DARS 2017 is the 2nd international workshop in a series dedicated to the Design and Analysis of Robust Systems. Robustness refers to the ability of a system to behave reliably in the presence of perturbation in the system parameters or irregularities in the system's operating environment. This is particularly important in the context of embedded systems and software, which interact with a physical environment through sensors and actuators and communicate over wired or wireless networks. Such systems are routinely subject to deviations arising from sensor or actuation noise, quantization and sampling of data, uncertainty in the physical environment, and delays or packet drops over unreliable network channels. When deployed in safety critical applications, system robustness in the presence of uncertainty is not just desirable, but crucial.
The goal of DARS is to foster dialogue and exchange of ideas and techniques across several disciplines with an interest in robustness such as formal verification, programming languages, fault-tolerance, control theory and hybrid systems.
Domains of interest include, but are not limited to: reactive, timed, hybrid or
probabilistic programs/circuits/systems/networks, approximate computing and
fault tolerance of distributed systems.
All deadlines are AOE (Anywhere on Earth)
|Extended abstract submission
||May 5, 2017
||April 28, 2017
||May 25, 2017
||July 22, 2017
Robustness, Mori-Zwanzig Model Reduction, and Statistical Validation of Hybrid Systems.
Geir E. Dullerud, University of Illinois
Abstract: In this talk we first discuss different notions of robustness for hybrid systems, using concepts from control theory as motivation. We then introduce a new approach for classifying robustness in hybrid systems and present a method for statistical verification. In particular, we focus on a computational approach to verification of mathematical models that are formed when combining physics-based models with discrete-transition models, such as those which model software algorithms. Namely, the mathematical models that arise when for instance considering Cyberphysical Systems, or the Internet of Things. During the first part of the talk will concentrate on system verification, and will present a new verification algorithm for continuous-time stochastic hybrid systems, whose specifications are expressed in metric interval temporal logic (MITL), by deploying a novel model reduction method. By partitioning the state space of the hybrid system and computing the optimal transition rates between partitions, we provide a procedure to both reduce the system to a continuous-time Markov chain, and the associated specification formulas. We prove that the unreduced formulas hold (or do not) if the corresponding reduced formula on the Markov chain is robustly true (or false) under certain perturbations. In addition, a stochastic algorithm to complete the verification has been developed. We have extended the approach of this algorithm, and have developed a direct stochastic algorithm for probabilistically verifying a certain hybrid system class, and applied this technique to an extensive benchmark problem with realistic dynamics. The second part of the talk: In many game theory and filtering problems it is not possible to analytically obtain solutions for statistical properties of systems under study, and here we will describe our recent work on principle-based numerical approaches to obtaining estimates of these properties, and the application of the techniques developed in the verification framework above. Monte Carlo simulation of Markov processes allows the numerical estimation of their statistical properties from an ensemble of sample system paths. We present methods for generating reduced-variance path ensembles for the tau-leaping discrete-time simulation algorithm, which allows mean stochastic process dynamics to be estimated with substantially smaller ensemble sizes. Our methods are based on antithetic and stratified sampling of Poisson random variates, and we provide a combination of analytical proofs and numerical evidence for their performance, which can frequently be a 2-3 orders of magnitude improvement over standard Monte Carlo. Application examples will be discussed.
Bio: Geir E. Dullerud is the W. Grafton and Lillian B. Wilkins Professor in Mechanical Engineering at the University of Illinois at Urbana-Champaign. There he is also a member of the Coordinated Science Laboratory, where he is Director of the Decision and Control Laboratory (21 faculty); he is an Affiliate Professor of both Computer Science, and Electrical and Computer Engineering. He has held visiting positions in Electrical Engineering KTH, Stockholm (2013), and Aeronautics and Astronautics, Stanford University (2005-2006). Earlier he was on faculty in Applied Mathematics at the University of Waterloo (1996-1998), after being a Research Fellow at the California Institute of Technology (1994-1995), in the Control and Dynamical Systems Department. He has published two books: "A Course in Robust Control Theory", Texts in Applied Mathematics, Springer, 2000, and "Control of Uncertain Sampled-data Systems", Birkhauser 1996. His areas of current research interests include hybrid dynamical systems, convex optimization in control, cyber-physical system security, cooperative robotics and stochastic simulation. In 1999 he received the CAREER Award from the National Science Foundation, and in 2005 the Xerox Faculty Research Award at UIUC. He is a Fellow of both IEEE (2008) and ASME (2011).
Analyzing Neural Network Robustness with Constraints
Dimitrios Vytiniotis, Microsoft Research
Abstract: Despite having high accuracy, neural nets have been shown to be susceptible to adversarial examples, where a small perturbation to an input can cause it to become
misclassified. In this talk I will give some metrics for measuring the robustness of a neural net and present an algorithm for approximating these metrics based on an encoding of robustness as a linear program. I will describe how one implements an “abstract interpreter” specifically tailored to neural networks, and how to use ideas from formal verification to scale the constraint generation and solving as needed for modern deep learning architectures. We show that our algorithm generates more informative estimates of robustness compared to estimates based on existing algorithms in the deep learning community. Our techniques were used to additionally marginally improve neural net robustness both according to the metrics that we propose, but also according to previously proposed metrics, although effective solutions to improving robustness remain elusive.
