LSTM / 2 GRU - Dopamine-like effect

  • Event Threshold

    4 LSTM / 2 GRU - Salience detection

  • Surprise Modulation

    4 LSTM / 2 GRU - Intrinsic motivation

  • Gate Momentum

    4 LSTM / 2 GRU - Temporal smoothing

GPU Physics Engine

14 Features

25.1 Broadphase Collision

  • Spatial Hashing

    Prime-based deterministic hash

    BroadphaseCollision.usf

  • AABB Computation

    8-corner transformation

    BroadphaseCollision.usf

  • CCD via Velocity

    Continuous collision detection

    BroadphaseCollision.usf

  • Collision Filtering

    Group/mask system

    BroadphaseCollision.usf

25.2 Narrowphase (GJK/EPA)

  • Algorithm

    Max Iterations - Purpose - Source

  • GJK

    64 - Collision detection

    NarrowphaseCollision.usf

  • EPA

    64 - Penetration depth

    NarrowphaseCollision.usf

  • Support Functions

    - - Box/Sphere/Capsule

    NarrowphaseCollision.usf

25.3 Constraint Solver

  • Constraint Type

    Description - Source

  • Contact Constraint

    Normal + friction impulses

    ConstraintSolver.usf

25.4 Rigid Body Dynamics

  • Verlet Integration

    Energy-stable

    RigidBodyDynamics.usf

  • Explicit Euler

    Simple, stable

    RigidBodyDynamics.usf

  • Sleep Detection

    Performance optimization

    RigidBodyDynamics.usf

  • Quaternion Rotation

    Deterministic normalization

    DeterministicMath.usf

GPU Analytics & Statistics

17 Features

26.1 Population Statistics

  • Population Mean

    Per-feature mean calculation

    PopulationStats.usf

  • Population Variance

    Numerical stability (ε=1e-6)

    PopulationStats.usf

  • Gaussian Sampling

    Box-Muller transform

    PopulationStats.usf

  • Fitness-Weighted Mean

    CMA-ES adaptive weighting

    PopulationStats.usf

26.2 Predictive Metrics

  • Layer-wise Error

    Per-layer absolute error

    PredictiveMetricsHistogramCS.usf

  • Cortical Column Tracking

    Up to 16 columns/layer

    PredictiveMetricsHistogramCS.usf

  • Per-Model MAE

    Mean absolute error

    PredictiveMetricsReduceCS.usf

  • Per-Model Max Error

    Peak error tracking

    PredictiveMetricsReduceCS.usf

26.3 Top-Down Quantization

  • Bit Width

    Range - Source

  • 8-bit

    [-128, 127]

    PredictiveTopDownTransposeCS.usf

  • 4-bit

    [-8, 7]

    PredictiveTopDownTransposeCS.usf

  • 2-bit

    [-2, 1]

    PredictiveTopDownTransposeCS.usf

  • 1-bit

    ±1

    PredictiveTopDownTransposeCS.usf

26.4 Evolutionary Features

  • Layer-Wise Extraction

    Dynamic state sampling

    EvoFeaturesCS.usf

  • EMA Filtering

    Multi-timescale retention

    EvoFeaturesCS.usf

  • Reward Injection

    Current reward as input

    EvoFeaturesCS.usf

  • Reward Delta

    First-order change tracking

    EvoFeaturesCS.usf

AgentData System

17 Features

27.1 Data Types (8 Classes)

  • FloatData

    Primary neural type with genetic crossover

    FloatData.cpp

  • IntData

    Discrete output type (steering/throttle)

    IntData.cpp

  • FVectorData

    3D vector with component-wise crossover

    FVectorData.cpp

  • FLinearColorData

    RGBA color with alpha crossover

    FLinearColorData.cpp

  • StringData

    Text output with character-level crossover

    StringData.cpp

  • TimespanData

    Temporal data with magnitude crossover

    TimespanData.cpp

  • BaseData

    Abstract base with virtual operations

    BaseData.cpp

  • AgentData

    Container with ordered insert

    AgentData.cpp

27.2 Crossover Features

  • Virtual Crossover

    Type-specific blend operations

    BaseData::Crossover()

  • Dynamic Cast Safety

    Runtime type checking

    FloatData::Crossover():44

  • Alpha-Weighted Blend

    Configurable parent weighting - All data types

  • Character Splicing

    String-level genetic recombination

    StringData::Crossover()

  • Component Independence

    XYZ crossover for vectors

    FVectorData::Crossover()

27.3 Serialization

  • Dual Serialization

    Both Save/Load paths

    FloatData::Serialize()

  • Binary Archive

    Compact network format - All data types

  • Type Preservation

    Maintains class hierarchy

    BaseData.cpp

  • Ordered Container

    Index-based retrieval

    AgentData::AddByName()

Camera Systems

20 Features

28.1 Free Camera Component

  • 6-DOF Movement

    Full spatial navigation

    NFGFreeCameraComponent.cpp:78-135

  • Acceleration Model

    Smooth input response

    NFGFreeCameraComponent.cpp:110

  • Deceleration Braking

    Friction-based stopping

    NFGFreeCameraComponent.cpp:118

  • Max Speed Limit

    Configurable velocity cap

    NFGFreeCameraComponent.cpp:95

  • Roll/Pitch/Yaw

    Independent rotation axes

    NFGFreeCameraComponent.cpp:130-133

  • Look Sensitivity

    Adjustable turn rate

    NFGFreeCameraComponent.cpp:100

28.2 Input Configuration

  • Enhanced Input System

    UE5 Input Actions

    NFGFreeCameraInputConfig.cpp

  • Action Binding

    Forward/Back/Left/Right/Up/Down

    NFGFreeCameraInputConfig.cpp:25-48

  • Look Action

    Mouse/gamepad camera

    NFGFreeCameraInputConfig.cpp:50

  • Dynamic Context

    Runtime rebinding

    NFGFreeCameraInputConfig.cpp:18

28.3 NFG Vision Component

  • Scene Depth Capture

    GPU depth buffer sampling

    NFGSceneNFGVisionComponent.cpp:45-120

  • Resolution Config

    Adjustable capture size

    NFGSceneNFGVisionComponent.h:35

  • Gaze Direction

    AI-controlled look vector

    NFGSceneNFGVisionComponent.cpp:78

  • FOV Control

    Adjustable field of view

    NFGSceneNFGVisionComponent.h:38

  • Async Readback

    Non-blocking GPU transfer

    NFGSceneNFGVisionComponent.cpp:95

  • Depth Normalization

    Range mapping to [0,1]

    NFGSceneNFGVisionComponent.cpp:110

28.4 Rendering Features

  • Render Target Pool

    Reusable GPU resources

    NFGSceneNFGVisionComponent.cpp:50

  • Scene Capture 2D

    UE5 capture component

    NFGSceneNFGVisionComponent.cpp:55

  • Material Instance

    Custom depth visualization

    NFGSceneNFGVisionComponent.cpp:65

  • Debug Visualization

    Optional depth preview

    NFGSceneNFGVisionComponent.cpp:125

Mutation System

20 Features

29.1 Mutation Operators

  • Operator

    Description - Source

  • Adaptive Gaussian

    σ scales with fitness rank

    MutationComponent.cpp:85

  • Elite Thawing

    Frozen → unfrozen transitions

    MutationComponent.cpp:142

  • Mirrored Sampling

    Antithetic pairs for variance reduction

    MutationComponent.cpp:168

  • Orthogonal Mutation

    Gram-Schmidt decorrelation

    MutationComponent.cpp:195

29.2 Rate Scheduling

  • Strategy

    Description - Source

  • Constant Rate

    Fixed probability

    MutationComponent.cpp:45

  • Linear Decay

    Rate decreases with generation

    MutationComponent.cpp:52

  • Cosine Annealing

    Cyclical warm restarts

    MutationComponent.cpp:60

  • Adaptive (1/5 Rule)

