Genetic Optimization for Self-Driving Vehicles Based on Automated Behavior Cloning Convolutional Neural Network

Abstract

Recently, autonomous driving technology has learned to drive safely and smoothly. Nonetheless, using convolutional neural networks (CNNs) has significantly influenced designs of autonomous driving technology. Most current designs of CNN-based autonomous driving technology are built manually by specialists in both CNNs and the investigated topics. Thus, finding the optimum CNN designs for learning safe driving behavior is complex. This study uses genetic algorithms to construct automated genetic behavior cloning (AGBC). The "Automated" features of the proposed algorithm require little knowledge of CNNs. Using a front-facing camera and an experienced driver's steering directions, researchers can acquire a strong CNNs architecture for learning safe driving behavior. Furthermore, it can be used to teach AGBC-CNNs to emulate human driving behavior. We compared our proposed technique AGBC-CNN with four manually adjustment behavior cloning CNNs (MABC-CNNs) and one automated genetic manually adjustment behavior cloning CNNs (AGMABC-CNNs) to validate its effectiveness and robustness. Studies show that AGBC-CNN outperforms existing architecture design techniques (MABC-CNNs and AGMABC-CNN) in terms of classification accuracy.