A Comprehensive Analysis of Methods to Write Requirements for Machine Learning Components used in Autonomous Vehicles

Machine learning components are widely used in autonomous vehicles for implementing functionalities related to perception and planning. To verify if the vehicle-level functionalities are as specified, one of the widely used approaches is requirements-based testing. However, writing testable requirements for machine learning components can be challenging since the machine learning outcomes are seldom known in advance. Nevertheless, it is important to have a specification that details the expected behavior from machine learning components. In this paper, we discuss different approaches to write a specification for machine learning algorithms that are used in autonomous vehicles. These approaches include natural language requirements, user stories, use case specifications, behavioral diagrams, data as requirements, and formal specification methods. We also propose a tabular specification method for specifying requirements of machine learning algorithms. We use a sample operational design domain (ODD) and system architecture to discuss the advantages and disadvantages of each of the techniques. We also discuss which approaches can aid with testing as well as error analysis of the model generated using the machine learning algorithms.