The problem
Mining Signal Temporal Logic requirements from a dataset of regular (good) and anomalous (bad) trajectories of a dynamical system. The training set is labelled by human experts, and only a limited, typically noisy, amount of data is available.
The approach
A systematic, two-step procedure that synthesises both the structure and the parameters of the formula:
- a novel evolutionary algorithm learns the structure of the formula;
- parameter synthesis works on a statistical emulation of the average robustness of a candidate formula with respect to its parameters.
Results
The method is compared with previous work and with a recent decision-tree based method, on two case studies, including anomalous trajectory detection in naval surveillance.