Failure Safety Analysis of Artificial Intelligence Systems for Smart/Autonomous Vehicles
Faranak Pourdanesh, Dinh Quang Truong, Fulvio Tagliabò, Phill Whiffin · 2021
Up to now failures in artificial intelligence systems, specifically machine learning algorithms which are their software components, are considered as systematic failures. The goal of this paper is to introduce a new concept of quantitative failure analysis for machine learning algorithms which can be used in smart/autonomous vehicles to guarantee sufficiently low risk of residual errors in this application. Firstly, a coincidence in evaluating impacts of unpredictable behaviours of machine learning algorithms and hardware components is introduced in order to statistically estimate failure rate based on a given number of data points. Next, a metric utilising this randomic failure rate is proposed to assess functional safety of smart and/or autonomous vehicles and evaluate their safeness according to ISO 26262:2018, and ISO/PAS 21448.