Safety exploration using Gaussian process classification for uncertain systems

In this paper, a novel method for identifying safe and unsafe regions of the system’s uncertain parameter space is proposed. For a given set of performance requirements, such estimation can be obtained by means of binary classification in which uncertain parameters are classified as either safe or unsafe in the sense that the given performance requirements are met or not. Hence, using Gaussian process classification it is possible to obtain (non-convex) safe and unsafe regions supported by minimum confidence levels of the corresponding estimations. We adopt active learning to update the Gaussian process classification model and to make more accurate predictions by selecting informative observations sequentially. The effectiveness of the proposed algorithm is demonstrated on various illustrative examples.