Project Summery

This project aims to use the evolutionary algorithm for training a neural network in an environment where there is not enough data to train the network. One of these environments is the game, where something new is always happening; therefore, it is impossible to generate enough data for training the model.

To run the game with the help of neural evolution, we need to design a neural network that takes important decision-making parameters in its input and then produces the corresponding output. In the end, the output produced is like pressing the space button and will guide the game agent on its way to getting higher scores.

This project consists of 2 main steps:

  1. Designing a proper neural network

    a. Defining important decision-making parameters

    b. Initialization of weights and biases

    c. Implementing feedforward algorithm

    d. Designing a good activation function

2. Designing a proper evolutionary algorithm

    a. Designing the next population selection algorithm(Roulette wheel, SUS, and Q-tournament)

    b. Generating new population using cross-over and mutation algorithms