Northeastern University · Siami Lab
I'm a PhD student in Electrical and Computer Engineering at Northeastern University, advised by Prof. Milad Siami. My research builds provable stability and safety certificates for systems controlled by neural networks — combining positive-systems theory, absolute stability, and sector-bounded analysis so that learned controllers come with formal guarantees rather than empirical hope.
The recurring theme is cheap certificates: replacing semidefinite-programming analysis of neural feedback loops with elementwise sector bounds plus Metzler/Hurwitz conditions. This turns a solver call into a matrix test that runs orders of magnitude faster than SDP-based IQC methods — and certifies regimes those methods cannot reach.
Always glad to talk about safe learning-based control, positive systems, or potential collaborations.
- Certified neural-network control — stability and safety guarantees for NN-in-the-loop systems
- Positive systems — Metzler / Hurwitz conditions and positive Aizerman-type results
- Absolute stability — sector-bounded nonlinearities, integral quadratic constraints, Lyapunov certificates
- Region of attraction and verification — local sector bounds via layer-wise linear relaxation
- Robustness under delay and uncertainty — delay-independent certificates, interval matrices, stability radius
- Learning-based control — imitating optimal or expert controllers while preserving guarantees
Two open-source libraries, in Python and MATLAB, for certifying feedback loops that contain a neural network. nnsector bounds the network, and nnstab uses those bounds to check that the closed loop is stable.
nnsector: sector bounds for feedforward neural networks
Give it a trained network, a point y* and a box around it. It returns two slopes gamma1 <= gamma2, and on the box the network stays between the two lines through (y*, pi(y*)) with these slopes. This is the form that absolute-stability tests need (positive Aizerman, circle criterion, Lur'e systems), so you can use it to certify a feedback loop with a neural network controller, observer or sensor.
- Local sector bounds on a box around any point (CDC 2025), and global bounds that hold everywhere (L-CSS 2024)
- Works with or without biases, for tanh, sigmoid, ReLU, leaky ReLU and softplus activations
- Reads PyTorch
nn.Sequentialmodels - The same functions in Python and MATLAB (the MATLAB version also runs in GNU Octave)
nnstab: stability checks for feedback loops with a neural network controller
Give it a plant x' = A x + B u, y = C x and the trained network u = pi(y) that closes the loop. It checks whether the loop is exponentially stable, using the positive Aizerman conditions on the sector bounds from nnsector: A + B gamma1 C must be Metzler and A + B gamma2 C must be Hurwitz. No SDP solver is needed.
- A global test for every starting state
x0 >= 0(L-CSS 2024), and if it fails, the largest certified region of attraction around 0 (CDC 2025) - A guaranteed decay rate in the region, and a simulator to test the result
- For positive loops:
B >= 0,C >= 0andpi(0) = 0(for example, a network without biases with tanh, ReLU or leaky ReLU activations) - Reads PyTorch models; the same functions in Python and MATLAB (the MATLAB version also runs in GNU Octave)
pip install git+https://github.com/hmon91/nnsector.git # sector bounds only
pip install git+https://github.com/hmon91/nnstab.git # stability checks (also installs nnsector)import nnsector as ns
import nnstab as nst
net = ns.demo_network() # small tanh network that ships with nnsector
sb = ns.local_bound(net, -1, 1) # box [-1, 1] around y* = 0
print(sb.gamma1, sb.gamma2) # about -2.786 and -1.200
A, B, C, net = nst.example_system("local") # an unstable plant with a NN controller
res = nst.verify(A, B, C, net) # global test, then a region of attraction
print(res.status) # 'local': every x0 >= 0 with sum(x0) <= 1.706 goes to 0Journal
- Ensuring Both Positivity and Stability Using Sector-Bounded Nonlinearity for Systems with Neural Network Controllers H. Montazeri Hedesh, M. Siami IEEE Control Systems Letters (L-CSS), vol. 8, pp. 1685–1690, 2024 IEEE Xplore · arXiv:2406.12744 · code · libraries: nnsector, nnstab
Conference
-
Delay-Independent Safe Control with Neural Networks: Positive Lur'e Certificates for Risk-Aware Autonomy H. Montazeri Hedesh, M. Siami American Control Conference (ACC), New Orleans, LA, USA, 2026, pp. 134-141 IEEE Xplore · arXiv:2510.06661 · code
-
Robust Stability Analysis of Positive Lur'e System with Neural Network Feedback H. Montazeri Hedesh, M. K. Wafi, B. Shafai, M. Siami IEEE Conference on Control Technology and Applications (CCTA), San Diego, CA, pp. 976–981, 2025 IEEE Xplore · arXiv:2505.18912 · code
-
Local Stability and Region of Attraction Analysis for Neural Network Feedback Systems under Positivity Constraints H. Montazeri Hedesh, M. K. Wafi, M. Siami 64th IEEE Conference on Decision and Control (CDC), Rio de Janeiro, Brazil, 2025, pp. 1142-1149 IEEE Xplore · arXiv:2505.22889 · code · libraries: nnsector, nnstab
-
Predator-Swarm-Guide Dynamics: A Hybrid Approach to Crowd Modeling and Guidance in Mass Shooting Scenarios A. Darabi, H. Montazeri Hedesh, M. Siami, M. Sznaier American Control Conference (ACC), Toronto, ON, Canada, 2024, pp. 4402–4408 IEEE Xplore · code
Full, up-to-date list on my Google Scholar.
Each repository reproduces the figures and tables of its paper. The set traces one research program: from global stability outward to local, robust, and delay-tolerant settings.
| Repository | Paper | Stack |
|---|---|---|
| Global stability of NN-controlled systems | Closed-form sector bound + Metzler/Hurwitz test for global exponential stability, with an IQC run-time baseline. (L-CSS 2024) | MATLAB Python |
| Local stability of NN-controlled systems | Localized Aizerman conditions and layer-wise linear relaxations for region-of-attraction estimation. (CDC 2025) | Jupyter |
| Robustness of NN-controlled systems | Closed-form stability radius for positive Lur'e systems under structured uncertainty, plus sector-bound refinement. (CCTA 2025) | MATLAB |
| Delay-independent safety of NN controllers | Certificates holding across state/input delays and interval-matrix uncertainty. (ACC 2026) | MATLAB |
| PSGModel | Predator–Swarm–Guide dynamics for crowd modeling and guidance. (ACC 2024) | MATLAB |
Also: CVX · MOSEK · YALMIP for the SDP baselines.

