Scalable Synthesis of Formally Verified Neural Value Function for Hamilton-Jacobi Reachability Analysis
Abstract: Hamilton-Jacobi (HJ) reachability analysis provides a formal method for guaranteeing safety in constrained control problems. It synthesizes a value function to represent a long-term safe set called feasible region. Early synthesis methods based on state space discretization cannot scale to high-dimensional problems, while recent methods that use neural networks to approximate value functions result in unverifiable feasible regions. To achieve both scalability and verifiability, we propose a framework for synthesizing verified neural value functions for HJ reachability analysis. Our framework consists of three stages: pre-training, adversarial training, and verification-guided training. We design three techniques to address three challenges to improve scalability respectively: boundary-guided backtracking (BGB) to improve counterexample search efficiency, entering state regularization (ESR) to enlarge feasible region, and activation pattern alignment (APA) to accelerate neural network verification. We also provide a neural safety certificate synthesis and verification benchmark called Cersyve-9, which includes nine commonly used safe control tasks and supplements existing neural network verification benchmarks. Our framework successfully synthesizes verified neural value functions on all tasks, and our proposed three techniques exhibit superior scalability and efficiency compared with existing methods.
- Fossil: a software tool for the formal synthesis of lyapunov functions and barrier certificates using neural networks. In Proceedings of the 24th International Conference on Hybrid Systems: Computation and Control, pp. 1–11.
- Formal synthesis of lyapunov neural networks. IEEE Control Systems Letters, 5(3), 773–778.
- A minimum discounted reward hamilton–jacobi formulation for computing reachable sets. IEEE Transactions on Automatic Control, 69(2), 1097–1103.
- Formally verified neural network control barrier certificates for unknown systems. IFAC-PapersOnLine, 56(2), 2431–2436.
- Hamilton-jacobi reachability: A brief overview and recent advances. In 2017 IEEE 56th Annual Conference on Decision and Control (CDC), pp. 2242–2253. IEEE.
- Deepreach: A deep learning approach to high-dimensional reachability. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pp. 1817–1824. IEEE.
- Predictive control for linear and hybrid systems. Cambridge University Press.
- First three years of the international verification of neural networks competition (vnn-comp). International Journal on Software Tools for Technology Transfer, 25(3), 329–339.
- Openai gym..
- Neural lyapunov control. Advances in neural information processing systems, 32.
- Fast reachable set approximations via state decoupling disturbances. In 2016 IEEE 55th Conference on Decision and Control (CDC), pp. 191–196. IEEE.
- Verification-aided learning of neural network barrier functions with termination guarantees..
- Lyapunov-stable neural-network control..
- Overcoming the curse of dimensionality for some hamilton–jacobi partial differential equations via neural network architectures. Research in the Mathematical Sciences, 7(3), 20.
- Ehlers, R. (2017). Formal verification of piece-wise linear feed-forward neural networks. In Automated Technology for Verification and Analysis: 15th International Symposium, ATVA 2017, Pune, India, October 3–6, 2017, Proceedings 15, pp. 269–286. Springer.
- Bridging hamilton-jacobi safety analysis and reinforcement learning. In 2019 International Conference on Robotics and Automation (ICRA), pp. 8550–8556. IEEE.
- Explaining and harnessing adversarial examples..
- Agile but safe: Learning collision-free high-speed legged locomotion..
- The safety filter: A unified view of safety-critical control in autonomous systems. Annual Review of Control, Robotics, and Autonomous Systems, 7.
- Safety and liveness guarantees through reach-avoid reinforcement learning. In 17th Robotics: Science and Systems, RSS 2021. MIT Press Journals.
- Verisig: verifying safety properties of hybrid systems with neural network controllers. In Proceedings of the 22nd ACM International Conference on Hybrid Systems: Computation and Control, pp. 169–178.
- Learning multiple layers of features from tiny images..
- Mnist handwritten digit database..
- Algorithms for verifying deep neural networks. Foundations and Trends® in Optimization, 4(3-4), 244–404.
- Safe control under input limits with neural control barrier functions. In Conference on Robot Learning, pp. 1970–1980. PMLR.
- Towards deep learning models resistant to adversarial attacks..
- Mitchell, I. M. (2007). Comparing forward and backward reachability as tools for safety analysis. In International Workshop on Hybrid Systems: Computation and Control, pp. 428–443. Springer.
- Mitchell, I. M. (2008). The flexible, extensible and efficient toolbox of level set methods. Journal of Scientific Computing, 35, 300–329.
- Neural network dynamics for model-based deep reinforcement learning with model-free fine-tuning. In 2018 IEEE international conference on robotics and automation (ICRA), pp. 7559–7566. IEEE.
- Acas xu: Integrated collision avoidance and detect and avoid capability for uas. In 2019 IEEE/AIAA 38th Digital Avionics Systems Conference (DASC), pp. 1–10. IEEE.
- Automated and formal synthesis of neural barrier certificates for dynamical models. In International conference on tools and algorithms for the construction and analysis of systems, pp. 370–388. Springer.
- Learning a formally verified control barrier function in stochastic environment..
- Evaluating robustness of neural networks with mixed integer programming..
- Nnv: the neural network verification tool for deep neural networks and learning-enabled cyber-physical systems. In International Conference on Computer Aided Verification, pp. 3–17. Springer.
- Beta-crown: Efficient bound propagation with per-neuron split constraints for neural network robustness verification. Advances in Neural Information Processing Systems, 34, 29909–29921.
- Simultaneous synthesis and verification of neural control barrier functions through branch-and-bound verification-in-the-loop training..
- Improve certified training with signal-to-noise ratio loss to decrease neuron variance and increase neuron stability..
- Wolsey, L. A. (2020). Integer programming. John Wiley & Sons.
- Fast and complete: Enabling complete neural network verification with rapid and massively parallel incomplete verifiers..
- Safe model-based reinforcement learning with an uncertainty-aware reachability certificate..
- Exact verification of relu neural control barrier functions. Advances in Neural Information Processing Systems, 36.
- Efficient neural network robustness certification with general activation functions. In Advances in neural information processing systems, Vol. 31.
Paper Prompts
Sign up for free to create and run prompts on this paper using GPT-5.
Top Community Prompts
Collections
Sign up for free to add this paper to one or more collections.