Lab Conditions for Research on Explainable Automated Decisions

Abstract

Artificial neural networks are being proposed for automated decision making under uncertainty in many visionary contexts, including high-stake tasks such as navigating autonomous cars through dense traffic. Against this background, it is imperative that the decision making entities meet central societal desiderata regarding dependability, perspicuity, explainability, and robustness. Decision making problems under uncertainty are typically captured formally as variations of Markov decision processes (MDPs). This paper discusses a set of natural and easy-to-control abstractions, based on the Racetrack benchmarks and extensions thereof, that altogether connect the autonomous driving challenge to the modelling world of MDPs. This is then used to study the dependability and robustness of NN-based decision entities, which in turn are based on state-of-the-art NN learning techniques. We argue that this approach can be regarded as providing laboratory conditions for a systematic, structured and extensible comparative analysis of NN behavior, of NN learning performance, as well as of NN verification and analysis techniques.