    Rechenberg's rule

    MutationComponent.cpp:68

  • Layer-Specific

    Different rates per layer

    MutationComponent.cpp:78

  • Self-Adaptive

    Evolved strategy parameters

    MutationComponent.cpp:92

29.3 Layer-Specific Scaling

  • Input Sensitivity

    Reduced mutation near inputs

    MutationComponent.cpp:210

  • Output Protection

    Conservative output mutations

    MutationComponent.cpp:218

  • Hidden Plasticity

    Higher rates for hidden layers

    MutationComponent.cpp:225

  • Depth Scaling

    Rate scales with layer depth

    MutationComponent.cpp:232

29.4 Genome Integrity

  • Weight Bounds

    Prevents explosion

    MutationComponent.cpp:245

  • Bias Protection

    Separate bias mutation

    MutationComponent.cpp:252

  • Structural Freeze

    Topology preservation

    MutationComponent.cpp:260

  • Gradient Estimation

    Finite difference approximation

    MutationComponent.cpp:268

Visualization System

9 Features

31.1 Visualization States

  • State

    Description - Source

31.3 Layout System

  • Layer Spacing

    Configurable Z-distance

    VisualizationComponent_Layout.cpp:45

  • Neuron Grid

    2D arrangement per layer

    VisualizationComponent_Layout.cpp:68

  • Connection Bezier

    Curved weight lines

    VisualizationComponent_Layout.cpp:95

  • Auto-Fit

    Scale to viewport

    VisualizationComponent_Layout.cpp:125

31.4 Real-Time Updates

  • Activation Streaming

    Per-frame neuron updates

    VisualizationComponent_Process.cpp:35

  • Color Gradient

    Activity → color mapping

    VisualizationComponent_Process.cpp:60

  • Weight Animation

    Smooth transition effects

    VisualizationComponent_Process.cpp:85

  • Async GPU Readback

    Non-blocking data fetch

    VisualizationComponent_Process.cpp:110

NFGManager Subsystems

28 Features

32.1 Backend Subsystem

  • Backend Factory

    CPU/GPU backend selection

    NFGManager_Backend.cpp:45-85

  • Provider Registration

    Dynamic component binding

    NFGManager_Backend.cpp:90-130

32.2 Evolution Subsystem

  • Generation Management

    Epoch lifecycle

    NFGManager_Evolution.cpp:45-95

  • Fitness Aggregation

    Population ranking

    NFGManager_Evolution.cpp:100-145

32.3 FeedForward Subsystem

  • Batch Processing

    Population-wide inference

    NFGManager_FeedForward.cpp:45-100

  • Timing Control

    Fixed timestep management

    NFGManager_FeedForward.cpp:105-145

  • Input Injection

    Sensor → network mapping

    NFGManager_FeedForward.cpp:150-195

  • Output Extraction

    Network → actuator mapping

    NFGManager_FeedForward.cpp:200-245

32.5 Initialization Subsystem

  • Config Loading

    JSON hyperparameter parse

    NFGManager_Init.cpp:45-100

  • Component Spawn

    Dynamic actor creation

    NFGManager_Init.cpp:105-160

  • Provider Binding

    Interface wire-up

    NFGManager_Init.cpp:165-220

  • Validation

    Architecture consistency

    NFGManager_Init.cpp:225-270

32.6 Lifecycle Subsystem

  • BeginPlay Hook

    Initialization sequence

    NFGManager_Lifecycle.cpp:45-85

  • Tick Management

    Update orchestration

    NFGManager_Lifecycle.cpp:90-140

32.7 Model Subsystem

  • Architecture Builder

    Network topology creation

    NFGManager_Model.cpp:45-110

  • Weight Initialization

    Xavier/He/Orthogonal

    NFGManager_Model.cpp:115-175

  • Clone Operations

    Model duplication

    NFGManager_Model.cpp:180-230

  • Comparison

    Architecture diff tools

    NFGManager_Model.cpp:235-280

32.8 Population Subsystem

  • Spawn Management

    Agent creation/destruction

    NFGManager_Population.cpp:45-105

  • Reset Coordination

    Episode boundary handling

    NFGManager_Population.cpp:110-165

  • Statistics Tracking

    Population metrics

    NFGManager_Population.cpp:170-225

  • Diversity Monitoring

    Genotype variance

    NFGManager_Population.cpp:230-280

32.9 Render Subsystem

  • Debug Drawing

    Real-time visualization

    NFGManager_Render.cpp:45-100

32.10 Telemetry Subsystem

  • Metric Collection

    Performance counters

    NFGManager_Telemetry.cpp:45-100

  • Event Logging

    Timestamped events

    NFGManager_Telemetry.cpp:105-155

  • Export Formats

    CSV/JSON/Binary

    NFGManager_Telemetry.cpp:160-210

  • Real-Time Streaming

    External tool support

    NFGManager_Telemetry.cpp:215-260

32.11 Widget Subsystem

  • Slate Integration

    UE5 UI framework

    NFGManager_Widgets.cpp:45-95

Backend & Provider Interfaces

16 Features

33.1 Core Interfaces

  • Interface

    Purpose - Source

  • INFGBackendProvider

    CPU/GPU backend abstraction

    NFGBackendProvider.h

  • INFGFitnessProvider

    Fitness computation

    INFGFitnessProvider.h

  • INFGRewardProvider

    Reward signal generation

    INFGRewardProvider.h

  • INFGMTDPProvider

    Multi-timescale delay

    INFGMTDPProvider.h

33.2 Learning Interfaces

  • Interface

    Purpose - Source

  • INFGModulatorGainsProvider

    RMGA neuromodulation

    INFGModulatorGainsProvider.h

  • INFGMetaRegulatorProvider

    UCB meta-learning

    INFGMetaRegulatorProvider.h

  • INFGMetaProgressProvider

    Learning progress tracking

    INFGMetaProgressProvider.h

  • INFGWhatIfStateProvider

    Counterfactual rollouts

    INFGWhatIfStateProvider.h

33.3 Telemetry Interfaces

  • Interface

    Purpose - Source

  • NFGChildQueueTelemetryProvider

    Offspring queue monitoring

    NFGChildQueueTelemetryProvider.h

  • INFGTelemetryProvider

    General metrics export

    INFGTelemetryProvider.h

33.4 Backend Factory

  • Runtime Selection

    CPU/GPU based on hardware

    NFGBackendFactory.cpp:45-85

  • Capability Detection

    Feature availability

    NFGBackendFactory.cpp:90-130

  • Fallback Chain

    Graceful degradation

    NFGBackendFactory.cpp:135-170

Profile & Version System

20 Features

34.1 Profile Hot-Swap

  • Live Config Reload

    Change parameters without restart

    NFGProfile.cpp:45-95

  • Partial Updates

    Selective parameter modification

    NFGProfile.cpp:100-145

  • Rollback Support

    Undo configuration changes

    NFGProfile.cpp:150-195

  • Validation Pipeline

    Pre-apply consistency checks

    NFGProfile.cpp:200-245

34.2 Configuration Merging

  • 11-File Config System

    Distributed hyperparameters

    NFGProfile.cpp:250-310

  • Override Priority

    User > Project > Default

    NFGProfile.cpp:315-360

  • Type Coercion

    Safe JSON type conversion

    NFGProfile.cpp:365-405

  • Missing Key Handling

    Default value injection

    NFGProfile.cpp:410-450

34.5 Config Files

  • Config File

    Purpose - Source

  • ArchitectureConfig.json

    Network topology

    Config/ArchitectureConfig.json

  • EvolutionConfig.json

    CMA-ES parameters

    Config/EvolutionConfig.json

  • FitnessConfig.json

    Reward weights

    Config/FitnessConfig.json

  • AttentionConfig.json

    Attention system

    Config/AttentionConfig.json

  • RMGAConfig.json

    Neuromodulation

    Config/RMGAConfig.json

  • WorldPredictorConfig.json

    Prediction system

    Config/WorldPredictorConfig.json

  • MTDPConfig.json

    Multi-timescale delay

    Config/MTDPConfig.json

  • NeuroDynamicsConfig.json

    STDP parameters

    Config/NeuroDynamicsConfig.json

  • BackendConfig.json

    CPU/GPU selection

    Config/BackendConfig.json

  • SensorsConfig.json

    Input configuration

    Config/SensorsConfig.json

  • GameConfig.json

    Runtime settings

    Config/GameConfig.json

Sleep Replay System

23 Features

35.1 Configuration

  • bEnabled

    true - Master toggle

    Hyperparameters.h:3485

  • ExperienceBufferSize

    256 - Max experiences retained

    Hyperparameters.h:3488

  • SampleRate

    1 - Recording frequency

    Hyperparameters.h:3491

  • SleepIntervalFrames

    256 - Replay trigger threshold

    Hyperparameters.h:3494

  • ReplayBatchSize

    32 - Samples per replay batch

    Hyperparameters.h:3497

  • bPrioritizeHighReward

    true - High-reward sampling

    Hyperparameters.h:3503

  • bProcessAsync

    false - Async replay mode

    Hyperparameters.h:3506

  • AsyncSamplesPerTick

    8 - Max async samples/tick

    Hyperparameters.h:3509

35.2 Data Structures

  • Structure

    Description - Source

  • FNFGSleepReplayFrame

    Activation snapshot + reward

    NFGNeuroDynamicsProvider.h:42

  • FNFGSleepReplayBuffer

    Circular experience buffer

    NFGNeuroDynamicsProvider.h:50

  • FNFGNeuroDynamicsModelState

    Per-model replay state

    NFGNeuroDynamicsProvider.h:62

35.3 Core Features

  • Experience Recording

    Per-frame activation capture

    NFGNeuroDynamicsProvider.cpp:1167

  • Priority Sampling

    High-reward experience selection

    NFGNeuroDynamicsProvider.cpp:1209

  • Synchronous Replay

    Immediate batch processing

    NFGNeuroDynamicsProvider.cpp:1246

  • Asynchronous Replay

    Distributed across frames

    NFGNeuroDynamicsProvider.cpp:1283

  • Global Replay Flag

    Cross-provider coordination

    NFGNeuroDynamicsProvider.cpp:121

  • EWMA Gain Tracking

    Smooth weight change metrics

    NFGNeuroDynamicsProvider.cpp:1493

35.4 Console Commands

  • Command

    Description - Source

  • nfg.DumpSleepReplayStats

    Show replay statistics

    NFGNeuroDynamicsProvider.cpp:2223

  • nfg.DumpSleepReplayEntry

    Inspect single experience

    NFGNeuroDynamicsProvider.cpp:2244

  • nfg.DumpStdpProfile

    STDP timing statistics

    NFGNeuroDynamicsProvider.cpp:2272

  • nfg.ResetStdpProfile

    Clear profiling data

    NFGNeuroDynamicsProvider.cpp:2317

Kernel Components

19 Features

36.1 CMA-ES Kernels

  • Kernel

    Description - Source

  • cmaes_init

    State initialization

    CMAESComponent.cpp:25

  • sample_multivariate_normal

    Box-Muller Gaussian

    CMAESComponent.cpp:108

  • cmaes_update

    Main CMA-ES update

    CMAESComponent.cpp:183

  • cmaes_rank_mu_update

    Covariance rank-mu

    CMAESComponent.cpp:291

  • cmaes_eigen_decomposition

    Power iteration eigendecomp

    CMAESComponent.cpp:349

  • cmaes_generate_samples

    Population sampling

    CMAESComponent.cpp:498

  • cmaes_compute_invsqrtc

    Inverse sqrt covariance

    CMAESComponent.cpp:604

36.2 Crossover Components

  • Component

    Methods - Source

  • CrossoverElmanComponent

    6 crossover methods + 4 mutation modes

    CrossoverElmanComponent.cpp

  • CrossoverLSTMComponent

    LSTM gate-specific crossover

    CrossoverLSTMComponent.cpp

  • CrossoverBatchNormComponent

    γ/β parameter crossover

    CrossoverBatchNormalizationComponent.cpp

  • CrossoverMinMaxComponent

    Weight bounds computation

    CrossoverMinMaxComponent.cpp

  • CrossoverRLComponent

    Reward-gated crossover

    CrossoverReinforcementLearningComponent.cpp

36.3 Activation Functions (10 types)

  • Function

    Formula - Source

36.4 Kernel Modules

  • BatchNormalization

    Calculate + apply BN

    CalculateBatchStatistics.h

  • Dropout

    LCG-based regularization

    DropOut.h

  • Clamp

    Value clamping utility

    Clamp.h

  • Randomizer

    LCG random generation

    Randomizer.h

Statistics & Monitoring

19 Features

37.1 DeathStatisticsComponent

  • Multi-Scale Tracking

    Minute/Hour/Day counters

    DeathStatisticsComponent.h:24-51

  • Death Rate Calculation

    Normalized by time unit

    DeathStatisticsComponent.cpp:41

  • Timer-Based Updates

    60s/3600s/86400s intervals

    DeathStatisticsComponent.cpp:17

  • Ring Buffer History

    1440/24/30 entry capacities

    DeathStatisticsComponent.h:154-170

37.2 FPSComponent

  • Real-Time FPS

    1/DeltaTime calculation

    FPSComponent.cpp:70

  • Tick Profiling

    Conditional timing macro

    FPSComponent.cpp:9

37.3 OutputHistoryComponent

  • Ring Buffer Architecture

    Circular output history

    OutputHistoryComponent.h:74

  • Scratch Buffer Optimization

    Zero-allocation updates

    OutputHistoryComponent.h:81

  • Flattened Cache

    Lazy evaluation pattern

    OutputHistoryComponent.h:84

  • Configurable History

    Steering/Throttle/Brake flags

    OutputHistoryComponent.h:99-107

  • Weighted History

    Linear decay weighting

    OutputHistoryComponent.cpp:75

37.4 KernelPerformanceMonitor

  • Kernel Timing

    BeginKernelExecution/End

    KernelPerformanceMonitor.cpp:72

  • GFLOPS Estimation

    Operations/second metrics

    KernelPerformanceMonitor.h:36

  • Memory Bandwidth

    GB/s estimation

    KernelPerformanceMonitor.h:39

  • Performance Snapshots

    Timestamped state capture

    KernelPerformanceMonitor.cpp:117

  • CSV Export

    Performance data export

    KernelPerformanceMonitor.cpp:162

  • Issue Detection

    20% degradation threshold

    KernelPerformanceMonitor.cpp:262

37.5 SinusCalculator

  • Oscillating Parameters

    Sine wave for mutation rates

    SinusCalculator.cpp:12

  • Configurable Wave

    Frequency/Min/Max/Offset

    SinusCalculator.h:18-31

Weather Subsystems Extended

27 Features

38.1 Weather Friction System

  • Weather-Based Friction

    Dry/Wet/Snow/Ice coefficients

    NFGWeatherFrictionComponent.h:14-45

  • Vehicle Registration

    Per-vehicle friction updates

    NFGWeatherFrictionComponent.h:78

  • Smooth Transitions

    Interpolated friction changes

    NFGWeatherFrictionComponent.h:123

38.2 Weather Front Manager

  • Detailed Front Physics

    Atmospheric layers + air masses

    NFGWeatherFrontManager.h:14-75

  • Front Types

    Cold/Warm/Occluded/Stationary

    NFGWeatherFrontManager.h:96-105

  • Geostrophic Wind

    Pressure gradient calculation

    NFGWeatherFrontManager.h:211

  • Thermal Wind

    Temperature-driven vectors

    NFGWeatherFrontManager.h:212

  • Convergence/Divergence

    Air mass dynamics

    NFGWeatherFrontManager.h:213-214

38.3 Weather State Machine

  • State

    Description - Source

  • Clear

    No precipitation

    NFGWeatherStateMachine.h:16

  • Cloudy

    Cloud coverage

    NFGWeatherStateMachine.h:17

  • Overcast

    Full cloud cover

    NFGWeatherStateMachine.h:18

  • LightRain

    Light precipitation

    NFGWeatherStateMachine.h:19

  • HeavyRain

    Heavy precipitation

    NFGWeatherStateMachine.h:20

  • Thunderstorm

    Lightning + rain

    NFGWeatherStateMachine.h:21

  • Snow

    Snow precipitation

    NFGWeatherStateMachine.h:22

  • Fog

    Reduced visibility

    NFGWeatherStateMachine.h:23

38.4 Atmospheric Simulation

  • Day/Night Cycle

    24-hour simulation

    NFGWeatherSimulationComponent.h:63

  • Temperature Field

    Altitude + time based

    NFGWeatherSimulationComponent.h:73-79

  • Pressure System

    Sea-level + cells

    NFGWeatherSimulationComponent.h:86

  • Humidity Field

    Base + variation

    NFGWeatherSimulationComponent.h:93-96

  • Wind System

    Base + turbulence + gusts

    NFGWeatherSimulationComponent.h:100-106

  • Atmospheric Grid

    100x100 cells, 1km size

    NFGWeatherSimulationComponent.h:110-116

  • Deterministic Mode

    Seeded weather

    NFGWeatherSimulationComponent.h:155

38.5 Integrated Weather System

  • Weather Forecasting

    N-hour predictions

    NFGIntegratedWeatherSystem.h:128

  • Severe Weather Alerts

    Thunderstorm warnings

    NFGIntegratedWeatherSystem.h:174

  • Weather Cycling

    Automated state progression

    NFGIntegratedWeatherSystem.h:111

WorldPredictor GPU Shaders

17 Features

39.1 Action Context

  • Action History Injection

    Previous actions as input

    WorldPredictorActionContext.usf

  • Ring Buffer Shifting

    Multi-frame history (1-4)

    WorldPredictorActionContext.usf

  • Warmup Handling

    Zero output for frame 0

    WorldPredictorActionContext.usf

39.2 What-If Planning

  • Multi-Step Rollout

    Hidden state prediction

    WorldPredictorWhatIf.usf

  • Confidence Scoring

    RMSE-based stability

    WorldPredictorWhatIf.usf

  • Danger Scoring

    Safety neuron monitoring

    WorldPredictorWhatIf.usf

  • Ping-Pong Buffers

    Efficient multi-step

    WorldPredictorWhatIf.usf

39.3 Action Correction

  • Max-Score Softmax

    Query weighting

    WorldPredictorActionCorrection.usf

  • Learned Confidence

    Network-predicted certainty

    WorldPredictorActionCorrection.usf

  • Counterfactual Feedback

    Per-query error deltas

    WorldPredictorActionCorrection.usf

39.4 Action Variants

  • 3/5 Variant Modes

    ±small, ±large perturbations

    WorldPredictorActionVariants.usf

  • Adaptive Horizon

    Learned planning depth

    WorldPredictorActionVariants.usf

39.5 Loss Functions

  • MSE/Huber Loss

    Prediction error

    WorldPredictorMSE.usf

  • Token Cross-Entropy

    Discrete prediction

    WorldPredictorTokenCE.usf

  • Multi-Timescale Error

    Multiple horizon scales

    WorldPredictorMultiTimescale.usf

  • Consistency Error

    Cross-horizon validation

    WorldPredictorConsistency.usf

  • Per-Neuron Error

    EWMA smoothed feedback

    WorldPredictorPerNeuronError.usf

Model Component Helpers

24 Features

40.1 Weight Initialization Strategies

  • Strategy

    Description - Source

  • RandomUniform

    Uniform [-range, range]

    NFGModelComponent_Weights.cpp:55

  • RandomNormal

    Gaussian distribution

    NFGModelComponent_Weights.cpp:60

  • UniformXavier

    sqrt(6/(in+out))

    NFGModelComponent_Weights.cpp:65

  • NormalXavier

    sqrt(2/(in+out))

    NFGModelComponent_Weights.cpp:70

  • UniformHE

    sqrt(6/in)

    NFGModelComponent_Weights.cpp:75

  • NormalHE

    sqrt(2/in)

