TNeuralNetwork

Source: source/classes/tneuralnet.prg

TNeuralNetwork is FiveWin's feedforward neural network implemented in pure Harbour. It provides a multi-layer perceptron (MLP) with backpropagation training. The network consists of three layer types: an input layer, one or more hidden layers, and an output layer. Each neuron uses the sigmoid activation function. The class is ideal for simple pattern recognition tasks such as the XOR problem.

Architecture

flowchart LR subgraph "TNeuralNetwork" A[oInputLayer
TNeuralLayer] B[aHiddenLayers
Array of TNeuralLayer] C[oOutputLayer
TNeuralLayer] end subgraph "TNeuralLayer" D[aNeurons
Array of TNeuron] end subgraph "TNeuron" E[nBias] F[aWeights] G[nValue] H[nDeltaError] end A --> D B --> D C --> D D --> E D --> F D --> G D --> H

Key DATA Members

TNeuralNetwork

DATA	Type	Default	Description
`oInputLayer`	Object		TNeuralLayer for the input layer
`aHiddenLayers`	Array	`{}`	Array of TNeuralLayer objects (one or more hidden layers)
`oOutputLayer`	Object		TNeuralLayer for the output layer
`nLearningRate`	Numeric	`0.1`	Learning rate for backpropagation weight updates

TNeuralLayer

DATA	Type	Default	Description
`aNeurons`	Array	`{}`	Array of TNeuron objects in this layer

TNeuron

DATA	Type	Default	Description
`nBias`	Numeric	`hb_Random()`	Bias term for this neuron
`aWeights`	Array		Array of connection weights from the previous layer, initialized randomly
`nValue`	Numeric		Current activation value after forward propagation
`nDeltaError`	Numeric		Error gradient computed during backpropagation

Methods

Method	Description
`New( nInputs, aHiddenLayers, nOutputs )`	Create a neural network with `nInputs` input neurons, an array of hidden layer sizes (e.g. `{ 4, 3 }` for two hidden layers), and `nOutputs` output neurons.
`Learn( aInputs, aOutputs, lDebug )`	Train the network on one sample: runs forward propagation, then backpropagation to adjust weights and biases.
`Propagation( aInputs )`	Run forward propagation: compute activation values from input through hidden layers to output. Uses the sigmoid activation function.
`BackPropagation( aInputs, aOutputs )`	Run backward propagation: compute error gradients and update all weights and biases based on the learning rate.

Example: XOR Training

#include "FiveWin.ch"

function Main()

   local oNN, n

   // Network: 2 inputs, hidden layer with 2 neurons, 1 output
   oNN := TNeuralNetwork():New( 2, { 2 }, 1 )

   // Train for 10000 iterations
   for n = 1 to 10000
      oNN:Learn( { 0, 0 }, { 0 } )
      oNN:Learn( { 1, 0 }, { 1 } )
      oNN:Learn( { 0, 1 }, { 1 } )
      oNN:Learn( { 1, 1 }, { 0 } )
   next

   // Test the trained network
   ? "XOR Truth Table:"
   ? "0 XOR 0 =", oNN:Propagation( { 0, 0 } )[ 1 ]  // ~0.0
   ? "1 XOR 0 =", oNN:Propagation( { 1, 0 } )[ 1 ]  // ~1.0
   ? "0 XOR 1 =", oNN:Propagation( { 0, 1 } )[ 1 ]  // ~1.0
   ? "1 XOR 1 =", oNN:Propagation( { 1, 1 } )[ 1 ]  // ~0.0

return nil

Example: Training with Debug Output

#include "FiveWin.ch"

function Main()

   local oNN := TNeuralNetwork():New( 2, { 4 }, 1 )

   // Debug mode prints inputs, outputs, expected values, and errors
   oNN:Learn( { 1, 0 }, { 1 }, .T. )

return nil

Notes

Weights and biases are initialized to random values using hb_Random().
The sigmoid activation function maps values to the range (0, 1). The derivative is used during backpropagation for gradient computation.
The nLearningRate controls how much weights change during each update. Lower values (e.g. 0.01 - 0.1) provide more stable training, while higher values may converge faster but risk oscillation.
Multiple hidden layers can be created by passing an array with multiple values: { 4, 3 } creates two hidden layers with 4 and 3 neurons respectively.
This is a pure Harbour implementation with no external dependencies. The Math_E() function is implemented via Harbour's C extension API.
For more advanced neural network capabilities, see the Transformer class in source/classes/transformer.prg.

TNeuralNetwork

Architecture

Key DATA Members

TNeuralNetwork

TNeuralLayer

TNeuron

Methods

Example: XOR Training

Example: Training with Debug Output

Notes

See Also