We present a Symmetry-based abstraction refinement algorithm SymAR that is directed towards safety verification of large-scale scenarios with complex dynamical systems. The abstraction maps modes with symmetric dynamics to a single abstract mode and refinements recursively split the modes when safety checks fail. We show how symmetry abstractions can be applied effectively to closed-loop control systems, including non-symmetric deep neural network (DNN) controllers. For such controllers, we transform their inputs and outputs to enforce symmetry and make the closed loop system amenable for abstraction. We implemented SymAR in Python and used it to verify paths with 100s of segments in 2D and 3D scenarios followed by a six dimensional DNN-controlled quadrotor, and also a ground vehicle. Our experiments show significant savings, up to 10x in some cases, in verification time over existing methods.
翻译:我们提出了一个基于对称的抽象精细算法Symar,该算法针对的是以复杂的动态系统对大型假设情景进行安全核查。抽象地图模式,对称动态为单一抽象模式,当安全检查失败时,对称性图解会回溯式分割模式。我们展示了如何有效地将对称抽象模型应用到闭环控制系统,包括非对称深神经网络控制器。对于这些控制器,我们改造其输入和输出,以强制对称并使闭环系统适合抽象化。我们在Python实施了Symar,并用Symar在2D和3D情景中以100个区段进行校验,然后是六维DNN控制的孔控象管和地面飞行器。我们的实验显示,在有些情况下,在对现有方法进行核查的时间里,可以节省10倍以上。