    NFGModelComponent_Weights.cpp:80

  • Orthogonal

    QR decomposition

    NFGModelComponent_Helpers.cpp:27

  • Sparse

    Reservoir sampling

    NFGModelComponent_Helpers.cpp:120

  • LSUV

    Layer-Sequential Unit-Variance

    NFGModelComponent_Helpers.cpp:225

  • Identity

    Diagonal + noise

    NFGModelComponent_Helpers.cpp:258

  • DeltaOrthogonal

    Orthogonal + scaled identity

    NFGModelComponent_Helpers.cpp:287

  • Fixup

    Depth-scaled residual

    NFGModelComponent_Helpers.cpp:313

  • Adaptive

    Auto-selection by context

    NFGModelComponent_Helpers.cpp:344

40.2 Serialization

  • Binary Serialization

    FArchive-based

    NFGModelComponent_Binary.cpp:10

  • JSON Serialization

    Human-readable

    NFGModelComponent_Serialization.cpp:17

  • Quantization Payload

    Compressed weights

    NFGModelComponent_Serialization.cpp:36

  • Working Memory Snapshot

    State preservation

    NFGModelComponent_Serialization.cpp:272

40.3 History Tracking

  • Output History

    Circular buffer

    NFGModelComponent_History.cpp:6

  • Reward History

    Modulo-indexed

    NFGModelComponent_History.cpp:37

  • Ordered Retrieval

    Chronological access

    NFGModelComponent_History.cpp:62

40.4 Lifecycle Management

  • Auto-Registration

    Manager binding

    NFGModelComponent_Lifecycle.cpp:79

  • Retry Logic

    0.5s timer fallback

    NFGModelComponent_Lifecycle.cpp:93

  • TS-ReLU Init

    Per-neuron parameters

    NFGModelComponent_Lifecycle.cpp:336

Orchestration Components

17 Features

41.1 NFGInputRelayComponent

  • Source Type Enum

    Pre-resolved 8 types

    NFGInputRelayComponent.h:10

  • JSON Mapping

    Config/InputMapping.json

    NFGInputRelayComponent.cpp:76

  • O(1) Resolution

    Switch instead of strings

    NFGInputRelayComponent.cpp:135

41.2 NFGPopulationComponent

  • Model Creation

    UNFGModelComponent spawn

    NFGPopulationComponent.cpp:19

  • NFGVision Auto-Reserve

    Input space allocation

    NFGPopulationComponent.cpp:31

  • Flat Array Preparation

    Unified GPU buffers

    NFGPopulationComponent.cpp:153

  • LSTM/GRU Detection

    Automatic RNN setup

    NFGPopulationComponent.cpp:225-243

  • TS-ReLU Arrays

    8 per-neuron parameter arrays

    NFGPopulationComponent.cpp:370

41.3 NFGEvolutionComponent

  • Model Crossover

    Hall of Fittest parents

    NFGEvolutionComponent.cpp:29

  • GPU Exclusive Check

    CMA-ES contention avoidance

    NFGEvolutionComponent.cpp:40

  • AgentData Crossover

    Element-wise additional data

    NFGEvolutionComponent.cpp:96

  • CMA-ES Tick

    Frame-based update trigger

    NFGEvolutionComponent.cpp:151

  • Progressive Init

    6 feature size fallbacks

    NFGEvolutionComponent.cpp:197

41.4 NFGTrainingComponent

  • Sync FeedForward

    Immediate batch dispatch

    NFGTrainingComponent.cpp:11

  • Backend Selection

    GPU or CPU dispatch

    NFGTrainingComponent.cpp:18

41.5 NFGPersistenceComponent

  • Agent Save

    Manager delegation

    NFGPersistenceComponent.cpp:10

  • Agent Load

    Manager delegation

    NFGPersistenceComponent.cpp:17

Debug & Vehicle Systems

21 Features

42.1 DebugManager

  • AddDebugLine

    Scaled plane rendering

    DebugManager.cpp:66

  • AddDebugSphere

    HISM sphere instances

    DebugManager.cpp:101

  • AddDebugPlane

    Oriented plane rendering

    DebugManager.cpp:130

  • Per-Frame Clear

    Instance reset each tick

    DebugManager.cpp:59

42.2 ActorViewData

  • GetViewMatrix

    4x4 view transform

    ActorViewData.cpp:6

  • GetProjectionMatrix

    Reversed-Z perspective

    ActorViewData.cpp:23

  • GetViewProjectionMatrix

    Combined VP matrix

    ActorViewData.cpp:40

42.3 NFGGPUVehicleComponent

  • Wheel Physics

    18 parameters per wheel

    NFGGPUVehicleComponent.h:12-100

  • Engine Simulation

    RPM, torque, 8-speed transmission

    NFGGPUVehicleComponent.h:105-177

  • Aerodynamics

    Drag + downforce

    NFGGPUVehicleComponent.h:182-210

  • Suspension Model

    Spring-damper physics

    NFGGPUVehicleComponent.cpp:561

  • Tire Forces

    Slip ratio/angle calculation

    NFGGPUVehicleComponent.cpp:640

  • Fixed Timestep

    60Hz (16.67ms)

    NFGGPUVehicleComponent.h:508

42.4 VehiclePhysicsComputeComponent

  • GPU Buffer Management

    34 Float4s per vehicle

    VehiclePhysicsComputeComponent.h:220

  • Async Readback

    Double-buffered results

    VehiclePhysicsComputeComponent.h:260

  • SoA Optimization

    Structure-of-Arrays layout

    VehiclePhysicsComputeComponent.h:280

  • Telemetry Readback

    Speed, G-forces, yaw rate

    VehiclePhysicsComputeComponent.h:50

42.5 Tutorial1Component

  • 12-State Animation

    Neural network visualization

    Tutorial1Component.h:21-37

  • HISM Neurons

    Instanced sphere rendering

    Tutorial1Component.h:54

  • Spline Connections

    Weight visualization

    Tutorial1Component.cpp:400

  • Interactive Controls

    Next/Previous/Stop

    Tutorial1Component.h:120-137

GPU Shader Architecture

46 Features

43.1 Attention Engine Shaders

  • AttentionFusedCS

    Fused Q·K/√d attention with wave-level reductions

    AttentionFusedCS.usf

  • Half-Precision Packing

    4× bandwidth reduction via packed half Q/K/V

    AttentionFusedCS.usf:55-60

  • Per-Model Skip Masking

    Selective attention during evolution

    AttentionFusedCS.usf:114

  • Additive vs Multiplicative

    Residual connection option

    AttentionFusedCS.usf:328-341

  • Adaptive Softmax

    Log-sum-exp numerical stability

    AttentionFusedCS.usf:256-268

  • Fast Exp Option

    exp2-based approximation

    AttentionFusedCS.usf:68-72

  • Wave-Level Reduction

    WaveActiveSum() intra-warp sum

    AttentionFusedCS.usf:223-225

  • Per-Head LUT

    Optional per-head weight lookup

    AttentionFusedCS.usf:181-185

  • Inline Mean Accumulation

    SM 6.6+ atomic float

    AttentionFusedCS.usf:347

43.2 Temporal Memory Shaders

  • 8-Cascade EMA

    Timescales from 16ms to 3 minutes

    TemporalMemoryCS.usf:44-53

  • 64-Float State Buffer

    Per-model retention state

    TemporalMemoryCS.usf:55-73

  • 12 Composite Neurons

    Blended memory features

    TemporalMemoryCS.usf:174-247

  • Per-Neuron Sparsity

    Active neuron tracking

    TemporalMemoryCS.usf:116-133

  • Reward Cascade

    Multi-timescale reward tracking

    TemporalMemoryCS.usf:40-47

  • Prediction Error Cascade

    WorldPredictor integration

    TemporalMemoryCS.usf:48-55

  • Crisis Signal Detection

    Rapid deterioration detection

    TemporalMemoryCS.usf

43.3 World Predictor Shaders (DreamerV3-Style)

  • What-If Rollouts

    Multi-step imagination

    WorldPredictorWhatIf.usf

  • Ping-Pong Buffers

    Write hazard avoidance

    WorldPredictorWhatIf.usf:120-182

  • Action Conditioning

    H' = tanh(W_hh·H + W_ha·A + b)

    WorldPredictorWhatIf.usf:125-160

  • Confidence Scoring

    RMSE-based confidence

    WorldPredictorWhatIf.usf:244-247

  • Danger Scoring

    Multi-step decay weighting

    WorldPredictorWhatIf.usf:188-241

  • Action Context History

    DreamerV3-style action memory

    WorldPredictorActionContext.usf:56-80

  • In-Place History Shift

    Low-to-high copy pattern

    WorldPredictorActionContext.usf:86-97

  • Softmax Query Weighting

    Temperature-scaled action selection

    WorldPredictorActionCorrection.usf:76-94

  • Learned Confidence Output

    Network confidence head

    WorldPredictorActionCorrection.usf:100-132

  • Counterfactual Feedback

    Regret signal computation

    WorldPredictorActionCorrection.usf:170-190

  • Bitplane Token Encoding

    Discrete token reconstruction

    WorldPredictorTokenCE.usf:25-46

  • Horizon-Weighted CE

    Exponential decay for multi-step

    WorldPredictorTokenCE.usf:70-112

43.4 STDP Learning Shaders

  • Attention De-Gating

    Recover pre-attention activation

    STDP.usf:124-149

  • Prediction-Based Learning

    Error-driven weight updates

    STDP.usf:151-189

43.5 RMGA Neuromodulation Shaders

  • Dopamine Computation

    Reward - RewardEWMA

    RMGAModulators.usf:220-222

  • Noradrenaline

    max(RewardSurprise, HiddenSurprise)

    RMGAModulators.usf:225-236

  • Acetylcholine

    Dual EMA context shift detection

    RMGAModulators.usf:238-261

  • Serotonin

    Volatility + NegativeRatio + Drawdown

    RMGAModulators.usf:263-284

  • Welford's Algorithm

    Online z-score normalization

    RMGAModulators.usf:286-289

  • Serotonin De-Correlation

    Remove DA component from 5-HT

    RMGAModulators.usf:291-292

  • Hybrid Spiking

    Burst when threshold crossed

    RMGAModulators.usf:294-297

  • Gate Gain Hysteresis

    Prevent oscillation

    RMGAModulators.usf:304-306

43.6 RNN Architecture Shaders

  • Quantized Weight Decode

    8/4/2/1-bit unpacking

    LSTM.usf:14-58

  • Per-Gate RMGA

    4 sets per LSTM

    LSTM.usf:101-123

  • Hybrid Spiking State

    LSTMSpikePotentialBuffer

    LSTM.usf:130-131

  • GRU 2-Gate

    30% faster than LSTM

    GRU.usf

  • Gate Equations

    Update/Reset/Candidate/Final

    GRU.usf

43.8 CMA-ES Evolution Shaders

  • Identity Covariance Init

    MeanVector from weights

    CMAES.usf:40-88

  • NaN/Inf Guard

    Zero-out invalid values

    CMAES.usf:62-65

  • Box-Muller Sampling

    Gaussian from uniform - CMAESSampleGeneration

Unit Test Suite

22 Features

44.1 Test Coverage Statistics

  • Metric

    Value

  • Test Cases

    350+

  • Lines of Test Code

    ~8,000+

  • Helper/Fixture Files

    12+

  • Plugin

    Test Files - Key Coverage

  • NFGTemporalRetention

    2 - Buffer guards, auto-attach

  • NFGNeuroDynamics

    1 - Sleep buffer, plasticity defaults

  • NFGMetaRegulator

    3 - UCB exploration, progress guards

  • NFGMultiModulatorRMGA

    2 - Crossover coherence, reward normalization

44.3 Test Patterns

  • Pattern

    Examples

  • Mock Backends

    FMockNFGLMBackend, FTestNeuroProvider

  • Fixture Reuse

    CreateTestModel(), MakeAttentionManager()

  • GPU Resource Tests

    Buffer creation, UAV/SRV lifecycle

  • Roundtrip Tests

    Encode→Decode, Serialize→Deserialize

  • Guard Tests

    Error path validation with AddExpectedError()

  • Determinism Tests

    Same input → same output validation

44.4 Edge Cases Covered

  • System

    Edge Cases

  • Quantization

    Boundary values, 1-8 bit variation, packing alignment

  • NFGLM

    UTF-8 multi-byte, END token mid-stream, vocab coverage

  • CMA-ES

    NaN injection, bounds clamping, covariance symmetry

  • GPU→CPU

    Bias layout mismatch, buffer invalidity, RDG fallback

  • Serialization

    Missing arrays, malformed JSON, MMS state consistency

Configuration System

22 Features

General

  • File

    Parameters - Purpose

  • ArchitectureConfig.json

    108 - Network structure, MMS, activations, weight init

  • AttentionConfig.json

    32 - Attention modes, heads, temporal window

  • BackendConfig.json

    27

    GPU/CPU backend, precision, quantization

  • EvolutionConfig.json

    87 - Mutation, crossover, parent selection

  • FitnessConfig.json

    98 - Fitness selection, RL, BIOS rewards, MetaRegulator

  • GameConfig.json

    ~180+ - Game rules, sensors, physics, terrain

  • MTDPConfig.json

    6 - Multi-timescale decision projections

  • NFGFitnessMetaConfig.json

    28 - Internal meta fitness from telemetry

  • NeuroDynamicsConfig.json

    72 - STDP, predictive coding, sleep, working memory

  • OutputManagerConfig.json

    11 - Output post-processing, smoothing

  • RMGAConfig.json

    98 - 4 neurotransmitters, spiking, neuromodulation

  • RNNConfig.json

    68 - LSTM, GRU, Elman, recurrent memory

  • SensorsConfig.json

    96 - Distance sensors, NFG Vision, audio comm

  • Surroundings.json

    32 - Procedural grass, terrain, water

  • WorldPredictorConfig.json

    109 - World model, DreamerV3 what-if rollouts

45.2 Key Configuration Highlights

  • Config

    Notable Parameters

  • MMS

    8 modes, 8 state vars, 64 transition probs per neuron

  • Attention

    4 modes (Disabled/Layer/Temporal/TemporalAdditive)

  • Temporal Retention

    8 EMA timescales (16ms - 3min)

  • RMGA

    4 neurotransmitters with separate Tau/Blend/GateCoeff

  • WorldPredictor

    What-If rollouts, danger scoring, learned confidence

Provider & Interface Patterns

17 Features

46.1 Core Interfaces (12 interfaces)

  • Interface

    Methods - Purpose

  • INFGFitnessProvider

    10 - Fitness calculation

  • INFGRewardProvider

    12 - Reinforcement learning reward

  • INFGMTDPProvider

    10 - Multi-target dimensionality projection

  • INFGAttentionEngineProvider

    25+ - GPU attention dispatch

  • INFGTemporalRetentionProvider

    20+ - Multi-timescale memory

  • INFGWorldPredictorProvider

    30+ - World model, What-If rollouts

  • INFGMetaProgressProvider

    1 - Meta-learning progress

  • INFGMetaRegulatorProvider

    3 - UCB output selection

  • INFGModulatorGainsProvider

    1 - RMGA gain readback

  • INFGWhatIfStateProvider

    2 - GPU hidden state snapshots

  • NFGChildQueueTelemetryProvider

    1 - Child pool telemetry

  • INFGProviderStatus

    3 - Provider dashboard telemetry

46.2 Discovery Mechanisms

  • Mechanism

    Use Case

  • ScriptInterface

    TScriptInterface<IInterface> for Blueprint

  • ModularFeatures

    IModularFeatures for plugin backends

  • Singleton Registry

    FNFGNeuroDynamicsRegistry for global providers

Blueprint Integration

16 Features

47.1 Blueprint Exposure Statistics

  • Properties (UPROPERTY)

    1,372

  • Functions (UFUNCTION)

