Abstract
The development of autonomous driving technology necessitates sophisticated simulation systems that can accurately model human driving behavior under various conditions. This thesis explores the effectiveness of a highway driving simulator in inducingdifferent cognitive workloads in users and evaluates the performance of a Reinforcement
Learning (RL) model in emulating human driving behavior. The study is guided by two
primary hypotheses: (H1) the simulator can be utilized to induce different workloads,
as measured by NASA-TLX ratings and driving performance, and (H2) the RL model
can emulate human driving behavior regarding lane-keeping accuracy.
The highway driving simulator was developed using Unity and featured scenarios
with varying levels of traffic density and cognitive load. Human drivers’ performance was
measured and compared with that of an RL model trained using the same scenarios. The
findings supported H1, demonstrating that the simulator effectively induced different
cognitive workloads, as evidenced by significant variations in NASA-TLX ratings and
driving performance metrics across scenarios. However, H2 was only partially supported.
While the RL model performed comparably to human drivers in the first scenario, it
couldn’t learn to drive in more complex environments.
The study highlights the simulator’s potential as a tool for testing and training both
human drivers and RL models. It also underscores the need for further improvements
of the RL model to handle the complexities of the harder scenarios. Future research
should focus on expanding the simulation to include more varied driving environments,
improving traffic realism, enhancing control system feedback, and introducing additional
tasks to increase scenario complexity. Incorporating computational rational cognitive
models could further enhance the human-likeness of the RL model.
| Date of Award | 2024 |
|---|---|
| Original language | English (American) |
| Supervisor | Philipp Wintersberger (Supervisor) |