Autonomous virtual vehicles with FNN-GA and Q-learning in a video game environment

Driving autonomous virtual vehicles have been developed by the use of several artificial intelligence (AI) algorithms. However, it is possible to improve their results and find other applications through the use of new algorithms or combinations of them. Taking into account this, in this work, firstly Feed-forward Neural Network (FNN) and Genetic Algorithm (GA) were selected, and then, these algorithms were combined to have the new algorithm. In order to evaluate its performance, the most popular circuits, such as ovals and tracks with different types of curves, were selected from auto racing, which were implemented in the Unity 2D video game engine. After, as a way to compare the performance of the FNN-GA algorithm, the Q-learning algorithm was selected, and the agents to learn with them were placed on said tracks. As a result, after executing 100 generations of the agents, per algorithm, it is found that taking into account the better use of CPU and RAM resources, Q-learning is the best, on the contrary, taking into account the quantity of in learning generations, is the best since it learns in fewer generations. On the other hand, both algorithms learn easier on ovals than on circuits with different types of curves.