Neural network based AI for racing games

Abstract

The usage of neural network-based artificial intelligence in game industry is still minimal, particularly affected by the difficulty and the lengthy training of the neural network itself. Moreover, deciding neural network’s many parameters is very difficult. Yet, it has a lot of potential as a machine learning technique where it can learn by itself which will help the development of the AI. This report presents simplified topological evolution method for evolutionary neural network algorithm called NEAT by changing the topology structure to a fully-connected multi layered perceptron (MLP) structure. It also proposes the new mutation function to be used along with the new topological structure. The performance of the new method is then tested using a 2D racing car simulation model in three tracks of increasing difficulty. Our modified evolution method results in simpler neural network topology. Specifically for our experiment, our method results in significantly lesser number of neurons and layers than NEAT. It is further shown that the adaptation maintains the performance of the original algorithm while reducing the evolution time by 17 percent.