Design of an Artificial Intelligence Model Certification System for Untrained Operational Environments
Abstract
The accuracy of Deep learning-based models for traffic state estimation (TSE), has been severely hampered by problems related to deficiencies in data availability, data collected during inclement weather conditions and the presence of noised data. Several scholars have asserted that certification of Machine Learning (ML) models using physics laws is an essential component in safety critical applications, however none of these studies have specifically examined the certification of Deep-Learning based models with physics laws to advance TSE. To expedite the application of Deep Learning models, it is critical to know whether a pre-trained AI model can be used in an unobserved operational environment with little or even no new data. However, it is often difficult to understand the black-box models learned by Deep Learning techniques from data. Considering that scientific knowledge is available for many engineering problems, this paper proposes a science-based certification methodology to sanity check whether the pre-trained data driven models can be used in untrained operational environments. This research demonstrates the benefit of certification of Deep Learning based models built using a small training synthetic dataset and certified by Lighthill-Whitham-Richards (LWR) law of traffic physics, depicted using the fundamental Greenshields’ diagram. This study certifies whether a TSE model trained in different traffic state conditions can be employed to predict a new environment that the model has not been trained on. The study also aims at improving interpretability and enhancing reliability of Deep Learning Models all being governed by the conservation law of traffic.
Subject Area
Computer Engineering|Information science|Information Technology|Artificial intelligence
Recommended Citation
Daryl Mupupuni,
"Design of an Artificial Intelligence Model Certification System for Untrained Operational Environments"
(2023).
ETD Collection for Tennessee State University.
Paper AAI30691221.
https://digitalscholarship.tnstate.edu/dissertations/AAI30691221