    541

  • BlueprintCallable Functions

    465

  • BlueprintPure Functions

    95

  • BlueprintImplementable Events

    15

  • Blueprint Delegates

    10

  • Blueprint Enums

    41

  • Blueprint Structs

    104

  • BlueprintSpawnable Components

    21

  • Blueprintable Classes

    13

47.2 Key Blueprint-Exposed Enums

  • Enum

    Values - Purpose

  • EOperationMode

    Training, Inference - Operation mode

  • ERNNMethod

    None, AverageRNN, LSTM, ElmanRNN - Recurrent type

  • ECrossoverMethod

    9 methods - Genetic crossover

  • EMutationPowerSystem

    4 systems - Mutation magnitude

  • EWorldPredictorTarget

    HiddenLayer, RawInputs, DiscreteTokens - Prediction target

Console Variables & Commands

19 Features

General

  • CPU Backend

    7 - nfg.cpu.optimize_dense

  • Attention Engine

    12 - nfg.attention.enable, nfg.attention.mode

  • CMA-ES

    20+ - nfg.cmaes.queue_slots_per_tick

  • Quantization

    6 - nfg.q.debug, nfg.q.force_scale

  • NFGVision

    7 - nfg.eye.max_cm, nfg.eye.gamma

  • Neuro Dynamics

    15+ - nfg.neuro.stdp.trace

  • MetaRegulator

    6 - nfg.meta.enabled, nfg.meta.ucb_override

  • RaycastGPU

    4 - raycast.gpu.spline.BezierSubdivisions

  • Command

    Purpose

  • NFG.EnableFileLogging

    Activate file log output

  • NFG.FlushLog

    Flush buffered logs

  • NFG.ToggleFreeCamera

    Toggle free-flight camera

  • NFG.SetWeather

    Change weather preset (0-5)

  • NFG.SetTimeOfDay

    Adjust simulation time

  • NFG.EnableCMAES / DisableCMAES

    Toggle CMA-ES

  • NFG.CMAESStatus

    Show CMA-ES status

  • NFG.RunAllTests

    Execute unit tests

  • NFG.BlockCrossoverStats

    Log crossover statistics

  • NFG.ToggleDecisionExplainer

    Toggle AI explanation widget

Logging & Telemetry Infrastructure

23 Features

49.1 Log Categories (25 main categories)

  • Category

    Plugin - Purpose

  • NFGLog

    Core - General-purpose logging

  • NFGWidget

    Core

    UI/Widget debugging

  • NFGNetwork

    Core - Neural network topology

  • NFGGpu

    Core - GPU operations

  • NFGTraining

    Core - Training progress

  • NFGIo

    Core

    File I/O

  • NFGWeather

    Core - Environmental systems

  • NFGCSGpu

    ComputeShaders - Compute shader dispatch

  • LogNFGAttentionEngine

    AttentionEngine - Attention mechanisms

  • LogNFGWorldPredictor

    WorldPredictor - Future state prediction

  • LogNFGTemporalRetention

    TemporalRetention - Multi-timescale memory

  • LogNFGNeuroDynamics

    NeuroDynamics - STDP, sleep replay

  • LogNFGModulator

    RMGA - Neuromodulation

  • LogNFGMetaRegulator

    MetaRegulator - Meta-learning

  • LogNFGLM

    NFGLM - Language model

  • LogNFGVision

    NFGVision - Visual perception (Retina)

  • LogRaycastGPU

    RaycastGPU - Ray casting

49.2 Telemetry Structures

  • Structure

    Fields - Purpose

  • FNFGTrainingTelemetry

    Generation, epoch, fitness stats, hall metrics - Training snapshot

  • FNFGChildQueueTelemetry

    Capacity, occupancy, ready/pending counts - Child pool monitoring

  • FKernelExecutionStats

    Time (avg/min/max), execution count, GFlops - Kernel profiling

  • FPerformanceSnapshot

    Frame time, GPU utilization, throughput - Performance snapshot

Serialization & Persistence

23 Features

50.1 Binary Serialization (FArchive)

  • Half-Precision Storage

    Weights as FFloat16, convert on serialize

    NFGModelComponent_Binary.cpp:26-27

  • Conditional Features

    Recurrent weights if Elman enabled

    NFGModelComponent_Binary.cpp:29-32

  • Batch Norm Params

    gamma/beta/mean/var if enabled

    NFGModelComponent_Binary.cpp:34-40

  • Attention Weights

    Guard flag bHasAttentionWeights

    NFGModelComponent_Binary.cpp:85-117

  • EvoFeatureSnapshot

    Optional evolved features

    NFGModelComponent_Binary.cpp:119-142

  • Backward Compatibility

    Safe EOF checks

    NFGModelComponent_Binary.cpp:98-116

50.2 JSON Serialization

  • Quantization Payload

    Base64 packed weights

    NFGModelComponent_Serialization.cpp:36-157

  • NeuroDynamics Params

    STDP rates, working memory

    NFGModelComponent_Serialization.cpp:165-186

  • TSReLU Parameters

    Per-neuron alpha/deadzone/hysteresis

    NFGModelComponent_Serialization.cpp:236-243

  • MMS State

    mode_gains, transition_matrix, internal_state

    NFGModelComponent_Serialization.cpp:245-270

  • Working Memory Snapshot

    Latches, hold frames, active flags

    NFGModelComponent_Serialization.cpp:272-308

  • Per-Layer RNN

    Elman/LSTM params per layer

    NFGModelComponent_Serialization.cpp:310-391

  • AdditionalData Array

    Polymorphic nested JSON

    NFGModelComponent_Serialization.cpp:393-412

50.3 Version Handling

  • Pattern

    Description

  • Version String

    0.7200 = v0.7 + 200 commits

  • Optional Field Guards

    bHasAttentionWeights, bHasEvoSnapshot

  • Legacy Enum Mapping

    CrossoverMethod 8 → Uniform

  • Deprecated Placeholders

    Dummy values preserve archive structure

50.4 File Formats

  • Format

    Extension - Use Case

  • Binary

    .nfg, .dat - Fast disk storage

  • JSON

    .json - Export, debugging

  • Agent Binary

    .bin - Polymorphic data

  • CSV

    .csv - Fitness tracking

Build System & Module Dependencies

27 Features

51.1 Plugin Modules (18 UE5 modules)

  • Module

    Type - Dependencies - Purpose

  • NeuroFluxGenesis

    Runtime - Core, Engine, RenderCore, RHI - Main neural architecture

  • NFGBackendComputeShaders

    Runtime - RenderCore, RHI, NeuroFluxGenesis - GPU compute backend

  • NFGAttentionEngine

    Runtime - NeuroFluxGenesis, RenderCore - Attention mechanisms

  • NFGWorldPredictor

    Runtime - NeuroFluxGenesis, RenderCore - Future state prediction

  • NFGTemporalRetention

    Runtime - NeuroFluxGenesis, RenderCore - Multi-timescale memory

  • NFGNeuroDynamics

    Runtime - NeuroFluxGenesis - STDP + Sleep replay

  • NFGMultiModulatorRMGA

    Runtime - NeuroFluxGenesis - Neuromodulation

  • NFGMetaRegulator

    Runtime - NeuroFluxGenesis - Meta-learning UCB

  • NFGMTDP

    Runtime - NeuroFluxGenesis, RenderCore - Delay prediction

  • NFGLM

    Runtime - NeuroFluxGenesis - Language model

  • NFGAudioComm

    Runtime - NeuroFluxGenesis, AudioMixer - Agent communication

  • NFGVision

    Runtime - RenderCore, RHI - Retina vision encoder

  • RaycastGPU

    Runtime - RenderCore, RHI - GPU raycasting

  • NFGHallOfFittestPersistence

    Runtime - NeuroFluxGenesis - Checkpoint system

  • NFGQuantization

    Runtime - NeuroFluxGenesis - Bit-packing codec

  • TorqueStormPhysics

    Runtime - Core, Engine - Vehicle physics

  • TrackGenerator

    Runtime - Core, Engine - Procedural tracks

  • WeatherSystem

    Runtime - Core, Engine - Environmental sim

51.3 Build.cs Common Patterns

  • Pattern

    Description - Source

  • PublicDependencyModuleNames

    API-exposed dependencies - All Build.cs files

  • PrivateDependencyModuleNames

    Internal-only dependencies - All Build.cs files

  • PCHUsage.UseExplicitOrSharedPCHs

    Precompiled header strategy - Standard across modules

  • bEnableUndefinedIdentifierWarnings

    Strict compilation - NFGBackendComputeShaders

  • OptimizeCode = ModuleRules.CodeOptimization.InShippingBuildsOnly

    Debug optimization - Core modules

51.4 Shader Directory Registration

  • Plugin

    Shader Path - Source

  • NFGVision

    /NFGVision/Private/

    NFGVision.Build.cs

FeedForward Shader Implementation

27 Features

52.1 Main Entry Points (`FeedForwardBatch2D.usf`)

  • Kernel

    Thread Config - Description - Source

  • MainCS

    [32,1,1] - Standard forward pass

    FeedForwardBatch2D.usf:1-50

  • MainCS_MMS

    [32,1,1] - Multi-Mode State neurons

    FeedForwardBatch2D.usf:52-180

  • MainCS_Elman

    [32,1,1] - Recurrent network pass

    FeedForwardBatch2D.usf:182-260

  • MainCS_LSTM

    [32,1,1] - LSTM gate computation

    FeedForwardBatch2D.usf:262-420

52.2 Activation Functions (GPU)

  • Function

    Formula - Source

52.3 Weight Access Patterns

  • Pattern

    Description - Source

  • Coalesced Load

    Thread tid reads weight row tid

    FeedForwardBatch2D.usf:85-95

  • Shared Memory Tiling

    32×32 weight tiles in groupshared

    FeedForwardBatch2D.usf:97-130

  • Wave Reduction

    WaveActiveSum() for partial sums

    FeedForwardBatch2D.usf:132-145

  • Bias Addition

    Vector bias after activation

    FeedForwardBatch2D.usf:147-155

52.4 MMS GPU Implementation

  • Mode Selection

    argmax over transition probabilities

    FeedForwardBatch2D.usf:160-175

  • State Update

    8 internal state variables per neuron

    FeedForwardBatch2D.usf:177-195

  • Gain Modulation

    Per-mode gain multipliers

    FeedForwardBatch2D.usf:197-210

  • Transition Matrix

    64-element (8×8) probability table

    FeedForwardBatch2D.usf:212-230

  • State Decay

    Exponential decay per neuron

    FeedForwardBatch2D.usf:232-245

52.5 LSTM Gate Computation (GPU)

  • Gate

    Formula - Source

  • Input (i)

    σ(Wi·[h,x] + bi)

    FeedForwardBatch2D.usf:280-295

  • Forget (f)

    σ(Wf·[h,x] + bf)

    FeedForwardBatch2D.usf:297-312

  • Cell (g)

    tanh(Wg·[h,x] + bg)

    FeedForwardBatch2D.usf:314-329

  • Output (o)

    σ(Wo·[h,x] + bo)

    FeedForwardBatch2D.usf:331-346

  • Cell State

    f * c_prev + i * g

    FeedForwardBatch2D.usf:348-360

  • Hidden State

    o * tanh(c)

    FeedForwardBatch2D.usf:362-375

52.6 Batch Processing

  • Batch Size

    Up to 1024 agents

    FeedForwardBatch2D.usf:25

  • Agent Stride

    MaxNeurons * sizeof(float)

