Hyperparameter Optimization of Opaque Models for Autonomous Vehicle Algorithms

Abstract

Algorithms usually consist of many hyperparameters that need to be tuned to perform efficiently. It may be possible to tune a handful of parameters manually for simple algorithms however as the algorithm becomes more complex the number of hyper- parameters also increases which makes finding the optimal hyperparameters more difficult. As a result, automating the parameter tuning would be of great interest in many different applications by reducing manual labor while increasing the perfor- mance of the algorithm. In this research, we focused on automating the process of hyperparameter selection for any opaque model to enable fully automated learning. We surveyed different hyperparameter optimization algorithms, selected the most effi- cient ones in different scenarios, and developed a framework that can be easily utilized by different users. We tested our algorithm and framework on NVIDIA’s localization algorithm developed for Autonomous Vehicles. Additionally, we performed hyperpa- rameter optimization on different regression algorithms on the abalone [7] dataset to have another thorough comparison of the different optimization algorithm.