Comparison of assistance systems for autonomous driving using Convolutional Neural Networks

Abstract

After building the neural network, hyperparameters tuning is an important step in Machine Learning to improve the model performance or to customize model hyperparameters to better suit the dataset. There are different tools and packages that use grid or random search algorithms for hyperparameters optimization. But these algorithms do not indicate the importance of different hyperparameter combinations or the correlation between hyperparameters and the loss function. Deep learning models consist of multiple layers with fully-connected individual neurons that makes it complicated to understand why the model learns it that way. That is why finding hyperparameters importance is necessary to define which factors have positive or negative impacts on the model. A deep learning model in this project will take images from the camera in the simulator as input and predict steering values. The aim of this work is to optimize the hyperparameters tuning process of CNN model. Instead of choosing and combining randomly, different sets of hyperparameters are selected systematically through multivariate quadratic regression.