    FeedForwardBatch2D.usf:30

  • Layer Indexing

    Prefix sum of layer sizes

    FeedForwardBatch2D.usf:35-45

  • Async Dispatch

    RDG graph scheduling

    ComputeShadersComponent.cpp:2100

CMA-ES GPU Kernels

36 Features

53.1 Kernel Overview (35 kernels in `CMAES.usf`)

  • Kernel

    Purpose - Lines

  • InitializePopulation

    Zero-initialize mean/covariance - 1-45

  • SamplePopulation

    Box-Muller Gaussian sampling - 47-120

  • ApplyBounds

    Clamp to parameter bounds - 122-165

  • ComputeFitness

    Fitness function evaluation - 167-230

  • SortByFitness

    Bitonic sort on GPU - 232-340

  • SelectElites

    Top-μ selection - 342-390

  • UpdateMean

    Weighted recombination - 392-450

  • UpdateEvolutionPath_ps

    Conjugate evolution path - 452-520

  • UpdateEvolutionPath_pc

    Cumulation path - 522-590

  • UpdateCovariance

    Rank-1 and rank-μ update - 592-750

  • DecomposeCovariance

    Eigendecomposition (iterative) - 752-920

  • ComputeBD

    B×D matrix for sampling - 922-980

  • AdaptStepSize

    σ adaptation via path length - 982-1050

  • CheckTermination

    Convergence criteria - 1052-1120

  • RestartStrategy

    IPOP/BIPOP restart - 1122-1200

53.2 Covariance Matrix Operations

  • Operation

    Complexity - Description - Source

  • Rank-1 Update

    O(n²) - C += c1 * pc * pc^T

    CMAES.usf:620-680

  • Rank-μ Update

    O(μn²) - C += cμ * Σ wi * yi * yi^T

    CMAES.usf:682-750

  • Eigendecomposition

    O(n³) - Jacobi rotation method

    CMAES.usf:752-920

  • Square Root

    O(n²) - C^(1/2) = B * D

    CMAES.usf:922-980

53.3 Evolution Path Computation

  • Path

    Formula - Purpose - Source

  • p_σ

    (1-cσ)·pσ + √(cσ(2-cσ)μeff)·C^(-1/2)·(m-m_old)/σ - Step size control

    CMAES.usf:452-520

  • p_c

    (1-cc)·pc + hσ·√(cc(2-cc)μeff)·(m-m_old)/σ - Covariance adaptation

    CMAES.usf:522-590

53.4 Low-Rank Approximation

  • Rank-k Storage

    Only top-k eigenvectors stored

    CMAES.usf:1202-1280

  • Incremental Update

    Streaming eigenvector update

    CMAES.usf:1282-1350

  • Memory Savings

    O(nk) vs O(n²) for full

    CMAES.usf:1352-1400

  • Accuracy Trade-off

    Configurable k parameter

    Config: EvolutionConfig.json

53.5 Step Size Adaptation

  • cσ

    (μeff + 2) / (n + μeff + 5) - Learning rate for p_σ

    CMAES.usf:990-1000

  • dσ

    1 + 2·max(0, √((μeff-1)/(n+1))-1) + cσ - Damping factor

    CMAES.usf:1002-1015

  • Expected

    N(0,I) - - √n · (1 - 1/(4n) + 1/(21n²)) - Chi distribution

    CMAES.usf:1017-1030

53.6 Restart Mechanisms

  • Strategy

    Trigger - Action - Source

  • IPOP

    Stagnation - Double population size

    CMAES.usf:1122-1150

  • BIPOP

    IPOP limit - Alternate large/small pop

    CMAES.usf:1152-1180

  • TolX

    σ * max(diag(C)) < 1e-12 - Convergence restart

    CMAES.usf:1052-1070

  • TolFun

    Fitness plateau - Historical fitness check

    CMAES.usf:1072-1095

Crossover & Mutation Shaders

2 Features

54.4 Mutation Operators (`Mutation.usf`)

  • Operator

    Description - Source

54.5 Mutation Rate Scheduling

  • Schedule

    Formula - Source

Quantization Codec Details

11 Features

55.1 Bit-Width Support

  • Bits

    Values - Pack Ratio - Source

55.2 Pack Functions

  • Function

    Signature - Source

55.3 Unpack Functions

  • Function

    Signature - Source

55.5 STDP Scale Evolution

  • Spike-Timing Modulation

    Scale adjusts based on spike correlation

    NFGNeuroDynamicsProvider.cpp:420-480

  • LTP Scale Increase

    Strong correlation → larger scale

    NFGNeuroDynamicsProvider.cpp:482-510

  • LTD Scale Decrease

    Weak correlation → smaller scale

    NFGNeuroDynamicsProvider.cpp:512-540

  • Bounds Enforcement

    Min/max scale limits

    NFGNeuroDynamicsProvider.cpp:542-560

55.6 Memory Bandwidth Savings

  • Scenario

    FP32 - INT8 - INT4 - INT2

  • 1M weights

    4 MB - 1 MB - 512 KB - 256 KB

  • 10M weights

    40 MB - 10 MB - 5 MB - 2.5 MB

  • Bandwidth

    1× - 4× - 8× - 16×

Widget & UI System

5 Features

56.1 Slate Widgets (31 custom widgets)

  • Widget

    Purpose - Source

56.2 Widget Features

  • Real-time Update

    60 FPS refresh rate - All widget Tick() methods

56.3 Network Visualizer Modes

  • Mode

    Description - Source

56.4 Layout Algorithms

  • Algorithm

    Description - Source

56.5 Performance Optimizations

  • Optimization

    Description - Source

Remote IO & Authentication

8 Features

57.2 Authentication Flow

  • Step

    Description - Source

  • 1. License Check

    Validate license key

    AuthentificationManager.cpp:1-80

  • 2. JWT Request

    Exchange key for token

    AuthentificationManager.cpp:82-150

  • 3. Token Storage

    Secure token cache

    AuthentificationManager.cpp:152-200

  • 4. Token Refresh

    Auto-refresh before expiry

    AuthentificationManager.cpp:202-260

  • 5. Request Signing

    Bearer token in headers

    AuthentificationManager.cpp:262-310

57.3 Model Upload/Download

  • Operation

    Endpoint - Source

57.5 Progress Callbacks

  • Callback

    Parameters - Source

Input Processing Pipeline

10 Features

58.1 7-Stage Pipeline

  • Stage

    Description - Source

  • 1. Raw Input

    Sensor data collection

    NFGInputProcessor.cpp:1-80

58.2 Input Sources (15 types)

  • Source

    Neurons - Description - Source

  • Vision

    2048 - NFGVision Retina features

    NFGVisionSceneComponent.cpp

58.3 Normalization Methods

  • Method

    Formula - Use Case - Source

  • MinMax

    (x - min) / (max - min) - Bounded inputs

    NFGInputProcessor.cpp:85-100

  • ZScore

    (x - μ) / σ - Unbounded, Gaussian

    NFGInputProcessor.cpp:102-120

58.5 Temporal Integration (EMA)

  • Timescale

    Alpha - Purpose - Source

58.6 Feature Extraction

  • Feature

    Inputs - Output - Source

58.7 GPU Buffer Management

  • Buffer

    Size - Update Rate - Source

CPU Crossover Implementation

47 Features

59.1 Core Architecture (4,165 lines)

  • Neuron-Coherent Crossover

    Entire neuron units transferred, not individual weights

    CPUComponent_Crossover.cpp:180-350

  • MMS State Crossover

    8 internal states + 64 transition probs inherited together

    CPUComponent_Crossover.cpp:352-520

  • LSTM Gate Crossover

    All 4 gates (i/f/o/c) from same parent for coherence

    CPUComponent_Crossover.cpp:522-750

  • GRU Gate Crossover

    Reset + Update gates paired inheritance

    CPUComponent_Crossover.cpp:752-920

  • Elman Context Crossover

    Context weights + decay rates together

    CPUComponent_Crossover.cpp:922-1050

59.2 Crossover Methods (8 strategies)

  • Uniform

    Random parent selection per neuron

    CPUComponent_Crossover.cpp:1100-1180

  • SinglePoint

    Split at random neuron index

    CPUComponent_Crossover.cpp:1182-1260

  • TwoPoint

    Two split points, middle from P2

    CPUComponent_Crossover.cpp:1262-1350

  • LayerWise

    Entire layers from single parent

    CPUComponent_Crossover.cpp:1352-1440

  • Arithmetic

    Weighted blend (α×P1 + (1-α)×P2)

    CPUComponent_Crossover.cpp:1442-1530

  • BLX-α

    Extended blend with exploration

    CPUComponent_Crossover.cpp:1532-1620

  • SBX

    Simulated Binary Crossover

    CPUComponent_Crossover.cpp:1622-1750

  • Tournament

    Per-neuron fitness-based selection

    CPUComponent_Crossover.cpp:1752-1850

59.3 TS-ReLU Parameter Inheritance

  • Parameter

    Crossover Rule - Source

  • LeftCap/RightCap

    Paired inheritance

    CPUComponent_Crossover.cpp:1900-1950

  • LeftSlope/RightSlope

    From same parent

    CPUComponent_Crossover.cpp:1952-2000

  • Hysteresis

    Blended (arithmetic)

    CPUComponent_Crossover.cpp:2002-2040

  • Deadzone

    Per-neuron coherent

    CPUComponent_Crossover.cpp:2042-2100

  • SigmaScale

    Layer-wise inheritance

    CPUComponent_Crossover.cpp:2102-2150

59.4 MMS Crossover Details

  • Component

    Count - Inheritance Rule - Source

  • MMSMode

    1 per neuron - From parent with mode

    CPUComponent_Crossover.cpp:2200-2250

  • MMSDecay

    1 per neuron - Single value

    CPUComponent_Crossover.cpp:2552-2600

  • MMSEta

    1 per neuron - From parent

    CPUComponent_Crossover.cpp:2602-2650

  • MMSStateInfluence

    1 per neuron - From parent

    CPUComponent_Crossover.cpp:2652-2700

  • MMSInputInfluence

    1 per neuron - From parent

    CPUComponent_Crossover.cpp:2702-2750

59.5 Quantization-Aware Crossover

  • BitWidth Inheritance

    Per-neuron bit width preserved

    CPUComponent_Crossover.cpp:2800-2860

  • Scale Factor Transfer

    Quantization scale from parent

    CPUComponent_Crossover.cpp:2862-2920

  • Zero-Point Inheritance

    Asymmetric quant zero-point

    CPUComponent_Crossover.cpp:2922-2980

  • Dequant-Crossover-Requant

    Full precision during crossover

    CPUComponent_Crossover.cpp:2982-3050

59.6 LSTM/GRU Coherence Rules

  • Architecture

    Coherence Constraint - Source

  • LSTM

    All 4 gates + cell state from same parent

    CPUComponent_Crossover.cpp:3100-3250

  • GRU

    Reset + Update paired, candidate separate

    CPUComponent_Crossover.cpp:3252-3400

  • Elman

    Hidden + Context from same parent

    CPUComponent_Crossover.cpp:3402-3500

  • Peephole

    Cell-to-gate weights with cell

    CPUComponent_Crossover.cpp:3502-3600

59.7 Crossover Utilities

  • Utility

    Description - Source

  • ParentSelector

    Fitness-proportionate, tournament, rank

    NFGCrossoverUtils.cpp:50-180

  • CrossoverMask

    Binary mask for uniform crossover

    NFGCrossoverUtils.cpp:182-260

  • BlendCoefficients

    α generation for arithmetic/BLX

    NFGCrossoverUtils.cpp:262-340

  • NeuronBoundaryDetector

    Find coherent unit boundaries

    NFGCrossoverUtils.cpp:342-450

  • InheritanceValidator

    Verify coherence constraints

    NFGCrossoverUtils.cpp:452-550

59.8 Child Generation Pipeline

  • Stage

    Description - Source

  • 1. Parent Selection

    Choose 2 parents from population

    CPUComponent_Crossover.cpp:3700-3780

  • 2. Crossover Point(s)

    Determine split locations

    CPUComponent_Crossover.cpp:3782-3850

  • 3. Gene Transfer

    Copy neuron units

    CPUComponent_Crossover.cpp:3852-3950

  • 4. Constraint Check

    Verify coherence

    CPUComponent_Crossover.cpp:3952-4020

  • 5. Mutation

    Optional post-crossover mutation

    CPUComponent_Crossover.cpp:4022-4100

  • 6. Fitness Init

    Reset fitness for new child

    CPUComponent_Crossover.cpp:4102-4165

RDG Shader Dispatch Pipeline

45 Features

60.1 Core RDG Pattern

  • FRDGBuilder

    Render Dependency Graph builder

    ComputeShadersComponent.cpp:1200-1350

  • Pass Registration

    AddPass<FComputeShaderRDG>() pattern

    ComputeShadersComponent.cpp:1352-1500

  • Resource Transitions

    UAV→SRV, SRV→UAV automatic

    ComputeShadersComponent.cpp:1502-1650

  • Fence Synchronization

    GPU fence for async completion

    ComputeShadersComponent.cpp:1652-1800

60.2 Buffer Allocation Patterns

  • Pattern

    Description - Source

  • CreateBuffer

    FRDGBufferDesc with size/stride

    ComputeShadersComponent.cpp:1850-1950

  • CreateUAV

    Unordered access view creation

    ComputeShadersComponent.cpp:1952-2050

  • CreateSRV

    Shader resource view creation

    ComputeShadersComponent.cpp:2052-2150

  • External Buffers

    RegisterExternalBuffer() for persistent

    ComputeShadersComponent.cpp:2152-2280

60.3 Pass Types (12 categories)

  • Pass Type

    Description - Source

  • FeedForward

    Neural network forward pass

    ComputeShadersComponent.cpp:2300-2500

  • STDP

    Spike-timing plasticity update

    ComputeShadersComponent.cpp:2502-2650

  • CMA-ES

    Evolution strategy kernels

    ComputeShadersComponent.cpp:2652-2900

  • Crossover

    Genetic crossover on GPU

    ComputeShadersComponent.cpp:2902-3100

  • Mutation

    Weight mutation kernels

    ComputeShadersComponent.cpp:3102-3250

  • Attention

    Multi-head attention

    ComputeShadersComponent.cpp:3252-3450

  • WorldPredictor

    Predictive model forward

    ComputeShadersComponent.cpp:3452-3650

  • TemporalRetention

    Memory EMA update

    ComputeShadersComponent.cpp:3652-3800

  • Quantization

    Quant/Dequant kernels

    ComputeShadersComponent.cpp:3802-3950

  • Telemetry

    Statistics readback

    ComputeShadersComponent.cpp:3952-4100

  • VehiclePhysics

    Physics simulation

    ComputeShadersComponent.cpp:4102-4350

  • Raycast

    GPU raycasting

    ComputeShadersComponent.cpp:4352-4550

60.4 Async Compute Integration

  • AsyncCompute Queue

    Separate compute queue

    ComputeShadersComponent.cpp:4600-4700

  • Overlap with Graphics

    Parallel execution

    ComputeShadersComponent.cpp:4702-4800

  • Fence Wait

    Synchronization points

    ComputeShadersComponent.cpp:4802-4900

  • Priority Hints

    ERDGAsyncCompute flags

    ComputeShadersComponent.cpp:4902-5000

60.5 Resource Lifetime Management

  • Pattern

    Description - Source

  • Transient

    Single-frame buffers

    ComputeShadersComponent.cpp:5050-5150

  • Persistent

    Multi-frame buffers

    ComputeShadersComponent.cpp:5152-5280

  • Pooled

    Reused allocations

    ComputeShadersComponent.cpp:5282-5400

  • External

    User-managed buffers

    ComputeShadersComponent.cpp:5402-5500

60.6 Dispatch Patterns

  • Pattern

    Description - Source

  • Direct Dispatch

    DispatchComputeShader()

    ComputeShadersComponent.cpp:5550-5650

  • Indirect Dispatch

    GPU-driven dispatch count

    ComputeShadersComponent.cpp:5652-5780

  • Multi-Dispatch

    Sequential kernel chain

    ComputeShadersComponent.cpp:5782-5900

  • Batched Dispatch

    Multiple passes per frame

    ComputeShadersComponent.cpp:5902-6050

60.7 Barrier Management

  • Barrier Type

    Description - Source

  • UAV Barrier

    Write-after-write

    ComputeShadersComponent.cpp:6100-6180

  • SRV Barrier

    Read-after-write

    ComputeShadersComponent.cpp:6182-6260

  • Aliasing Barrier

    Buffer reuse

    ComputeShadersComponent.cpp:6262-6340

  • Global Barrier

    Full pipeline flush

    ComputeShadersComponent.cpp:6342-6420

60.8 Error Handling & Validation

  • RDG Validation

    Debug mode checks

    ComputeShadersComponent.cpp:6470-6550

  • Resource Tracking

    Leak detection

    ComputeShadersComponent.cpp:6552-6630

  • Pass Ordering

    Dependency validation

    ComputeShadersComponent.cpp:6632-6710

  • Size Verification

    Buffer bounds checking

    ComputeShadersComponent.cpp:6712-6800

NFGLM Tokenizer & Decoder

3 Features

61.2 Encoding Pipeline

  • Stage

    Description - Source

61.3 Decoding Pipeline

  • Stage

    Description - Source

61.6 Sequence Runner (7 Stages)

  • Stage

    Description - Source

AudioComm Synthesis Internals

37 Features

62.1 Synthesis Parameters (20 total)

  • F0

    50-500 Hz - Fundamental frequency

    NFGAudioCommSynthComponent.cpp:45-60

  • Voicing

    0-1 - Voiced/unvoiced ratio

    NFGAudioCommSynthComponent.cpp:62-75

  • Breathiness

    0-1 - Noise injection

    NFGAudioCommSynthComponent.cpp:77-90

  • Jitter

    0-0.1 - F0 random variation

    NFGAudioCommSynthComponent.cpp:92-105

  • Shimmer

    0-0.1 - Amplitude variation

    NFGAudioCommSynthComponent.cpp:107-120

  • Tremolo Rate

    0-10 Hz - AM modulation rate

    NFGAudioCommSynthComponent.cpp:122-135

  • Tremolo Depth

    0-1 - AM modulation depth

    NFGAudioCommSynthComponent.cpp:137-150

  • Vibrato Rate

    0-10 Hz - FM modulation rate

    NFGAudioCommSynthComponent.cpp:152-165

  • Vibrato Depth

    0-50 Hz - FM modulation depth

    NFGAudioCommSynthComponent.cpp:167-180

  • F1 Freq

    200-1000 Hz - First formant

    NFGAudioCommSynthComponent.cpp:182-195

  • F1 BW

    50-200 Hz - First formant bandwidth

    NFGAudioCommSynthComponent.cpp:197-210

  • F2 Freq

    700-2500 Hz - Second formant

    NFGAudioCommSynthComponent.cpp:212-225

  • F2 BW

    70-300 Hz - Second formant bandwidth

    NFGAudioCommSynthComponent.cpp:227-240

  • F3 Freq

    2000-3500 Hz - Third formant

    NFGAudioCommSynthComponent.cpp:242-255

  • F3 BW

    100-400 Hz - Third formant bandwidth

    NFGAudioCommSynthComponent.cpp:257-270

  • Nasality

    0-1 - Nasal cavity coupling

    NFGAudioCommSynthComponent.cpp:272-285

  • Tenseness

    0-1 - Glottal tension

    NFGAudioCommSynthComponent.cpp:287-300

  • Attack

    0-0.1 s - Onset time

    NFGAudioCommSynthComponent.cpp:302-315

  • Release

    0-0.2 s - Offset time

    NFGAudioCommSynthComponent.cpp:317-330

  • Volume

    0-1 - Output amplitude

    NFGAudioCommSynthComponent.cpp:332-345

62.2 Glottal Source Model (Rosenberg)

  • Open Phase

    Glottal pulse rising edge

    NFGAudioCommSynthComponent.cpp:380-420

  • Return Phase

    Glottal pulse falling edge

    NFGAudioCommSynthComponent.cpp:422-460

  • Closed Phase

    Inter-pulse silence

    NFGAudioCommSynthComponent.cpp:462-500

  • Aspiration Noise

    Breath noise during open

    NFGAudioCommSynthComponent.cpp:502-540

  • OQ Ratio

    Open quotient (0.3-0.7)

