ProjectSMILE - Safety analysis and verification/validation of MachIne LEarning based systems

SMILE - Safety analysis and verification/validation of MachIne LEarning based systems

The SMILE project has explored challenges while introducing machine learning-based systems in automated driving (AD) applications. It also explored strategies to cope with those challenges to guide the industry and thus, be able to realize the potential to apply machine learning in safety critical systems. 

The project elaborated on the following research questions:

What available methods are there that can guarantee safe use of machine learning algorithms and trained neural networks within safety critical vehicular systems?

Within which safety critical vehicular functions is there a need for machine learning? What are the safety requirements for these systems and which requirements need to be allocated to elements based on machine-learning?

What are the major barriers for introduction of deep machine learning in safety critical systems? What are the available concepts/strategies to improve the level of safety integrity achievable in deep machine learning based systems? Which of these concepts/strategies are worthwhile to develop and evaluate further?

A literature review study based on (a) the initial research questions and (b) discussions with the industrial domain experts within safety critical systems within automation was the first part of the project. The second part concerned a workshop series. Six workshops have been arranged to (i) present the findings from the literature study (ii) give inspirational talks to the project members from researchers outside the project group (iii) cross fertilizing with other research projects, ESPLANADE, (iv) highlight the industrial needs from the project partners, (v) discussions about the industrial needs and their relation to state-of-art, and (vi) concluding the findings from the project and formulate the continuation, SMILE II, project. 

A new research agenda is created based on the results from our research. Three main areas were identified: data, models and verification and validation. Data concerns collection and management approaches for training to make robust DNN components. Models is related to data and refers to methods of incorporating data from multiple sources and contexts into the vehicle control systems. The area verification and validation concerns defining test cases, verification/validation environments and procedures, accuracy and safety targets, and other topics related to creating standards for developing DNN-based software used in safety-critical systems.