mlp

implementation of a multilayer perceptron which can be trained using the backpropagation algorithm

installation

npm install mlp

Using browserify, mlp can be used in a browser as well.

how to use

initialize a new perceptron

// create the perceptron
var MLP = require('mlp');
var mlp = new MLP(2,3);

// add hidden layers and initialize
mlp.addHiddenLayer(5);
mlp.addHiddenLayer(5);
mlp.init();

The amount and dimensions of the hidden layers has to correspond with the problem, you want to solve (think about overfitting)...

train the perceptron

// create a training set
mlp.addToTrainingSet([2, 1], [1, 0, 0]);
mlp.addToTrainingSet([0, 1], [1, 0, 0]);
mlp.addToTrainingSet([3, 4], [0, 1, 0]);
mlp.addToTrainingSet([2, 3], [0, 1, 0]);
mlp.addToTrainingSet([1, 0], [0, 0, 1]);
mlp.addToTrainingSet([3, 1], [0, 0, 1]);

// train the perceptron
var learnRate = 0.5;
var error = Number.MAX_VALUE;
while (error > 0.01) {
	error = mlp.train(learnRate);
}

Depending on the training set, the learn rate and the structure of the perceptron, this step may take a while. If necessary, you can add additional exit conditions to the loop (e.g. a maximum number of iterations).

If you want to restart the training, you have to clear the training set and to reset the weights:

mlp.clearTrainingSet();
mlp.resetWeights();

classify element

var elementToClassify = [1, 1];
var classification = mlp.classify(elementToClassify);

The vector classification contains in each component the probability that the element belongs to the associated class.

export and import of mlp data

After training the perceptron, you can export the weights to save them for later use:

var data = mlp.exportToJson();

The export is a simple JSON object that can be imported again as follows:

mlp.setWeights(data);

Changelog

1.1.0 / 2016-06-30

Add methods exportToJson and setWeights to export and import a perceptron.

License

Copyright (c) 2015-2016 Ulf Biallas. Licensed under the [MIT license][MIT]. [MIT]: https://github.com/ulfbiallas/npm-mlp/blob/master/LICENSE