Bio: Dimitrios Vytiniotis is a researcher at Microsoft Research, working on programming languages. He holds a PhD from the University of Pennsylvania (2008). Dimitrios has worked on type inference and type systems, functional programming, domain-specific languages, and memory management. Some of his work is on problems in the intersection of programming languages and machine learning: using formal method techniques to assess the robustness of neural networks, and more recently, asynchronous dataflow for training of dynamic neural networks.
A CEGAR approach for stability verification of linear hybrid systems
Miriam García Soto.
Abstract: This document summarizes results related to an algorithmic approach
for stability analysis of linear hybrid systems.
Classical approaches rely on Lyapunov function search and suffer
from numerical issues. In addition, an unsuccessful template for the
Lyapunov function does not provide insights on the choice of a better
To overcome these issues, we present a counterexample guided-abstraction
refinement (\cegar) approach which iteratively searches for a stability
certification over an abstract system, and provides insights to obtain
more accurate abstract systems if needed.
Robust Linear Temporal Logic.
Paulo Tabuada and Daniel Neider.
Abstract: Although it is widely accepted that every system should be robust, in the sense that "small" violations of environment assumptions should lead to "small" violations of system guarantees, it is less clear how to make this intuitive notion of robustness mathematically precise. In our work, we address the problem of how to specify robustness in temporal logic. Our solution is a robust version of Linear Temporal Logic, which we term Robust Linear Temporal Logic (rLTL). Formulas in this logic are syntactically identical to LTL formulas but are endowed with a many-valued semantics that encodes robustness. In particular, the semantics of the rLTL formula \phi implies \psi is such that a "small" violation of the environment assumption \phi is guaranteed to only produce a "small" violation of the system guarantee \psi. In addition to the definition of rLTL, we study the corresponding verification and synthesis problems. Similarly to LTL, we show that both problems are decidable, with the verification problem being solvable in exponential time and the synthesis problem being solvable in doubly exponential time.
Stabilizing Numeric Programs against Platform Uncertainties.
Yijia Gu and Thomas Wahl.
Abstract: Floating-point arithmetic (FPA) is a loosely standardized approximation of real arithmetic available on many computers today. The use of approximation incurs commonly underestimated risks for the reliability of numeric software, including reproducibility issues caused by the relatively large degree of freedom for FPA implementers offered by the IEEE 754 standard. If left untreated, such problems can seriously interfere with program portability.
In this paper we discuss numeric programs' lack of robustness against platform variations, including irreproducible control flow and invariants that hold on some platforms but not others. We also demonstrate how, using information on the provenance of platform dependencies, such reproducibility violations can be repaired, which results in stabilized program execution. We illustrate the use of this technique on decision-making and other numeric programs, and present an outlook to its applicability to solving reproducibility issues among CPU and GPU versions
of kernel support vector machines.
Based in part on work published at EUROPAR 2015 and VMCAI 2017.
Martingales for Probabilistic Termination and Safety.
Krishnendu Chatterjee, Petr Novotný and Djordje Zikelic.
Abstract: We address the problem of automated proofs of reachability and safety properties for linear-arithmetic probabilistic programs with nondeterminism. We define the notion of stochastic invariants, which are constraints along with a probability bound that the constraints hold. We introduce a concept of repulsing supermartingales. First, we show that repulsing supermartingales can be used to obtain bounds on the probability of the stochastic invariants. Second, we show the effectiveness of repulsing supermartingales in the following three ways: (1)~With a combination of ranking and repulsing supermartingales we can compute lower bounds on the probability of termination; (2)~repulsing supermartingales provide witnesses for refutation of almost-sure termination; and (3)~with a combination of ranking and repulsing supermartingales we can establish persistence properties of probabilistic programs.
We also present results on related computational problems and an experimental evaluation of our approach on academic examples.
Recent Advances in Designing Robust Probabilistic Systems.
Radu Calinescu, Milan Ceska, Simos Gerasimou, Marta Kwiatkowska and Nicola Paoletti.
Abstract: Robust systems are of great interest because they can withstand changes in the system parameters and
do not expose users to large variations in quality attributes. This paper presents our recent results on automating the design of robust probabilistic systems. We integrate search-based design strategies with formal analysis of parametric Markov models, and introduce sensitivity-aware synthesis of Pareto-optimal sets of designs providing useful tradeoffs between quality attributes and sensitivity. The synthesis algorithms are implemented in our publicly available tool RODES and evaluated on several case studies from different domains. The experiments demonstrate that our approach provides unique insights into how the parameters affect the system dynamics and robustness.