Deep Behavioral Replication of Markov Models for Autonomous Cars using Neural Networks

With autonomous vehicles on the brink of a revolution, there is an increased need for reliable individual driving functions and the overarching vehicle. Classical assessment techniques for such systems, which assume that future states depend only on the current state, include Failure Modes, Effects and Diagnostic Analysis (FMEDA), or classic Markov models, also recommended by IEC 61508, ISO 26262, and SOTIF ISO 21448. More realistic are memoryless approaches like the Monte Carlo simulation of Markov chains. However, these are computationally expensive. As an in-the-loop component to assess the safety of autonomous vehicles, Markov models essentially become the bottlenecks of the toolchain. In this context, trained neural networks are excellent tools to replicate the behavior of the Markov models rendering them an excellent in-the-loop component. To this end, the current work demonstrates that the deep learning models are capable of learning and generalizing the behavior of the Markov models.

Keywords: Autonomous driving, Deep learning, Markov chain, Safety of the Intended Functionality (SOTIF), Failure probability, Operational failure.

Download PDF