2021/01/05 by Michael Everett, Everett, Michael, Golnaz Habibi +3 · 4 citations
Computer Science · Engineering · #Adversarial Robustness in Machine Learning #FOS: Computer and information sciences #FOS: Electrical engineering #Fault Detection and Control Systems #Machine Learning (cs.LG) #Machine Learning and Algorithms #Robotics (cs.RO) #Systems and Control (eess.SY) #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.2101.01815
openalex publication_date 2021/01/05 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Neural Networks (NNs) can provide major empirical performance improvements\nfor robotic systems, but they also introduce challenges in formally analyzing\nthose systems' safety properties. In particular, this work focuses on\nestimating the forward reachable set of closed-loop systems with NN\ncontrollers. Recent work provides bounds on these reachable sets, yet the\ncomputationally efficient approaches provide overly conservative bounds (thus\ncannot be used to verify useful properties), whereas tighter methods are too\nintensive for online computation. This work bridges the gap by formulating a\nconvex optimization problem for reachability analysis for closed-loop systems\nwith NN controllers. While the solutions are less tight than prior semidefinite\nprogram-based methods, they are substantially faster to compute, and some of\nthe available computation time can be used to refine the bounds through input\nset partitioning, which more than overcomes the tightness gap. The proposed\nframework further considers systems with measurement and process noise, thus\nbeing applicable to realistic systems with uncertainty. Finally, numerical\ncomparisons show 10\× reduction in conservatism in \(1)/(2) of the\ncomputation time compared to the state-of-the-art, and the ability to handle\nvarious sources of uncertainty is highlighted on a quadrotor model.\n