Fast Koopman Surrogate Falsification Using Linear Relaxations and Weights

Abstract

Recent work demonstrated that using Koopman surrogate models to falsify black-box models against signal temporal logic specifications is highly effective. However, the bottleneck of this approach arises from the mixed-integer linear program optimization used to synthesize the falsifying trajectory. The complexity of mixed-integer linear programming can be prohibitive, increasing exponentially with the number of binary variables. In this work, we introduce a new weighted robustness encoding that eliminates the need for binary variables. We also propose a new weighting scheme for Koopman operator linearization that aims to compensate for inaccuracies in the learned model. We evaluate our approach using a set of benchmarks from the ARCH falsification competition. Our weighting methods significantly improve computational efficiency and reduce the number of simulations needed to find falsifying traces.

Publication
22nd International Symposium on Automated Technology for Verification and Analysis (ATVA 2024)