Context: Future automotive customer functions will be highly personalisable and adapt their settings proactively in an intelligent way. Aim: We aim at designing generic architectural patterns for functional architectures containing machine learning components. Method: We first formalise a new architectural model. Based on this model, we present and discuss three alternative architectural patterns: (1) concurrent learning, (2) end-to-end learning, and (3) user shadow learning. For these patterns, three alternative integration approaches are discussed: (i) centralised holistic approach, (ii) domain-specific approach, and (iii) dedicated approach. Moreover, we conduct an evaluation using real car data for different customer functions. Conclusion: We propose the use of the user shadow learning pattern in the dynamic architectural model. The user shadow learning pattern is not affected by safety constraints, as is usually the case for integrating artificial intelligence, as it only models user behaviour while leaving the original function intact. To integrate the multitude of models, we propose a domain-specific approach. This approach provides a balance between the trade-offs in the dedicated approach and the holistic approach, being high computational overhead and design complexity, respectively.

Context: As we move towards higher levels of automation in autonomous driving, we see an increase in functionality that either assists or takes over in both normal and emergency scenarios. These new functionalities can be intentionally switched off by the user, but can also be deactivated unintentionally by (i) accident, (ii) a software malfunction, (iii) a hardware defect, or (iv) an intrusion. Aim: In addition to already applied methods at design time, we aim to recognise mistimed and/or unintended deactivation of vehicle functions, in particular, driver assistance functions (ADAS), at run-time. Upon recognition of the occurrence, we propose to inform the user and the original equipment manufacturer (OEM) in order to improve both the future and the current system behaviour, to support development processes, and to support already conducted safety measures. Method: Based on a feature subset, selected by streaming feature selection, we learn the nominal behaviour of the driver in the interaction with ADAS functions in order to find deviations. The approach considers the technical challenges of automotive E/E architectures and is optimised to reduce communication and computational complexity. We evaluate this approach with recorded real car data from customers participating in a field study. Results: Based on eight datasets, we traced a total of 17 state-of-the-art ADAS functions per participant, yielding to a total of 136 runs. We observed that during 24 among them, the user deactivated the functions at least once for more than a few seconds. For 13 of these 24 runs, we were able to detect and flag possible non-nominal behaviour. Conclusion: As at least one participant configured a convincingly large number of ADAS functions, we need a dynamic system to monitor the configuration of these functions actively. Our approach was capable of detecting potential non-nominal behaviour in up to 52% (13/24) out of these reconfigurations. This result is promising and will receive further attention in future work.

More and more software-intensive systems employ machine learning and runtime optimization to improve their functionality by providing advanced features (e. g. personal driving assistants or recommendation engines). Such systems incorporate a number of smart software functions (SSFs) which gradually learn and adapt to the users' preferences. A key property of SSFs is their ability to learn based on data resulting from the interaction with the user (implicit and explicit feedback)-which we call trainability. Newly developed and enhanced features in a SSF must be evaluated based on their effect on the trainability of the system. Despite recent approaches for continuous deployment of machine learning systems, trainability evaluation is not yet part of continuous integration and deployment (CID) pipelines. In this paper, we describe the different facets of trainability for the development of SSFs. We also present our approach for automated trainability evaluation within an automotive CID framework which proposes to use automated quality gates for the continuous evaluation of machine learning models. The results from our indicative evaluation based on real data from eight BMW cars highlight the importance of continuous and rigorous trainability evaluation in the development of SSFs.