Training Genetic Neural Networks Algorithms for Autonomous Cars with the LAOP Platform

摘要

The challenge with self-driving cars is to create a model that converts sensors data (such as cameras or proximity sensors) into actions. This way the car can react to its changing environment and make the right decisions. In the literature, Neural Networks is the most promising technique used to parse these sensors data. A well trained and designed neural network can take the sensors values and output the right actions. In this paper, we introduce a Way to train Efficiently Neural Networks with Genetic principles, called WENNG. Moreover, we propose a comparative study between all the variations of WENNG to highlight the best-performing ones. To evaluate our WENNG training variation, we implement two well known neural network algorithms: the FullyConnected one and the NEAT algorithm. Through extensive simulations, we demonstrate that the Natural Selection WENNG outperforms the Greedy WENNG at training the genetic neural networks with a low mutation rate. Finally, we show that an IMproved version of NEAT called IMNEAT, minimizes twice the number of generations to reach the maximum fitness value compared to the traditional NEAT algorithm.