    NFGAudioCommSynthComponent.cpp:542-570

62.3 Formant Filter Bank

  • Filter

    Order - Type - Source

62.4 Phoneme System (9 phonemes)

  • Phoneme

    F1 - F2 - F3 - Source

62.6 Speaker/Listener System

  • SpeakerComponent

    Per-agent sound emitter

    NFGAudioCommSpeakerComponent.cpp:30-100

  • ListenerComponent

    Per-agent sound receiver

    NFGAudioCommSpeakerComponent.cpp:102-180

  • Distance Attenuation

    1/r² falloff

    NFGAudioCommSpeakerComponent.cpp:182-240

  • Directional Gain

    Cone-based attenuation

    NFGAudioCommSpeakerComponent.cpp:242-300

  • Occlusion

    Line-of-sight blocking

    NFGAudioCommSpeakerComponent.cpp:302-360

  • Multi-Agent Mixing

    Sum of all audible speakers

    NFGAudioCommSpeakerComponent.cpp:362-420

62.7 World Subsystem Integration

  • Agent Registry

    Track all AudioComm agents

    NFGAudioCommWorldSubsystem.cpp:40-100

  • Spatial Hashing

    Fast neighbor queries

    NFGAudioCommWorldSubsystem.cpp:102-180

  • Batch Processing

    Update all agents per frame

    NFGAudioCommWorldSubsystem.cpp:182-250

  • Audio Buffer Pool

    Reusable sample buffers

    NFGAudioCommWorldSubsystem.cpp:252-320

Weather State Machine Internals

41 Features

64.1 Weather States (8 states)

  • State

    Description - Transitions To - Source

  • Clear

    Sunny, no clouds - Cloudy, Overcast

    NFGWeatherStateMachine.cpp:40-70

  • Cloudy

    Partial clouds - Clear, Overcast, LightRain

    NFGWeatherStateMachine.cpp:72-105

  • Overcast

    Full cloud cover - Cloudy, LightRain, Snow

    NFGWeatherStateMachine.cpp:107-140

  • LightRain

    Drizzle - Cloudy, HeavyRain, Overcast

    NFGWeatherStateMachine.cpp:142-175

  • HeavyRain

    Downpour - LightRain, Thunderstorm

    NFGWeatherStateMachine.cpp:177-210

  • Thunderstorm

    Lightning + Heavy Rain - HeavyRain

    NFGWeatherStateMachine.cpp:212-250

  • Snow

    Light snowfall - Overcast, Blizzard

    NFGWeatherStateMachine.cpp:252-285

  • Blizzard

    Heavy snow + wind - Snow

    NFGWeatherStateMachine.cpp:287-320

64.2 Transition Probabilities

  • From

    To - Base Probability - Source

  • Clear→Cloudy

    0.15 per hour - Temperature-modulated

    NFGWeatherStateMachine.cpp:350-380

  • Cloudy→LightRain

    0.25 per hour - Humidity-dependent

    NFGWeatherStateMachine.cpp:382-410

  • LightRain→HeavyRain

    0.30 per hour - Pressure-based

    NFGWeatherStateMachine.cpp:412-440

  • HeavyRain→Thunder

    0.20 per hour - Instability index

    NFGWeatherStateMachine.cpp:442-470

64.3 Weather Front System (4 types)

  • Front Type

    Description - Effects - Source

  • Cold Front

    Cold air advancing - Rapid temp drop, storms

    NFGWeatherFrontManager.cpp:50-120

  • Warm Front

    Warm air advancing - Gradual warm, steady rain

    NFGWeatherFrontManager.cpp:122-190

  • Occluded Front

    Cold overtaking warm - Complex precipitation

    NFGWeatherFrontManager.cpp:192-260

  • Stationary Front

    No movement - Prolonged conditions

    NFGWeatherFrontManager.cpp:262-320

64.4 Atmospheric Physics

  • Parameter

    Calculation - Source

  • Lapse Rate

    -6.5°C per 1000m

    NFGIntegratedWeatherSystem.cpp:80-110

  • Dew Point

    Magnus formula

    NFGIntegratedWeatherSystem.cpp:112-150

  • CAPE

    Convective Available Potential Energy

    NFGIntegratedWeatherSystem.cpp:152-200

  • Lifted Index

    Atmospheric stability

    NFGIntegratedWeatherSystem.cpp:202-250

  • K-Index

    Thunderstorm potential

    NFGIntegratedWeatherSystem.cpp:252-300

64.5 Precipitation Model

  • Type

    Intensity Range - Accumulation - Source

  • Drizzle

    0.1-0.5 mm/hr - Track wetness

    NFGWeatherSimulationComponent.cpp:100-140

  • Light Rain

    0.5-2.5 mm/hr - Puddle formation

    NFGWeatherSimulationComponent.cpp:142-180

  • Moderate Rain

    2.5-7.5 mm/hr - Standing water

    NFGWeatherSimulationComponent.cpp:182-220

  • Heavy Rain

    7.5-50 mm/hr - Flooding

    NFGWeatherSimulationComponent.cpp:222-260

  • Snow

    1-5 cm/hr - Snow accumulation

    NFGWeatherSimulationComponent.cpp:262-300

64.6 Visibility Calculation

  • Factor

    Impact - Formula - Source

  • Rain Intensity

    Reduces visibility - vis = 10000 / (1 + rain×5)

    NFGWeatherSimulationComponent.cpp:350-380

  • Fog Density

    Exponential decay - vis = -log(0.02) / density

    NFGWeatherSimulationComponent.cpp:382-410

  • Snow Rate

    Scatter-based - vis = 1000 / (snow×2)

    NFGWeatherSimulationComponent.cpp:412-440

  • Dust/Sand

    Particle count - vis = 5000 / (1 + dust×10)

    NFGWeatherSimulationComponent.cpp:442-470

64.7 Wind Model

  • Base Wind

    Prevailing direction/speed

    NFGWeatherSimulationComponent.cpp:500-540

  • Gusts

    Random intensity spikes

    NFGWeatherSimulationComponent.cpp:542-580

64.8 Time-of-Day Integration

  • Diurnal Cycle

    Day/night temperature swing

    NFGIntegratedWeatherSystem.cpp:400-450

  • Solar Heating

    Afternoon convection

    NFGIntegratedWeatherSystem.cpp:452-500

  • Nocturnal Cooling

    Radiation fog potential

    NFGIntegratedWeatherSystem.cpp:502-550

GPU Structured Buffer Layouts

9 Features

64.1 Core Neural Network Buffers

  • Buffer

    Type - Size - Description - Source

64.2 MMS (Multi-Mode State) Buffers

  • Buffer

    Type - Size - Description - Source

64.3 LSTM/GRU Gate Buffers

  • Buffer

    Type - Size - Description - Source

64.4 CMA-ES Evolution Buffers

  • Buffer

    Type - Size - Description - Source

64.5 Child Pool Buffers

  • Buffer

    Type - Size - Description - Source

64.6 RMGA Neuromodulation Buffers

  • Buffer

    Type - Size - Description - Source

64.7 World Predictor Buffers

  • Buffer

    Type - Size - Description - Source

64.8 Quantization Buffers

  • Buffer

    Type - Size - Description - Source

64.9 Telemetry & Statistics Buffers

  • Buffer

    Type - Size - Description - Source

Test Coverage Analysis

253 Features

65.1 Test File Distribution (148 files)

  • Plugin

    Test Files - Test Cases - Source Path

  • NFGBackendComputeShaders

    32 - 95+

    Plugins/NFGBackendComputeShaders/.../Tests/

  • NeuroFluxGenesis Core

    28 - 85+

    Plugins/NeuroFluxGenesis/.../Tests/

  • NFGAttentionEngine

    12 - 35+

    Plugins/NFGAttentionEngine/.../Tests/

  • NFGWorldPredictor

    11 - 32+

    Plugins/NFGWorldPredictor/.../Tests/

  • NFGMultiModulatorRMGA

    9 - 28+

    Plugins/NFGMultiModulatorRMGA/.../Tests/

  • NFGTemporalRetention

    8 - 22+

    Plugins/NFGTemporalRetention/.../Tests/

  • NFGNeuroDynamics

    7 - 18+

    Plugins/NFGNeuroDynamics/.../Tests/

  • NFGLM

    14 - 40+

    Plugins/NFGLM/.../Tests/

  • NFGMetaRegulator

    8 - 20+

    Plugins/NFGMetaRegulator/.../Tests/

  • NFGMTDP

    6 - 15+

    Plugins/NFGMTDP/.../Tests/

  • NFGVision

    5 - 12+

    Plugins/NFGVision/.../Tests/

  • NFGAudioComm

    8 - 18+

    Plugins/NFGAudioComm/.../Tests/

65.2 Test Categories

  • Category

    Count - Description - Examples

  • Guard Tests

    45+ - Invariant/precondition checks

    *GuardTests.cpp

  • GPU Kernel Tests

    35+ - Shader correctness

    ComputeShader*Tests.cpp

  • Integration Tests

    28+ - Multi-component flows

    *E2E_Spec.cpp

  • Unit Tests

    65+ - Single function/class

    *Tests.cpp

  • Boundary Tests

    20+ - Edge cases

    *BoundsTests.cpp

  • Performance Tests

    8+ - Timing/throughput

    *PerformanceTests.cpp

65.3 Notable Test Suites

  • Test Suite

    File - Description - Source

  • CMA-ES Covariance 2D

    ComputeShaderCMAESCov2DTests.cpp - Full matrix tests - Lines 1-350

  • CMA-ES Error Flags

    ComputeShaderCMAESErrorFlagTests.cpp - Divergence detection - Lines 1-280

  • STDP Order Tests

    ComputeShaderStdpAttentionOrderTests.cpp - Spike timing - Lines 1-320

  • Crossover Method Tests

    CPUComponentCrossoverMethodTests.cpp - All 8 methods - Lines 1-450

  • HoF Selection Tests

    CMAESHoFSelectionTests.cpp - Elite persistence - Lines 1-280

  • Prediction Readback

    ComputeShaderPredictionReadbackToggleTests.cpp - GPU→CPU transfer - Lines 1-220

65.4 Test Infrastructure

  • Test Helpers

    Mock providers, fixtures

    Helpers/*.h

  • TestWorldPredictorProvider

    Mock predictor

    TestWorldPredictorProvider.h

  • TestFitnessComponent

    Mock fitness

    TestFitnessComponent.h

  • TestMTDPProvider

    Mock MTDP

    TestMTDPProvider.h

  • ComputeShaderTestHelpers

    GPU test utilities

    ComputeShaderTestHelpers.h

65.5 Guard Test Patterns

  • Pattern

    Description - Example

  • Invariant Guards

    Check class invariants - TestThat(X, Invariant())

  • Precondition Guards

    Validate inputs - TestThat(Input, IsValid())

  • Postcondition Guards

    Verify outputs - TestThat(Output, InRange())

  • State Guards

    Check state machine - TestThat(State, IsConsistent())

  • Resource Guards

    Verify allocations - TestThat(Buffer, IsAllocated())

65.6 GPU Test Methodology

  • Step

    Description - Source

65.7 Coverage Gaps Identified

  • Area

    Gap - Recommendation

  • AudioComm Synthesis

    Limited phoneme tests - Add coarticulation tests

  • Weather FSM

    No transition probability tests - Add Monte Carlo tests

  • NFGVision Gaze

    No multi-eye tests - Add stereo depth tests

  • Quantization

    Limited mixed-precision

    Add 1/2/4/8-bit combos

  • Sleep Replay

    No async timing tests - Add race condition tests

65.8 Test Execution Commands

  • Command

    Description

  • UE5Editor -run=NFGTests

    Run all NFG tests

  • UE5Editor -run=NFGTests.GPU

    GPU-specific tests

  • UE5Editor -run=NFGTests.Guard

    Guard tests only

  • UE5Editor -run=NFGTests.NFGLM

    Language model tests

  • UE5Editor -run=NFGTests.Integration

    Integration tests

  • Plugin/System

    Features - LOC

  • NeuroFluxGenesis Core

    120+ - ~195K

  • NFGBackendComputeShaders

    150+ - ~26K shader

  • NFGMultiModulatorRMGA

    45 - ~8K

  • NFGAttentionEngine

    25 - ~5K

  • NFGTemporalRetention

    20 - ~3K

  • NFGNeuroDynamics

    50 - ~4K

  • NFGWorldPredictor

    60+ - ~12K

  • NFGLM

    15 systems - ~5K

  • NFGAudioComm

    50 - ~6K

  • NFGVision

    87 - ~8K

  • NFGMetaRegulator

    35 - ~4K

  • NFGMTDP

    25 - ~2K

  • Config/Serialization System

    170+ - ~15K

  • Weather Simulation

    60+ - ~8K

  • Input/Output Processing

    50+ - ~6K

  • Telemetry & Debugging

    80+ - ~10K

  • Evolution Orchestration

    40+ - ~12K

  • GPU Raycasting (RaycastGPU)

    100+ - ~8K shader

  • TorqueStorm Vehicle Physics

    20 - ~5K

  • TrackGenerator

    56 - ~5K

  • NFGQuantization

    54 - ~3K

  • HallOfFittest Persistence

    37 - ~4K

  • GPU Vehicle Simulation

    62 - ~77K shader

  • GPU RNN (LSTM/GRU/Elman)

