Toward V&V of neural network based controllers

Abstract

Online adaptation is a powerful means to handle unexpected slow or catastrophic changes of the system's behavior (e.g., a stuck or broken rudder of an aircraft). Therefore, adaptation is one way for realizing a self-healing system. Substantial research and development has been made to use neural networks (NN) for such tasks (e.g., integrated in various unmanned helicopters and test-flown on a modified F-15 aircraft). Despite the advantages of adaptive neural network based systems, the lack of methods to perform certification, verification, and validation (V&V) of such systems severely restricts their applicability.In this paper, we report on ongoing work to develop V&V techniques and processes for NN-based safety-critical control systems, in our case an aircraft flight control system. Although the project ultimately aims at V&V of online adaptive systems, this paper focuses on the first part of this project dealing with so-called pre-trained neural networks (PTNN). V&V techniques developed here are important pre-requisites for handling the online adaptive case. In particular, we describe highlights of a process guide which has been developed within this project and discuss important V&V issues which need to be addressed during certification.