When applying genetic programming to our scenarios with neural networks, the genes and genomes represent our internal structure of our neural network, which means that it will contain the following information:
- Layer connections (in a convolutional layer scenario)
- Neuron connections
- Neuron weights
However, to keep things simple we currently assume that our neural network simply only has every neuron connected to its previous layers, so we assume the Feed-Forward Network structure for the sake of easiness.
Serialization
When talking about the identical reproduction of states of instances (or composites), we talk about serialization and deserialization in the programming world.
Assuming our latest code from the previous chapter, our
Brain will now get a serialize() method that pushes all
neuron weights into an array and returns it:
Brain.prototype.serialize = function() {
let weights = [];
this.layers.forEach(layer => {
layer.forEach(neuron => {
weights.push(neuron.value);
});
});
return weights;
};Deserialization
The deserialization part will assume that we have a simulation running with identical neural networks. For now this is just easier to implement.
Brain.prototype.deserialize = function(weights) {
let index = 0;
this.layers.forEach(layer => {
layer.forEach(neuron => {
neuron.value = weights[index++];
});
});
};As you might have guessed already, the weights array is in our scenario identical to the genome that our evolutionary algorithm will produce.