@misc{10481/107646,
year = {2025},
url = {https://hdl.handle.net/10481/107646},
abstract = {The mounting demand for autonomous steering systems that are both secure and
efficient reveals a critical gap in the precise prediction of steering angles, a prerequisite
for ensuring safe and effective navigation. The prevalent challenges in current
systems, namely their inability to generalize across diverse driving conditions,
underscore the necessity for advanced methodologies. This thesis proposes a novel
approach to predicting steering angles. The approach involves training and optimizing
a Convolutional Neural Network (CNN) model using a combination of advanced
optimization techniques and GAs and leveraging powerful, diverse datasets. Two
datasets were utilized in this research. The first dataset, designated as the ”original”
dataset, was collected during a 25-minute driving session on a simulated track featuring
diverse curves and terrain types. This initial training set comprised 9,428 images,
which served as the foundation for the subsequent analyses. To confront limitations
in diversity and size, a second expanded dataset was developed by incorporating a
wider range of simulator driving scenarios. This increased the total size to 45,282
images. To address challenges such as misalignment and steering angle imbalance,
preprocessing steps were applied, including image cropping, zooming, and marker
balancing. The methodology entailed the evaluation of several optimization techniques,
namely Nadam, Adam, RMSprop, Adamax, SGD, Adagrad, Adadelta, and
Ftrl. These techniques were utilized in conjunction with the Python programming
language, Google Colaboratory, and the NVIDIA convolutional architecture. Genetic
algorithms (GAs) were employed to refine model weights by leveraging genetic operators
such as selection, mutation, and replacement. The experimental findings suggest
that GAs demonstrated superior performance in minimizing the mean squared error
(MSE) when compared to conventional optimizers. However, the efficacy of GAs was
found to be contingent upon the precise configuration of the genetic operators employed.
The integration of GA optimization with conventional optimizers within the
final models resulted in enhanced accuracy and robustness. In summary, the findings
underscore the effectiveness of genetic algorithms as a transformative tool for model
weight optimization, advancing the development of precise and reliable autonomous
steering systems.},
abstract = {La creciente demanda de sistemas de direccion autonomos que sean seguros y eficientes revela una brecha critica en la prediccion precisa de los angulos de direccion, un requisito previo para garantizar una navegacion segura y eficaz. Los desafios prevalentes en los sistemas actuales, a saber, su incapacidad para generalizar en diversas condiciones de conduccion, subrayan la necesidad de metodologias avanzadas. Esta tesis propone un enfoque novedoso para predecir los angulos de direccion. El enfoque implica entrenar y optimizar un modelo de Red Neuronal Convolucional (CNN) utilizando una combinacion de tecnicas de optimizacion avanzadas y Algoritmos Geneticos (GA) y aprovechando poderosos y diversos conjuntos de datos. Se utilizaron dos conjuntos de datos en esta investigacion. El primer conjunto de datos, denominado "original", se recopilo durante una sesion de conduccion de 25 minutos en una pista simulada con diversas curvas y tipos de terreno. Este conjunto inicial de entrenamiento comprendia 9,428 imagenes, que sirvieron como base para los analisis posteriores. Para enfrentar las limitaciones de diversidad y tamano, se desarrollo un segundo conjunto de datos ampliado incorporando una gama mas amplia de escenarios de conduccion en simulador. Esto aumento el tamano total a 45,282 imagenes. Para abordar problemas como el desajuste y el desequilibrio en los angulos de direccion, se aplicaron pasos de preprocesamiento, que incluyeron recorte de imagenes, zoom y equilibrado de marcadores. La metodologia consistio en la evaluacion de varias tecnicas de optimizacion, a saber, Nadam, Adam, RMSprop, Adamax, SGD, Adagrad, Adadelta y Ftrl. Estas tecnicas se utilizaron en conjunto con el lenguaje de programacion Python, Google Colaboratory y la arquitectura convolucional de NVIDIA. Se emplearon GAs para refinar los pesos del modelo mediante operadores geneticos como seleccion, mutacion y reemplazo. Los resultados experimentales sugieren que los GA demostraron un rendimiento superior en la minimizacion del error cuadratico medio (MSE) en comparacion con los optimizadores convencionales. Sin embargo, la eficacia de los GA se encontro dependiente de la configuracion precisa de los operadores geneticos empleados. La integracion de la optimizacion GA con optimizadores convencionales dentro de los modelos finales resulto en una mayor precision y robustez. En resumen, los hallazgos subrayan la efectividad de los algoritmos geneticos como una herramienta transformadora para la optimizacion de pesos de modelos, avanzando en el desarrollo de sistemas de direccion autonomos precisos y confiables.},
organization = {Tesis Univ. Granada.},
publisher = {Universidad de Granada},
title = {Designing an Autonomous Driving System Based on Data-Extracted Behavior and Evolutionary Optimization},
author = {Khawaldeh, Bashar},
}