Refine
Is part of the Bibliography
- no (1)
Has Fulltext
- yes (1)
Year of publication
- 2023 (1)
Document Type
- Master's Thesis (1)
Institute
Language
- English (1)
The motivation behind this thesis is to investigate the potential of machine learning technique, specifically neural networks and Reinforcement Learning (RL) that can by-self learn to improve vehicle handling. With the advancement in machine learning techniques, RL has gained interest in recent years in the field of vehicle controls due to its ability in effectively handling complex tasks. The All-Wheel Steering (AWS) has shown potential in enhancing the vehicle stability and maneuverability by individually steering each wheel. The primary goal of this thesis is to leverage Deep Reinforcement Learning (DRL) to further optimize the AWS control. The algorithm was trained and evaluated in a virtual environment by altering the vehicle control module imported into the CarMaker for Simulink interface. This algorithm observed vehicle states and steering input to learn best action which regulates the steering at rear axle in addition to front wheel steering. The results from this study demonstrated that the trained RL agent led to decrease in vehicle sideslip, average lateral acceleration on vehicle and deviation from path compared to an existing control method without requiring knowledge of the model. Two different reward function were considered to investigate the learning behavior. When a function was guided to minimize the lateral velocity, it was observed that the trained agent in training scenario resulted in almost 6 times lower average vehicle sideslip than with the rear wheel steer controller from a state-of-the-art model.