Settings Reference
This is the lookup page for every neural-network configuration function. Each is shown
with its arguments, a one-line description, and its default value where it has one. For
fuller explanations, follow the cross-links to the themed pages. Call these functions
before you train, and call RestoreDefaults() at the end of
your formula.
Warning
Settings persist for the formula until you reset them. Always finish a neural-network
formula with RestoreDefaults() so leftover settings don't carry into the
next run or formula.
Model & architecture
See Network Architectures,
TCN & Transformer and
Activation Functions for detail.
| Function | Description — default |
|---|
| SetNeuralNetworkType(type) | Model: 0 = MLP, 1 = LSTM, 2 = GRU, 3 = TCN,
4 = Transformer. Default 0 (MLP). |
| SetRecurrentParams(hidden, layers, seqLen) | Hidden units, recurrent
layers, and window length for LSTM/GRU. layers builds a genuinely
stacked multi-layer network. Ignored for the MLP. |
| SetTcnParams(channels, kernelSize, levels) | TCN only. channels
≥ 1, kernelSize ≥ 2, levels ≥ 1. Default 16, 2, 4. Receptive field
1 + (kernelSize−1)·(2levels−1) is computed automatically. |
| SetTransformerParams(dModel, numHeads, numBlocks, seqLen) | Transformer
only. all ≥ 1 and dModel divisible by numHeads. Default 16, 2, 2, 16.
Feed-forward width is 4·dModel. |
| SetNetworkLayer1…4(n1[, n2[, n3[, n4]]]) | Hidden-layer neuron counts
(sigmoid activation on every layer). 1–4 layers, 1–100 neurons each. MLP. |
| SetNetworkWithActivationLayer1…4(n1, a1, …, outAct) | Hidden-layer
(neurons, activation) pairs plus the output-layer activation. Activation codes
0–9. MLP. |
| SetLayerNorm(flag) | MLP only. 0 = off (default), 1 = on. Layer
Normalization on every hidden layer (output layer never normalized). Disables AFL
export while on. |
| SetCategoricalInput(channelIndex, numCategories, embedDim) | MLP
only. channelIndex ≥ 0, numCategories ≥ 2, embedDim ≥ 1. Marks an input
channel as categorical (integer IDs) mapped to a learned embedding; bypasses input
scaling. Call once per categorical channel (accumulates). Disables AFL export.
Default: none declared. |
Training algorithm, rate & schedule
See Training Algorithms for detail and the
full optimizer code table. The learning-rate schedule is honoured by every
model type (MLP, LSTM, GRU, TCN, Transformer), not just the MLP.
| Function | Description — default |
|---|
| SetLearningAlgorithm(code) | Optimizer, codes 0–13. Default 0 (Backpropagation). |
| SetLearningRate(rate) | Weight-update step size (0–20). Default 0.2. Adam family wants ~0.001–0.005. |
| SetMomentum(m) | Momentum for Backprop / Nesterov (0–1). Default 0. |
| SetSarpropTemperature(t) | SARPROP temperature (0–1). Default 0.02. |
| SetLearningRateSchedule(type) | 0 const, 1 step, 2 cosine, 3 SGDR. Default 0. |
| SetLRScheduleStep(epochs) | Step size / cosine length / first SGDR cycle (≥1). Default 50. |
| SetLRStepDecayFactor(f) | Step-decay multiplier (0–1). Default 0.5. |
| SetLRScheduleMinPercent(p) | Schedule floor as a fraction of base rate (0–<1). Default 0. |
| SetSGDRCycleMultiplier(m) | SGDR cycle-length growth (≥1). Default 2.0. |
Optimizer hyperparameters
| Function | Description — default |
|---|
| SetAdamBeta1(b) | First-moment decay, 0<b<1. Default 0.9. |
| SetAdamBeta2(b) | Second-moment decay, 0<b<1. Default 0.999. |
| SetAdamEpsilon(e) | Numerical-stability constant, e>0. Default 1e-8. |
| SetAdamWeightDecay(wd) | AdamW weight decay, wd≥0. Default 0.01. |
| SetRMSPropDecay(r) | RMSProp decay, 0<r<1. Default 0.9. |
| SetAdadeltaRho(r) | Adadelta rho, 0<r<1. Default 0.95. |
| SetAdadeltaEpsilon(e) | Adadelta epsilon, e>0. Default 1e-6. |
Accuracy & regularization
See Accuracy & Avoiding Overfitting — the most
important settings here.
| Function | Description — default |
|---|
| SetPercentTestingData(pct) | Hold back the most recent pct% as unseen
test data and select the best test-error network (0–100). Default 0 (off). |
| SetEarlyStoppingPatience(n) | Stop if test error hasn't improved for n
epochs (≥0). Default 0 (off). |
| SetDropoutRate(p) | Hidden-unit dropout while training (0–<1). Default 0 (off). |
| SetWeightInit(mode) | 0 uniform, 1 Xavier, 2 He. Default 0. |
| SetGradientClipNorm(n) | Cap the gradient size to stabilise training (≥0). Default 0 (off). |
| SetErrorAlgorithm(code) | Loss: 0 linear, 1 tanh, 2 Huber. Default 1 (tanh). |
| SetHuberDelta(d) | Huber transition point, d>0 (with error code 2). Default 1.0. |
Scaling
See Input & Output Scaling for detail.
| Function | Description — default |
|---|
| SetScalingAlgorithm(algo) | 0 = Mean/StdDev (default), 1 = Min/Max. |
| SetInputScalingMinMax(min, max) | Internal range inputs are scaled into
(−2 to 2, min<max). Default 0..1. |
| SetOutputScalingMinMax(min, max) | Internal range the output is scaled
into (−2 to 2, min<max). Default 0..1. |
Data & training control
| Function | Description — default |
|---|
| SetMaximumEpochs(n) | Maximum training passes over the data. |
| SetBatchSize(n) | 0 = full batch (default); n>0 = minibatch of n.
Honoured by the sequence models (LSTM/GRU/TCN/Transformer) too — real
mini-batch over whole sequences — as well as the MLP. |
| EnableShuffleData() / DisableShuffleData() | Reshuffle data order each
epoch (for minibatch mode). Default disabled. |
| SetMaximumThreads(n) | Worker threads for the interactive train
(1 … number of CPUs). |
| SetEnsembleSize(k) | Train k seed-varied networks and
average their predictions (k ≥ 1; default 1). Any model type. Members save as
<file>.0…<file>.{k−1}; the run path discovers and averages
them automatically. Disables AFL export while k > 1. See
Ensembling. |
Reproducibility & versioning
| Function | Description |
|---|
| SetSeed(n) | Fix the random seed (n ≥ 1) so a run is repeatable. |
| SetVersion(n) | Version tag (n ≥ 0) used to bust the Walk-Forward
indicator cache — bump it to force a retrain. See
Neural Network Functions. |
Output, progress & cleanup
| Function | Description |
|---|
| EnableProgress() / DisableProgress() | Show or hide the training progress
window. |
| EnableNetworkToAFL() / DisableNetworkToAFL() | Export the trained MLP as
an AFL formula alongside the saved network. Auto-skipped (with a trace message) when
the config can't be a formula — any sequence model, LayerNorm on, categorical
embeddings declared, or ensemble size > 1. See
Neural Network Functions. |
| RestoreDefaults() | Reset every neural-network setting to its default.
Call this at the end of every formula. |
Note
Every setting defaults to "off / no change", so a formula that sets nothing trains a
plain MLP with the standard defaults. Out-of-range values are rejected with a trace
message and leave the setting unchanged.