feedforward

author: Arne (voidragon) Simon [[e-mail]]

A simple layered feedforward network with momentum based backward propergation.

It learns fast and it is easy to train a whole population of networks.

Please Donate

• BTC: 17voJDvueb7iZtcLRrLtq3dfQYBaSi2GsU
• ETC: 0x7bC5Ff6Bc22B4C6Af135493E6a8a11A62D209ae5

Example

Using Feedforward

const { Feedforward, Population } = require('feedforward');
 
const pattern = [
    [[0, 0], [0]],
    [[1, 0], [0]],
    [[0, 1], [0]],
    [[1, 1], [1]],
];
 
// --- train - feedforward ---
 
const net = new Feedforward({
    model: [2, 1],
    flexibility: 0.5
});
 
const measure = net.train({
    pattern,
    maxIterations: 70,
    minError: 0.1,
});
 
console.log(`iterations: ${measure.iterations} error: ${measure.error}`);
 
pattern.forEach(p => {
    const res = net.map(p[0]);
    console.log(`${p[1][0]} = ${res[0]} error: ${Math.abs(p[1][0] - res[0])}`);
});

Using Population

// --- train - population ---
 
const pop = new Population({
    model: [2, 1],
    spread: 10,
    flexibility: 0.5
});
 
const measure = pop.train({
    pattern,
    maxIterations: 70,
    minError: 0.1,
});
 
console.log(`iterations: ${measure.iterations} error: ${measure.error}`);
 
pattern.forEach(p => {
    const res = measure.net.map(p[0]);
    console.log(`${p[1][0]} = ${res[0]} error: ${Math.abs(p[1][0] - res[0])}`);
});

Save and load a network

// every thing that is needed is the config and the connections.
const data = JSON.stringify({
    config: net.config,
    connections: net.connections,
});
 
const loaded = JSON.parse(data);
const newNet = Feedforward.fromConnections(loaded.config, loaded.connections);

API

class Feedforward

A simple feedforward neuronal network.

fromConnections(config, connections)[static] -> net

Creates a new network for the config and initializes the connection values.

• config - A feedforward config object.
• connections - 2d array of connection values.
• net - The newly created feedforward network.

new Feedforward(config)

• config
  • model - Array of layer depths.
  • flexibility - How fast the net adepts.
  • activationThreshold - Idicates when an output cell identifies as on, aka 1.0 .

map(inputs) -> outputs

• inputs - Array of input values, normalized between 1.0 and 0.0 .
• outputs - An array of 0.0 or 1.0, indicating if the output neuron is off or on.

correct(outputs)

• outputs - Array of expected output values, normalized between 1.0 and 0.0 .

train(options) -> measure

• options - Training options.
  • pattern - The pattern to learn, aka approximate.
  • maxInterations - Maximum interations for the learning process.
  • minError - The minimum error at wich learning ends.
• measure -
  • error - The avarage error for the pattern.
  • iterations - How many iterations where nessecary for learning the pattern.

class Population

new Population(config)

• model - Array of layer depths.
• spread - How many networks will be generated.
• flexibility - How fast the net adepts.
• activationThreshold - Indicates when an output cell identifies as on, aka 1.0 .

map(inputs) -> outputs

• inputs - Array of input values, normalized between 1.0 and 0.0 .

correct(outputs)

• outputs - Array of expected output values, normalized between 1.0 and 0.0 .

train(options) -> measure

• options - Training options.
  • pattern - The pattern to learn, aka approximate.
  • maxInterations - Maximum interations for the learning process.
  • minError - The minimum error at wich learning ends.
• measure -
  • error - The avarage error for the pattern.
  • iterations - How many iterations where nessecary for learning the pattern.
  • net - The feedforward network with the lowest error.

nextGeneration()

Takes the best network from the last training and prodcues new population based on this network.