Advanced Architecture

NeuraLib_Models

The official architecture expansion for compact temporal, recurrent, attention, Transformer, residual, and reinforcement-learning research inside TradingView.

Updated July 12, 2026Publisher Alien_AlgorithmsPine Script v6

The expansion library

Import NeuraLib first, then NeuraLib_Models. The expansion registers fluent methods directly on NeuraLib Sequential models.

Temporal model with dueling Q-headpine

import Alien_Algorithms/NeuraLib/1 as nl
import Alien_Algorithms/NeuraLib_Models/1 as models

var nl.Sequential model = nl.sequential("advanced_model")

if barstate.isfirst
    model := model
      .input(array.from(8), "sequence")
      .temporalConvStack(4, 2, 3, 2, 2, 1, nl.ActivationKind.relu, 0.0, "temporal")
      .globalAvgPool1d(2, 3, "pool")
      .duelingQHead(4, 2, nl.ActivationKind.relu, "q_head")
      .build(nl.rng(7))

The result remains a normal NeuraLib model. .compile(), .trainOnBatch(), .predict(), .evaluate(), datasets, scalers, and weight utilities work unchanged.

Model families

Family	Methods	Purpose
Residual	`residualDense()`	Deeper feedforward paths that preserve input information.
Temporal convolution	`conv1d(), temporalConvStack()`	Local structure in short flattened sequences.
Pooling	`globalAvgPool1d(), globalMaxPool1d()`	Compress a sequence into feature summaries.
Recurrent	`lstm(), gru()`	Ordered state memory across a compact sequence.
Attention	`selfAttention(), multiHeadSelfAttention(), crossAttention()`	Context mixing across token rows.
Transformers	`transformerEncoder(), transformerEncoderStack(), transformerDecoder()`	Attention plus feedforward residual architecture.
Reinforcement learning	`duelingQHead(), prioritizedReplayBuffer()`	Action values and deterministic prioritized replay.

Temporal convolution and pooling

A sequence of four time steps with two features is represented as eight flattened inputs. Conv1D filters slide across time, while a temporal stack repeats compact convolution blocks. Global average pooling captures broad sequence behavior; max pooling emphasizes the strongest activation.

LSTM and GRU

.lstm() and .gru() scan flattened sequences in time order. Use them when ordering and memory matter more than a single snapshot. Keep time steps and unit counts small: recurrent backpropagation is expensive in Pine.

Attention and Transformers

selfAttention() mixes row context with one attention head.
multiHeadSelfAttention() splits the model dimension across up to eight heads.
crossAttention() uses packed query and memory rows.
transformerEncoder() adds attention, residual paths, normalization, and feedforward transformation.
transformerDecoder() combines causal target attention with memory attention.

Attention shape rule

modelDim must be divisible by headCount. The current implementation supports at most eight heads.

pushExperience() inserts or overwrites a row.
updatePriority() updates a sampled row after observing error.
Sampling is deterministic for the same buffer, arguments, and seed.

Shape rules

Conv1D, temporal stacks, LSTM, and GRU expect timeSteps × featureCount flattened features.
Attention expects each row to contain exactly modelDim columns.
Cross-attention and decoders require documented packed row layouts.
pushPositionalEncoding() adds sinusoidal position features to a FeatureBuilder.
Import the expansion after NeuraLib and avoid method-name collisions with core methods.