Trusted Machine Learning for Probabilistic Models

In several mission-critical domains (e.g., selfdriving cars, cybersecurity, robotics) where machine learning algorithms are being used heavily, it is becoming increasingly important to ensure that the learned models satisfy some domain properties (e.g., temporal constraints). Towards this goal, we propose Trusted Machine Learning (TML), wherein we combine the strengths of machine learning and model checking. If the desired logical properties are not satisfied by a trained model, we modify either the model (‘model repair’) or the data from which the model is learned (‘data repair’). We outline a concrete case study based on the Markov Chain model of a car controller for ‘lane changing’ — we demonstrate how we can ensure that such a model, learned from data, satisfies properties specified in Probabilistic Computation Tree Logic (PCTL).