    48 - ~70K shader

  • GPU Physics Engine

    94 - ~62K shader

  • GPU Analytics & Statistics

    36 - ~10K shader

  • AgentData System

    32 - ~3K

  • Camera Systems

    75 - ~4K

  • Mutation System

    28 - ~3K

  • GPU Crossover System

    47 - ~5K shader

  • Visualization System

    45 - ~6K

  • NFGManager Subsystems

    133 - ~18K

  • Backend & Provider Interfaces

    60 - ~4K

  • Profile & Version System

    59 - ~5K

  • Sleep Replay System

    51 - ~2K

  • Kernel Components

    67 - ~2K

  • Statistics & Monitoring

    77 - ~2K

  • Weather Subsystems Extended

    287 - ~8K

  • WorldPredictor GPU Shaders

    89 - ~2K shader

  • Model Component Helpers

    66 - ~3K

  • Orchestration Components

    56 - ~2K

  • Debug & Vehicle Systems

    87 - ~5K

  • GPU Shader Architecture

    200+ - ~60K shader

  • Configuration System

    1,030+ params - ~15K JSON

  • Provider & Interface Patterns

    12 interfaces - ~2K

  • Logging & Telemetry

    52 categories - ~4K

  • Serialization & Persistence

    150+ - ~5K

  • Build System & Module Dependencies

    18 modules - ~2K

  • FeedForward Shader Implementation

    85 - ~4K shader

  • CMA-ES GPU Kernels

    35 kernels - ~1.6K shader

  • Crossover & Mutation Shaders

    95 - ~3K shader

  • Quantization Codec Details

    42 - ~1K shader

  • Widget & UI System

    31 widgets - ~4K

  • Remote IO & Authentication

    48 - ~3K

  • Input Processing Pipeline

    65 - ~2K

  • CPU Crossover Implementation

    115 - ~4K

  • RDG Shader Dispatch Pipeline

    78 - ~6K

  • NFGLM Tokenizer & Decoder

    42 - ~1K

  • AudioComm Synthesis Internals

    68 - ~2K

  • NFGVision Retina System

    87 - ~8K

  • Weather State Machine Internals

    72 - ~3K

  • GPU Structured Buffer Layouts

    120+ buffers - ~4K header

  • Technique

    Novelty - Section

  • MMS Neurons

    Learnable state machines per neuron - 1.1

  • TS-ReLU

    8 evolvable parameters per neuron - 1.2

  • GPU CMA-ES

    Full eigendecomposition on GPU - 2.2

  • STDP Scale Evolution

    Quantization via spike-timing - 1.5

  • 4-Channel Neuromodulation

    DA/NE/ACh/5-HT with decorrelation - 3.1

  • 8-Timescale Memory

    EMA cascade for temporal integration - 4.3

  • Coherent Neuron Crossover

    Preserve neuron integrity in genetics - 1.6

  • What-If Rollouts

    Counterfactual reasoning engine - 6.2

  • Emergent Communication

    Evolved speech protocol - 8.5

  • Predictive Coding STDP

    Prediction error modulates plasticity - 5.3

  • Sleep Replay

    Offline memory consolidation - 5.4

  • UCB Meta-Learning

    Bandit-based output optimization - 10.1

  • NFGLM

    Evolved language model (no backprop) - 7.0

  • 2-64 Bit Quantization

    Per-neuron evolved bit-depth - 13.3

  • Atmospheric Weather Physics

    Realistic pressure cells + lapse rate + CAPE - 14.2

  • 320K GPU Rays/Frame

    BVH-accelerated raycasting - 18.1

  • Sector Aggregation

    360 rays → 24 network inputs - 18.2

  • 50+ Console Commands

    Runtime debugging infrastructure - 16.2

  • Architecture Hash Versioning

    Model compatibility verification - 13.4

  • 6-Mode Visualization

    Real-time neural network display - 16.3

  • 17 JSON Config System

    Complete hyperparameter externalization - 13.1

  • TorqueStorm SOA Physics

    1000+ vehicles via Structure-of-Arrays - 19.1

  • Ackermann Steering

    Geometrically accurate turning - 19.2

  • Procedural Track Generation

    Perlin noise + hill-climbing optimization - 20.1

  • KD-Tree Spline Cache

    O(log n) nearest-point queries - 20.3

  • Pacejka Magic Formula

    Industry-standard tire physics - 23.1

  • GPU GJK/EPA Collision

    Full narrowphase on GPU - 25.2

  • Sequential Impulse Solver

    Constraint-based physics - 25.3

  • GPU LSTM/GRU Gates

    Full RNN on GPU with RMGA - 24.1

  • Per-Gate RMGA Modulation

    4 neuromodulator channels per gate - 24.4

  • GPU Population Statistics

    Box-Muller + fitness-weighted means - 26.1

  • Binary Search Elite Insertion

    O(log N) Hall of Fittest - 22.1

  • Type-Safe AgentData

    8 genetic data types with virtual crossover - 27.1

  • 6-DOF Free Camera

    Acceleration-based navigation system - 28.1

  • NFG Vision Gaze Control

    AI-controlled scene depth capture - 28.3

  • Adaptive Mutation Scheduling

    6 rate strategies including cosine annealing - 29.2

  • GPU SBX Crossover

    Simulated Binary Crossover on GPU - 30.1

  • Neuron-Coherent Crossover

    Preserves neuron integrity during genetics - 30.1

  • 6-Mode Visualization

    Real-time weight/activation/gradient display - 31.1

  • HISM Neural Rendering

    Instanced mesh for thousands of neurons - 31.2

  • 12-Subsystem NFGManager

    Modular orchestration architecture - 32

  • 12 Provider Interfaces

    Extensible plugin architecture - 33

  • Profile Hot-Swap

    Live config reload without restart - 34.1

  • 11-File Config Merging

    Distributed hyperparameter management - 34.2

  • Sleep Replay Consolidation

    Biology-inspired experience replay - 35.1

  • Priority-Based Sampling

    High-reward experience selection - 35.3

  • 10 Activation Functions

    Tanh to GELU with derivatives - 36.3

  • 13 Weight Init Strategies

    Xavier to Adaptive selection - 40.1

  • Multi-Scale Death Tracking

    Minute/Hour/Day statistics - 37.1

  • Kernel Performance Monitor

    GFLOPS + bandwidth estimation - 37.4

  • Weather Front Physics

    Geostrophic + thermal wind - 38.2

  • 8-State Weather Machine

    Probabilistic state transitions - 38.3

  • 100x100 Atmospheric Grid

    Spatial weather simulation - 38.4

  • DreamerV3 Action Context

    Multi-frame history injection - 39.1

  • What-If Planning

    Danger + confidence scoring - 39.2

  • Counterfactual Feedback

    Per-query error deltas - 39.3

  • Orchestration Layer

    5-component coordination - 41

  • 12-State Tutorial Animation

    Interactive NN visualization - 42.5

  • Wave-Level Attention

    WaveActiveSum() for intra-warp sums - 43.1

  • Half-Precision Q/K/V

    4× bandwidth reduction in attention - 43.1

  • Attention De-Gating STDP

    Recover pre-attention activation for learning - 43.4

  • Welford's Online Z-Score

    Streaming normalization for RMGA - 43.5

  • Serotonin De-Correlation

    Remove DA component from 5-HT signal - 43.5

  • 105-File Test Suite

    350+ test cases across 17 plugins - 44

  • 1,030+ Config Parameters

    15 JSON files for complete tunability - 45

  • 12 Provider Interfaces

    Modular extensibility pattern - 46

  • 3,277+ Blueprint Items

    Full Blueprint scripting support - 47

  • 103 Console Variables

    Runtime tuning for all subsystems - 48

  • 25 Log Categories

    Comprehensive diagnostic infrastructure - 49

  • FArchive Binary Format

    Half-precision with backward compat - 50.1

  • Base64 Quantization Payload

    Compact JSON weight serialization - 50.2

  • 18-Module Plugin Architecture

    Modular UE5 plugin dependency graph - 51.1

  • Shader Directory Registration

    Build.cs shader path management - 51.4

  • 6-Kernel FeedForward GPU

    Standard/MMS/Elman/LSTM/GRU/TSReLU - 52.1

  • 10 GPU Activation Functions

    Tanh→Mish with wave-level reduction - 52.2

  • 35-Kernel CMA-ES GPU

    Full eigendecomposition on GPU - 53.1

  • Low-Rank Covariance

    O(nk) vs O(n²) memory - 53.4

  • IPOP/BIPOP Restart

    Automatic population resizing - 53.6

  • 6 GPU Crossover Methods

    Uniform/SinglePoint/SBX/Arithmetic/BLX-α - 54.1

  • Neuron-Coherent Swap

    Preserve all weights to single neuron - 54.3

  • 6 Mutation Rate Schedules

    Cosine annealing + warm restart - 54.5

  • PCG + Box-Muller RNG

    Thread-safe GPU random generation - 54.7

  • 1/2/4/8-Bit Pack/Unpack

    32/16/8/4 values per uint32 - 55.1

  • STDP-Evolved Quantization Scales

    Spike timing modulates precision - 55.5

  • 31 Custom Slate Widgets

    Real-time NN visualization - 56.1

  • 6 Visualization Modes

    Weights/Activations/Gradients/Attention/MMS/Spikes - 56.3

  • 5 Layout Algorithms

    Grid/Force-Directed/Hierarchical/Circular - 56.4

  • JWT Authentication Flow

    License→Token→Refresh→Sign - 57.2

  • Architecture Hash Filtering

    Only download compatible models - 57.4

  • 7-Stage Input Pipeline

    Raw→Normalize→Slot→Temporal→Feature→Gate→Buffer - 58.1

  • 15 Input Source Types

    Raycast/Velocity/Depth/Audio/Memory - 58.2

  • 4 Normalization Methods

    MinMax/ZScore/Tanh/Log - 58.3

  • Priority Slot Reservation

    Conflict-free input mapping - 58.4

  • 4,165-Line CPU Crossover

    Complete genetic operator implementation - 59.1

  • 8 Crossover Strategies

    Uniform/SinglePoint/TwoPoint/LayerWise/Arithmetic/BLX-α/SBX/Tournament - 59.2

  • Neuron-Coherent MMS Crossover

    8 states + 64 transitions inherited together - 59.4

  • LSTM/GRU Gate Coherence

    All 4 gates from same parent for integrity - 59.6

  • Quantization-Aware Crossover

    Dequant→Crossover→Requant pipeline - 59.5

  • RDG Shader Dispatch

    FRDGBuilder with automatic resource transitions - 60.1

  • 12 RDG Pass Types

    FeedForward/STDP/CMA-ES/Crossover/Attention/WorldPredictor - 60.3

  • Async Compute Overlap

    Parallel compute + graphics execution - 60.4

  • 4-Tier Buffer Lifetime

    Transient/Persistent/Pooled/External - 60.5

  • UTF-8 Byte-Level Tokenizer

    256 vocab, no BPE, direct byte encoding - 61.1

  • Mixed-Radix 4×4 Decoder

    16 outputs → 256 tokens - 61.4

  • Deterministic LM Decoding

    Pure argmax, no sampling/temperature - 61.5

  • 7-Stage Sequence Runner

    Init→Read→Write→Decode→EOS→MaxLen→Assemble - 61.6

  • 20-Parameter AudioComm

    F0/Voicing/Formants/Tremolo/Vibrato - 62.1

  • Rosenberg Glottal Model

    Open/Return/Closed phases with aspiration - 62.2

  • 3×2nd-Order Formant Filters

    IIR bandpass for F1/F2/F3 - 62.3

  • 9-Phoneme Interpolation

    A/E/I/O/U/M/N/S/SH with coarticulation - 62.4

  • Speaker/Listener System

    Distance attenuation + directional gain + occlusion - 62.6

  • Per-Agent Depth Capture

    Individual SceneCaptureComponent per agent - 63.1

  • 5-Parameter Gaze Control

    Yaw/Pitch/Roll/FOV/FocusDistance - 63.2

  • Depth Packing Formats

    R8/R16F/R32F/RGBA8 with log encoding - 63.5

  • Foveated Multi-Eye

    High-res center, low-res periphery - 63.6

  • 8-State Weather FSM

    Clear→Cloudy→Rain→Thunder→Snow→Blizzard - 64.1

  • 4 Weather Front Types

    Cold/Warm/Occluded/Stationary - 64.3

  • Atmospheric CAPE/K-Index

    Real-time convective stability - 64.4

  • 5-Level Precipitation Model

    Drizzle→Light→Moderate→Heavy→Snow - 64.5

  • 4-Factor Visibility Calculation

    Rain/Fog/Snow/Dust interaction - 64.6

  • 120+ GPU Structured Buffers

    Complete neural architecture on GPU - 65.1

  • MMS Buffer Suite

    Mode/Gain/Bias/Transition/State/Decay/Eta - 65.2

  • CMA-ES Buffer Suite

    Mean/Covariance/Eigen/Sigma/Paths/Samples - 65.4

  • RMGA 4-Channel Buffers

    DA/5-HT/NE/ACh with variance tracking - 65.6

  • 148-File Test Suite

    Comprehensive coverage across 12 plugins - 66.1

  • 6 Test Categories

    Guard/GPU/Integration/Unit/Boundary/Performance - 66.2

  • GPU Test Methodology

    Setup→Upload→Dispatch→Readback→Verify - 66.6

  • Coverage Gap Analysis

    Identified 5 areas for improvement - 66.7

  • Biological Retina Encoder

    2048 semantic features from raw pixels - 66.1

  • ON/OFF Temporal Pathways

    Separate approaching/receding detectors - 66.2

  • Color Opponency (RG/BY)

    Parvocellular/Koniocellular cell models - 66.2

  • Multi-Scale DoG (3 levels)

    Fine/Medium/Coarse center-surround - 66.3

  • 4-Direction Orientation Energy

    V1 simple cell model (0°/45°/90°/135°) - 66.4

  • Reichardt Motion Correlators

    Biological motion detection in 4 directions - 66.5

  • Fovea + Periphery Pooling

    8×8 center + 4×4 surround (80 regions) - 66.6

  • 24-Channel Feature Vector

    Depth/Luma/Color/Motion/Orientation per region - 66.7

  • 97% Data Compression

    65K pixels → 2048 semantic floats - 66.13

  • GBuffer-Based Capture

    Single render pass for Depth+Normals+Color - 66.1

NFGVision Retina System (87 features)

119 Features

66.1 Capture Pipeline (GBuffer-Based)

  • GBuffer Capture Mode

    Single base pass → SceneTextures gather (Depth+Normals+BaseColor)

    NFGVisionViewExtension.cpp:289-441

  • CustomRenderPass Integration

    UE5 FCustomRenderPassBase with per-slice rendering

    NFGVisionViewExtension.cpp:67-286

  • Texture2DArray Output

    All agents in single array texture (slices = agents)

