[论文] Hyperlyve: Parallelized Enumeration and Exact Lyapunov Stability Certi...
论文概要
研究领域: ML
作者: Brian Yarbrough, Jesse Wayment, Jingbo Wang, Philip E. Paré, Shreyas Sundaram
发布时间: 2026-11-16
arXiv: 2611.21599
中文摘要
我们引入 Hyperlyve(超平面分割 Lyapunov 验证与枚举),一个用于证明非线性动力系统渐近稳定性的框架,包括具有神经网络控制器的系统。Hyperlyve 在反例引导归纳合成(CEGIS)循环中学习并形式化验证连续分段仿射(CPA)神经 Lyapunov 函数。Hyperlyve 的关键创新是一种并行算法,它枚举深度网络 ReLU 激活形成的超平面排列,并验证 Lyapunov 条件在每个多面体区域内成立。我们推导了可靠且完备的区域测试,用以验证紧凑域内精确平滑动力学(而非分段线性近似)的渐近稳定性。此外,我们将 CPA 神经 Lyapunov 验证问题的计算复杂度形式化表征为 co-NP 完全,填补了文献中的关键空白。实证结果表明,Hyperlyve 枚举 ReLU 网络的速度比竞争工具快 150 倍以上,这使得 Hyperlyve 能够证明当前最先进验证器无法处理的系统的渐近稳定性。
原文摘要
We introduce Hyperlyve (Hyperplane Partitioned Lyapunov Verification and Enumeration), a framework for proving the asymptotic stability of nonlinear dynamical systems, including those with neural network controllers. Hyperlyve learns and formally certifies continuous piecewise-affine (CPA) neural Lyapunov functions within a counterexample-guided inductive synthesis (CEGIS) loop. Hyperlyve's key innovation is a parallel algorithm that enumerates the hyperplane arrangement formed by the ReLU activations of deep networks and certifies that Lyapunov conditions hold within each polyhedral region. We derive sound and complete regional tests to verify the asymptotic stability of the exact smooth dynamics (rather than a piecewise-linear approximation) within a compact domain of interest. Furthermo...
自动采集于 2026-11-18
#论文 #arXiv #ML #小凯