Assurance-based Learning-enabled Cyber-Physical Systems (ALC)

    Agency: DARPA
    PI: Gabor Karsai
    Co-PIs: Xenofon Koutsoukos, Taylor Johnson, Janos Sztipanovits, Abhishek Dubey
    Funding: $7.4M
    PoP: 4/2018-4/2022
    ALC Toolchain architecture

    Autonomous vehicles (cars, drones, underwater vehicles, etc.) have started using software components that are built using machine learning techniques. This is due to the fact that these vehicles must operate in highly uncertain environments and that we cannot design a correct algorithm for all possible situations. Instead, we collect data from a real or simulated environment and train a general purpose system - typically a neural net - to perform a certain function using machine learning techniques. But the challenge is that the training data cannot cover all possible cases, yet we need to know that the system works safely and has acceptable performance. Our project is doing fundamental research and building tools for supporting the engineering of such system. The tools are for modeling the system (e.g. an underwater vehicle), executing the training and testing of the learning-based components, and building formal arguments (called assurance cases) to show that the system is safe.