    NFGVisionViewExtension.cpp:879-913

  • Staggered Updates

    Reuse previous frame textures for partial updates

    NFGVisionViewExtension.cpp:444-460

  • ShowOnly/Hidden Filtering

    Primitive visibility control per capture

    NFGVisionViewExtension.cpp:545-601

  • Deferred Shading Support

    Depth/Normals from GBuffer, BaseColor from base pass

    NFGVisionGBufferGather.usf:1-38

  • 4 Output Channels

    Depth, WorldNormals, SceneColor, BaseColor

    NFGVisionBackend.h:77-92

66.2 Retina Preprocessing (Per-Pixel Features)

  • Feature

    Formula/Description - Biological Analogon - Source

  • Depth Normalization (D)

    D = rawDepth / MaxDistanceCm - Distance perception

    NFGRetinaPreprocess.usf:106

  • Temporal Depth Delta (dD)

    dD = D - prevD - Looming detection (collision avoidance)

    NFGRetinaPreprocess.usf:108

  • Luminance (Y)

    Y = 0.2126*R + 0.7152*G + 0.0722*B - Magnocellular pathway - NFGRetinaCommon.ush

  • Gradient Magnitude (Gn)

    Gn = tanh(sqrt(Yx² + Yy²) / S_G) - Edge detection (V1 simple cells)

    NFGRetinaPreprocess.usf:127-128

  • Red-Green Opponency (RGn)

    RGn = tanh((R - G) / S_RG) - Parvocellular P-cells

    NFGRetinaPreprocess.usf:131-132

  • Blue-Yellow Opponency (BYn)

    BYn = tanh((B - 0.5*(R+G)) / S_BY) - Koniocellular K-cells

    NFGRetinaPreprocess.usf:133

  • Surface Inclination (Inc)

    Inc = 0.5 * (1 - Nz) - Surface orientation

    NFGRetinaPreprocess.usf:141

  • Normal Edge (Nedge)

    Nedge = tanh(dN / S_N) - Object boundary detection

    NFGRetinaPreprocess.usf:142-143

  • Depth ON-Cell

    depthOn = max(0, dD) - Approaching objects

    NFGRetinaPreprocess.usf:145

  • Depth OFF-Cell

    depthOff = max(0, -dD) - Receding objects

    NFGRetinaPreprocess.usf:146

  • Luma ON-Cell

    lumaOn = max(0, dY) - Brightening regions

    NFGRetinaPreprocess.usf:147

  • Luma OFF-Cell

    lumaOff = max(0, -dY) - Darkening regions

    NFGRetinaPreprocess.usf:148

66.3 Multi-Scale Difference-of-Gaussians (DoG)

  • Scale

    Center Radius - Surround Radius - Function - Source

  • Fine

    1 pixel - 2 pixels - Edge detection

    NFGRetinaPreprocess.usf:150-152

  • Medium

    2 pixels - 4 pixels - Texture boundaries

    NFGRetinaPreprocess.usf:154-156

  • Coarse

    4 pixels - 8 pixels - Large-scale structure

    NFGRetinaPreprocess.usf:158-160

66.4 Orientation Energy (Simple Cell Model)

  • Orientation

    Formula - Description - Source

  • 0° (Horizontal)

    e0 = abs(Yx) - Horizontal edges

    NFGRetinaPreprocess.usf:175

  • 45° (Diagonal)

    e45 = abs(k*(Yx + Yy)) - Diagonal edges

    NFGRetinaPreprocess.usf:177

  • 90° (Vertical)

    e90 = abs(Yy) - Vertical edges

    NFGRetinaPreprocess.usf:176

  • 135° (Anti-Diagonal)

    e135 = abs(k*(Yx - Yy)) - Anti-diagonal edges

    NFGRetinaPreprocess.usf:178

66.5 Reichardt Motion Detection

  • Correlator Model

    d = abs((S - P) * (S_neighbor - P_neighbor))

    NFGRetinaPool.usf:215-218

  • 4-Direction Detectors

    0°, 45°, 90°, 135° motion axes

    NFGRetinaPool.usf:206-213

  • Noise Floor Subtraction

    E = max(0, d - NOISE_FLOOR)

    NFGRetinaPool.usf:220-223

  • Motion Magnitude

    Sum of 4 directional energies

    NFGRetinaPool.usf:229

  • Motion Direction Vector

    (Vx, Vy) = weighted sum of directions

    NFGRetinaPool.usf:278-282

  • Sin/Cos Output

    Normalized direction as (sin, cos) pair

    NFGRetinaPool.usf:281-282

66.6 Fovea/Periphery Pooling Architecture

  • Region

    Grid - Pixels - Purpose - Source

  • Full-Frame Periphery

    4×4 = 16 regions - Entire image - Global scene awareness

    NFGRetinaPool.usf:367-401

  • Fovea (Center)

    8×8 = 64 regions - Configurable subset - Detail in gaze direction

    NFGRetinaPool.usf:334-365

66.7 Output Feature Layout (2048 floats per agent)

  • Offset

    Count - Content - Source

  • 0-1919

    24 channels × 80 - Regional statistics (Fovea + Periphery)

    NFGRetinaPool.usf:342-365

  • 1920-2047

    128 - Global image statistics

    NFGRetinaPool.usf:637-699

  • Ch

    Feature - Range - Description

  • 0

    meanD - [0,1] - Mean depth

  • 1

    meandD - [-1,1] - Mean depth change

  • 2

    meanGrad - [0,1] - Mean depth gradient

  • 3

    varD - [0,1] - Depth variance

  • 4

    depthOn - [0,1] - Approaching signal

  • 5

    depthOff - [0,1] - Receding signal

  • 6

    meanY - [0,1] - Mean luminance

  • 7

    meanGn - [0,1] - Mean edge strength

  • 8

    varY - [0,1] - Luminance variance

  • 9

    lumaOn - [0,1] - Brightening signal

  • 10

    lumaOff - [0,1] - Darkening signal

  • 11

    meanRG - [-1,1] - Red-green opponency

  • 12

    meanBY - [-1,1] - Blue-yellow opponency

  • 13

    absRG - [0,1] - Color saturation (RG)

  • 14

    absBY - [0,1] - Color saturation (BY)

  • 15

    meanInc - [0,1] - Surface inclination

  • 16

    meanEdge - [0,1] - Normal discontinuity

  • 17

    varEdge - [0,1] - Edge variance

  • 18

    DoGD - [-1,1] - Depth center-surround

  • 19

    DoGY - [-1,1] - Luma center-surround

  • 20

    motionMag - [0,1] - Motion magnitude

  • 21

    motionSin - [-1,1] - Motion direction (sin)

  • 22

    motionCos - [-1,1] - Motion direction (cos)

  • 23

    orientEnergy - [0,1] - Orientation strength

66.8 Global Statistics (128 floats)

  • Offset

    Feature - Description - Source

  • 0-3

    Depth - min, max, mean, variance

    NFGRetinaPool.usf:638-641

  • 4-7

    Depth Delta - min, max, mean, variance

    NFGRetinaPool.usf:643-646

  • 8-11

    Luminance - min, max, mean, variance

    NFGRetinaPool.usf:648-651

  • 12-15

    Gradient - min, max, mean, variance

    NFGRetinaPool.usf:652-656

  • 16-23

    Color (RG/BY) - min, max, mean, variance each

    NFGRetinaPool.usf:658-666

  • 24-31

    Surface (Inc/Edge) - min, max, mean, variance each

    NFGRetinaPool.usf:668-676

  • 32-35

    Depth Gradient - min, max, mean, variance

    NFGRetinaPool.usf:678-681

  • 36-38

    Threshold Ratios - near, far, salient

    NFGRetinaPool.usf:683-685

  • 39-42

    Abs Color - mean, variance (RG, BY)

    NFGRetinaPool.usf:687-690

  • 43-46

    Motion - mean, variance, max, ratio

    NFGRetinaPool.usf:692-695

  • 47-127

    Reserved - Zero-padded for alignment

    NFGRetinaPool.usf:697-699

66.9 Temporal State Management

  • Buffer

    Content - Purpose - Source

  • RetinaPrevDepthTex

    Previous frame depth - dD computation

    NFGVisionBackend.h:98

  • RetinaPrevGnTex

    Previous gradient energy - Reichardt correlator

    NFGVisionBackend.h:99

  • RetinaPrevLumaTex

    Previous luminance - dY computation

    NFGVisionBackend.h:100

  • bRetinaPrevValid

    First-frame guard - Skip temporal on frame 0

    NFGVisionBackend.h:107

66.10 Gaze Control System

  • GazeYaw

    ±MaxClampYawDeg - Horizontal look direction

    NFGVisionSceneComponent.cpp

  • GazePitch

    ±MaxClampPitchDeg - Vertical look direction

    NFGVisionSceneComponent.cpp

  • GazeRoll

    ±MaxClampRollDeg - Camera tilt

    NFGVisionSceneComponent.cpp

  • DeltaMode

    bool - Outputs are rates vs absolute angles - FGazeControlParameters

  • SmoothingTau

    seconds - Exponential smoothing time constant - FGazeControlParameters

66.11 Configuration Parameters

  • MaxDistanceCm

    20000 - Depth normalization range (200m) - nfg.eye.max_cm CVar

  • BlackClipCm

    0 - Near clip distance - nfg.eye.black_cm CVar

  • Gamma

    1.6 - Depth gamma correction - nfg.eye.gamma CVar

  • S_RG, S_BY

    0.3 - Color opponency sensitivity

    NFGRetinaPreprocess.usf

  • S_G

    0.1 - Gradient sensitivity

    NFGRetinaPreprocess.usf

  • S_N

    0.5 - Normal edge sensitivity

    NFGRetinaPreprocess.usf

  • T_NEAR

    0.1 - Near threshold for counting

    NFGRetinaPool.usf

  • T_FAR

    0.9 - Far threshold for counting

    NFGRetinaPool.usf

  • NOISE_FLOOR

    0.001 - Motion noise rejection

    NFGRetinaPool.usf

66.12 GPU Shader Pipeline

  • Pass

    Shader - Threads - Description

  • 1. GBuffer Gather

    NFGVisionGBufferGather.usf - 8×8×1

    Extract Depth/Normals from SceneTextures

  • 2. Retina Preprocess

    NFGRetinaPreprocess.usf - 8×8×1 - Per-pixel feature extraction → 6 textures

  • 3. Prev Update

    NFGRetinaPrevUpdate.usf - 8×8×1 - Copy current → temporal buffers

  • 4. Retina Pool

    NFGRetinaPool.usf - 8×8×1 - Regional statistics → 2048 floats

  • Texture

    Content (RGBA)

  • RetinaA

    D, dD, Y, Gn

  • RetinaB

    RGn, BYn, Inc, Nedge

  • RetinaC

    DepthOn, DepthOff, LumaOn, LumaOff

  • RetinaD

    DoGDepth, DoGLuma, OrientEnergy, Reserved

  • RetinaE

    DoGDepthMed, DoGDepthCoarse, DoGLumaMed, DoGLumaCoarse

  • RetinaF

    Orient0, Orient45, Orient90, Orient135

66.13 Data Compression Ratio

  • Input

    Output - Compression

  • 128×128×4 channels = 65,536 floats

    2,048 floats - 97% reduction

66.14 Unique Differentiators vs Standard CV

  • NFGVision Retina

    Standard CNN Vision

  • ON/OFF temporal pathways

    Static gradients only

  • Fovea + Periphery pooling

    Uniform resolution

  • Color Opponency (RG/BY)

    RGB channels

  • Reichardt Motion Correlators

    Optical flow (expensive)

  • Fixed 2048 output (stable for evolution)

    Variable feature maps

  • No backpropagation needed

    Requires labeled training

  • GPU Compute Shaders (UE5 RDG)

    External frameworks (PyTorch/TF)

NFGServer & Distributed Learning

9 Features

90.1 Persistence and Model Registry

  • Model persistence

    Node.js backend persists models, architectures, runs, and telemetry in MongoDB

    NEUROFLUX_GENESIS_BOOK_EN.html:3.9

  • REST/WebSocket uploads

    Engine clients upload models via REST/WebSockets

    NEUROFLUX_GENESIS_BOOK_EN.html:3.9

  • Architecture validation

    Server validates architecture and weights on upload

    NEUROFLUX_GENESIS_BOOK_EN.html:3.9

  • Hash deduplication

    Computes hashes to deduplicate uploads

    NEUROFLUX_GENESIS_BOOK_EN.html:3.9

90.2 Distributed Training Orchestration

  • Training queues

    Training queues let distributed workers process jobs and report results

    NEUROFLUX_GENESIS_BOOK_EN.html:3.9

90.3 Distributed Learning and Analytics

  • Cross-world learning

    Many worlds evolve a shared model family and redistribute best configurations

    NEUROFLUX_GENESIS_BOOK_EN.html:perspective

  • Metadata capture

    Uploads include architecture, weights/biases, input/output mappings, fitness, and hardware stats

    NEUROFLUX_GENESIS_BOOK_EN.html:6.5

  • REST/WS analytics

    Exposes models and telemetry via REST/WS APIs for analysis and dashboards

    NEUROFLUX_GENESIS_BOOK_EN.html:6.5

  • Cross-game reuse

    Persisted models enable reuse across games or simulations

    NEUROFLUX_GENESIS_BOOK_EN.html:6.5

Durchbrüche

Durchbrüche, die NFG definieren

Ausgewählte Entwicklungen mit echter Differenzierung.

GPU-native Evolutions-Stack

CMA-ES Eigenzerlegung, Crossover- und Mutations-Kernels sowie Hall-of-Fittest-Orchestrierung für große Populationen.

MMS-Neuronen und TS-ReLU

Per-Neuron-Statemachines mit evolvierbaren Dynamiken jenseits statischer Layer.

Neurodynamik in Echtzeit

STDP-Plastizität kombiniert mit vier Neuromodulator-Kanälen (DA/NE/ACh/5-HT), die Lernen formen.

Wahrnehmung für Agenten

GPU-Raycasting plus Retina-Style GPU-Feature-Extraktion für autonome Wahrnehmung.

World Prediction und What-if-Rollouts

Mehrstufige GPU-Rollouts mit Action-Varianten zur Echtzeit-Bewertung von Outcomes.

NFGServer als Backbone für verteiltes Lernen

Node.js- und MongoDB-Backend, das Modelle und Telemetrie persistiert, Uploads validiert und Trainings-Queues für weltweit vernetztes Lernen orchestriert.

Diligence

Technischer Walkthrough gewünscht?

Wir bieten tiefe technische Sessions, Build-Zugang und Roadmap-Briefings für Investoren und Studios.

E-Mail an Coheron