Statistical Optimization of CNN-LSTM Network Architectures: A Case Study in Autonomous Vehicle Control

Abstract

This thesis introduces a novel framework for optimizing combined Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) architectures for kinematic control problems, with a specific focus on autonomous vehicle control, chosen for its combination of dynamics and scene recognition which bare similarities to other complex controls problems faced in mechatronics engineering.

A combined dataset and high-fidelity simulation environment is implemented using an an off-the-shelf game engine, a novel approach in the literature which is traditionally limited by the quality of openly available simulation environments, and enabling a hybrid approach of training neural network models via both Imitation Learning and Reinforcement Learning.

A comprehensive exploration of network structures and hyperparameters is undertaken using the Tree-structured Parzen Estimator (TPE) to systematically improve model performance, enabling more informed approaches to neural network structure and design. The research demonstrates the impact of varying temporal and spatial information through varying the emphasis on the CNN and LSTM layers of the network respectively, as well as the amount of context provided to the network.

The findings and methodology are adaptable to other problems in the kinematic optimization control space, and the particular similarities of other problems in the area are discussed.