Recurrent Neural Network Experiments

Introduction

This is a repo in which I would experiment with a brain like recurrent neural network.

Concept Explanation

Each neuron will be connected to a bunch of others including the input neurons that will activate every cycle with the desired input values and some output neurons that if activated will be the way to interact with the enviroment.

Network Learning

Taking into account the difficulty that it would be to use backpropagation in this type of network, it would basicly evolve natural selection, i.e, mutations and survival of the fittest

Algorithm (temporary basic explanation)

Every cycle the following will occur

for input_neuron in network.input_neurons:
    input_neuron.activated = True
    input_neuron.value = input_value

for neuron in network.neurons:
    if neuron.activated:
        for conected_neuron in neuron.conected_neurons:
            concected_neuron.value += neuron.value * connection.weight

for neuron in network.neurons:
    if neuron.value >= neuron.activation_threshold:
        neuron.activated = True

for output_neuron in network.output_neurons:
    if output_neuron.value >= output_neuron.activation_value:
        output_neuron.make_action()

First Experiment

After making the basics of the network class I would try it to make a small 2d car drive on its own in a little circuit.

WARNING

If you want to review the car game code beware that it is mainly in spanish (I may translate this in the future)

Potential Expansions

Add posible mutations to the number of neurons

Observation log

Survival of the living fittest

The cars would just learn to stay still after only a feww points

Survival of the best

Sometimes cars learn to drive the complete circuit but almost always follows the same strategy and never learn to go faster as just one genetic line is preserved per generation

Reproduction based on points

When multiple genetic lines are let to reproduce it stucks far less often and also it increases the optimizations in time

Another map with more point lines

It Does learn in other maps but drifting sometimes stucks the evolution

Change circuit in each generation

Learning is achieved in both circuits which proves that it doenst learn the circuit it *actually* learn to drive