ARCH-COMP21 Category Report: Artificial Intelligence and Neural Network Control Systems (AINNCS) for Continuous and Hybrid Systems Plants

Johnson, Taylor T.; Lopez, Diego Manzanas; Benet, Luis; Forets, Marcelo; Guadalupe, Sebastián; Schilling, Christian; Ivanov, Radoslav; Carpenter, Taylor J.; Weimer, James; Lee, Insup

doi:10.29007/kfk9

Cited by 16 publications

(15 citation statements)

References 23 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We consider the benchmark problems used in the competition on NNCS at ARCH-COMP 2021 (Johnson et al 2021). In total there are seven problems with various features.…”

Section: Discussionmentioning

confidence: 99%

“…We also compare to the result of NNV (Tran et al 2020b), which uses a black-box reachability method for the plant and hence loses many dependencies. In (Johnson et al 2021) the authors reported that their tool runs out of memory before the analysis finishes. The plot in Figure 5 contains intermediate results, which show that the precision declines quickly.…”

Section: Case Study: Unicycle Modelmentioning

confidence: 99%

“…The original specifications are not satisfied by the modified controllers and hence we only compare reachability results and run time here for these problems. Our implementation can solve all benchmark problems with the original controllers, as shown in the ARCH-COMP 2021 report (Johnson et al 2021).…”

Section: Other Problems From Arch-compmentioning

confidence: 99%

“…While our approach is conceptually simple, we demonstrate in our evaluation that it is effective and scalable in practice. We successfully analyzed all benchmark problems from an annual NNCS competition (Johnson et al 2021) for the first time.…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Verification of Neural-Network Control Systems by Integrating Taylor Models and Zonotopes

Schilling,

Forets,

Guadalupe

2021

Preprint

Self Cite

View full text Add to dashboard Cite

We study the verification problem for closed-loop dynamical systems with neural-network controllers (NNCS). This problem is commonly reduced to computing the set of reachable states. When considering dynamical systems and neural networks in isolation, there exist precise approaches for that task based on set representations respectively called Taylor models and zonotopes. However, the combination of these approaches to NNCS is non-trivial because, when converting between the set representations, dependency information gets lost in each control cycle and the accumulated approximation error quickly renders the result useless. We present an algorithm to chain approaches based on Taylor models and zonotopes, yielding a precise reachability algorithm for NNCS. Because the algorithm only acts at the interface of the isolated approaches, it is applicable to general dynamical systems and neural networks and can benefit from future advances in these areas. Our implementation delivers stateof-the-art performance and is the first to successfully analyze all benchmark problems of an annual reachability competition for NNCS.

show abstract

“…We consider the benchmark problems used in the competition on NNCS at ARCH-COMP 2021 (Johnson et al 2021). In total there are seven problems with various features.…”

Section: Discussionmentioning

confidence: 99%

Section: Case Study: Unicycle Modelmentioning

confidence: 99%

Section: Other Problems From Arch-compmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Verification of Neural-Network Control Systems by Integrating Taylor Models and Zonotopes

Schilling,

Forets,

Guadalupe

2021

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…Verification of the Closed-loop ACAS Xu System. Existing works have verified NNCS with a single neural network controller on a small set of initial states [15]. The closed-loop ACAS Xu system involves switching between multiple neural networks and has a large set of initial states, creating a unique challenge for verification.…”

Section: Related Workmentioning

confidence: 99%

Neural Network Compression of ACAS Xu Early Prototype is Unsafe: Closed-Loop Verification through Quantized State Backreachability

Bak,

Tran

2022

Preprint

View full text Add to dashboard Cite

ACAS Xu is an air-to-air collision avoidance system designed for unmanned aircraft that issues horizontal turn advisories to avoid an intruder aircraft. Due the use of a large lookup table in the design, a neural network compression of the policy was proposed. Analysis of this system has spurred a significant body of research in the formal methods community on neural network verification. While many powerful methods have been developed, most work focuses on open-loop properties of the networks, rather than the main point of the system-collision avoidance-which requires closed-loop analysis.In this work, we develop a technique to verify a closed-loop approximation of ACAS Xu using state quantization and backreachability. We use favorable assumptions for the analysis-perfect sensor information, instant following of advisories, ideal aircraft maneuvers and an intruder that only flies straight. When the method fails to prove the system is safe, we refine the quantization parameters until generating counterexamples where the original (non-quantized) system also has collisions.

show abstract

RINO: Robust INner and Outer Approximated Reachability of Neural Networks Controlled Systems

Goubault

Putot

2022

Lecture Notes in Computer Science

View full text Add to dashboard Cite

We present a unified approach, implemented in the RINO tool, for the computation of inner and outer-approximations of reachable sets of discrete-time and continuous-time dynamical systems, possibly controlled by neural networks with differentiable activation functions. RINO combines a zonotopic set representation with generalized mean-value AE extensions to compute under and over-approximations of the robust range of differentiable functions, and applies these techniques to the particular case of learning-enabled dynamical systems. The AE extensions require an efficient and accurate evaluation of the function and its Jacobian with respect to the inputs and initial conditions. For continuous-time systems, possibly controlled by neural networks, the function to evaluate is the solution of the dynamical system. It is over-approximated in RINO using Taylor methods in time coupled with a set-based evaluation with zonotopes. We demonstrate the good performances of RINO compared to state-of-the art tools Verisig 2.0 and ReachNN* on a set of classical benchmark examples of neural network controlled closed loop systems. For generally comparable precision to Verisig 2.0 and higher precision than ReachNN*, RINO is always at least one order of magnitude faster, while also computing the more involved inner-approximations that the other tools do not compute.

show abstract

ARCH-COMP21 Category Report: Artificial Intelligence and Neural Network Control Systems (AINNCS) for Continuous and Hybrid Systems Plants

Cited by 16 publications

References 23 publications

Verification of Neural-Network Control Systems by Integrating Taylor Models and Zonotopes

Verification of Neural-Network Control Systems by Integrating Taylor Models and Zonotopes

Neural Network Compression of ACAS Xu Early Prototype is Unsafe: Closed-Loop Verification through Quantized State Backreachability

RINO: Robust INner and Outer Approximated Reachability of Neural Networks Controlled Systems

Contact Info

Product

Resources

About