nntsreinforcement/lib/neuralnetwork.js

A Neural Network for Reinforcement Learning in TypeScript

"use strict";
Object.defineProperty(exports, "__esModule", { value: true });
exports.NeuralNetwork = void 0;
var layer_1 = require("./layer");
var NeuralNetwork = /** @class */ (function () {
    function NeuralNetwork(numOfNeurons, layers) {
        this.layers = [];
        if (layers == undefined) {
            // Create new layers
            for (var _i = 0, numOfNeurons_1 = numOfNeurons; _i < numOfNeurons_1.length; _i++) {
                var numOfNeuronsForLayer = numOfNeurons_1[_i];
                this.layers.push(new layer_1.Layer(numOfNeuronsForLayer));
            }
        }
        else {
            if (layers.length == numOfNeurons.length) {
                this.layers = layers;
            }
            else {
                console.warn('The length of layers is different to the length of neurons');
            }
        }
        this.numOfNeurons = numOfNeurons;
    }
    NeuralNetwork.prototype.initRandomConnections = function () {
        for (var i = 1; i < this.layers.length; i++) {
            for (var _i = 0, _a = this.layers[i].neurons; _i < _a.length; _i++) {
                var neuron = _a[_i];
                neuron.initRandomConnectionsFromLayer(this.layers[i - 1]);
            }
        }
    };
    NeuralNetwork.prototype.feedForward = function (inputs) {
        this.layers[0].setValues(inputs);
        for (var i = 1; i < this.layers.length; i++) {
            var layer = this.layers[i];
            layer.compute(this.layers[i - 1].getActivations());
        }
        return this.layers[this.layers.length - 1].getActivations();
    };
    NeuralNetwork.prototype.mutate = function (rate) {
        for (var _i = 0, _a = this.layers; _i < _a.length; _i++) {
            var layer = _a[_i];
            layer.mutate(rate);
        }
    };
    NeuralNetwork.prototype.deepCopy = function () {
        var layers = [];
        for (var i = 0; i < this.layers.length; i++) {
            var layer = new layer_1.Layer(this.numOfNeurons[i]);
            var neurons = [];
            for (var _i = 0, _a = this.layers[i].neurons; _i < _a.length; _i++) {
                var neuron = _a[_i];
                neurons.push(neuron.deepCopy());
            }
            if (i > 0) {
                //Connect to prev layer
                this.connectNeuronsToLayer(neurons, layers[i - 1]);
            }
            layer.neurons = neurons;
            layers[i] = layer;
        }
        return new NeuralNetwork(this.numOfNeurons, layers);
    };
    NeuralNetwork.prototype.crossover = function (brain) {
        var layers = [];
        for (var i = 0; i < this.layers.length; i++) {
            var layer = new layer_1.Layer(this.numOfNeurons[i]);
            var crossOverPoint = (Math.random() * (this.layers[i].neurons.length - 1)) / 2 -
                (this.layers[i].neurons.length - 1) / 4;
            var neurons = [];
            for (var j = 0; j < this.layers[i].neurons.length; j++) {
                var neuronCopy = this.layers[i].neurons[j].deepCopy();
                for (var k = 0; k < neuronCopy.weights.length; k++) {
                    neuronCopy.weights[k] =
                        j > crossOverPoint
                            ? brain.layers[i].neurons[j].weights[k]
                            : this.layers[i].neurons[j].weights[k];
                }
                neurons.push(neuronCopy);
            }
            if (i > 0) {
                //Connect to prev layer
                this.connectNeuronsToLayer(neurons, layers[i - 1]);
            }
            layer.neurons = neurons;
            layers.push(layer);
        }
        return new NeuralNetwork(this.numOfNeurons, layers);
    };
    NeuralNetwork.prototype.linearInterpolation = function (a, b, d) {
        return a + (b - a) * d;
    };
    NeuralNetwork.prototype.connectNeuronsToLayer = function (neurons, prevLayer) {
        for (var _i = 0, neurons_1 = neurons; _i < neurons_1.length; _i++) {
            var neuron = neurons_1[_i];
            neuron.prevNeurons = prevLayer.neurons;
        }
    };
    return NeuralNetwork;
}());
exports.NeuralNetwork = NeuralNetwork;