In this thesis, our objective is to investigate methodologies for predicting the parameters
of a PID-controlled cruise control system. Specifically, we delve into an
examination and comparison of three distinct approaches; one involving numerical
solutions, the second utilizing deep neural networks (DNN), and finally a combination
of the two.
We present several possible approaches for modeling the system as an ordinary
differential equation (ODE), a prerequisite for utilizing numerical solutions.
We outline the challenges of this process, as well as address issues unique to the
explored system. We also address the critical concern of selecting an appropriate
ODE solver, in addition to exploring optimization methods thereof. Additionally,
we investigate hyper-parameter optimization of neural networks using Weight and
Biases and delve into approaches for identifying and mitigating issues related to
model saturation and overfitting.
The ultimate goal is to compare these different methods. This comparison
encompasses an assessment of prediction accuracy, computation speed, and an
analysis of the respective strengths and weaknesses of the different approaches,
with the intent of finding the one best suited for our task.