Artificial neural networks are important in everyday life and are becoming more widespread. For this reason, it is crucial they are understood and tested. This paper tests and compares two training methods: reinforcement learning with backpropagation and an evolutionary method. The hypothesis is that the training method using backpropagation and reinforcement learning is more efficient in training a neural network to play a game than a model trained with the evolutionary algorithm. However, the model trained with backpropagation and reinforcement learning will have lower performance than a model trained with the evolutionary algorithm. To research the hypothesis, a feedforward neural network and how it works must first be explained.

Neural networks are systems inspired by the biological brain which enables a computer to predict, model, classify and many other applications. All this by learning from some set of training data to find general relations that can be applied to unseen data. A neural network model is essentially a function with potentially thousands of parameters. Just like any other function, input values are provided and with those, the output is calculated. In a feedforward neural network, this process is called feedforward.

The process of feedforward is meaningless with a model that has not yet been configured to do anything. A neural network must first be taught to perform a certain task. This is what is accomplished with machine learning. Backpropagation is an example of a machine learning method. For backpropagation two things are required: the input and the corresponding output. Backpropagation will adjust the parameters of a model so the next time the same input is provided, the output will be closer to the desired output. This is called optimisation.

Reinforcement learning is a way to teach a neural network by giving it positive reinforcement when it does something good and negative reinforcement when it does something bad. This is used when no desired output is known so backpropagation cannot directly be applied.

An evolutionary algorithm is much more intuitive than backpropagation. It is the imitation of natural selection in biology, but with self-determined factors deciding the fitness of a model. When training a neural network with an evolutionary algorithm, a large group of random models will be generated, all performing the same task. Some models, however, will be better suited for this task than others. How well they are suited to their environment is their fitness. This will be the determining factor of who survives and can therefore reproduce and create mutated offspring. This process is repeated as many times as required to reach the desired performance.

The hypothesis of this paper has been proven wrong. Neural networks trained with an evolutionary algorithm do end up performing at a higher level than models trained with reinforcement learning and backpropagation. However, Neural networks trained with an evolutionary algorithm are also more efficient with regard to not only the number of cycles needed to reach the same performance but also with regard to the time required.