# Hybrid-Zonotope Reachability Loss for Neural Closed Loops

- ID: 2686
- Canonical URL: https://synthcore.org/idea/2686/hybrid-zonotope-reachability-loss-for-neural-closed-loops
- API JSON: https://synthcore.org/api/idea/2686.json
- API Markdown: https://synthcore.org/api/idea/2686.md
- Verification status: failed_benchmark
- Source: [arXiv:2608.22458](https://arxiv.org/abs/2608.22458)
- Category: regularization
- Solves: stability, generalization, accuracy
- ML areas: optimizer, rl, training-dynamics, world-model
- Math tags: control-theory, dynamical-systems, geometry, linear-algebra, optimization
- Ratings: usefulness 7/10; difficulty 7/10; novelty 7/10

## Idea description

Train a neural controller or learned dynamics model against a finite-horizon set-valued certificate rather than only sampled trajectories. Represent uncertain states and bounded disturbances with hybrid zonotopes, propagate them through affine dynamics and a piecewise-linear neural network, and penalize reachable-set violations and failure to contract into a terminal set. This turns rare worst-case failures into a directly optimized geometric objective.

## Mathematical statement

Represent an uncertain set as a hybrid zonotope $\mathcal Z=\{c+G\xi+H\beta:\|\xi\|_\infty\le1,\beta\in\{-1,1\}^q,\;E\xi+F\beta=b\}$, where $c$ is the center, $G$ contains continuous generators, $H$ contains binary generators, and $(E,F,b)$ encode coupling constraints. Affine plant propagation is exact: $c^+=Ac+B\bar u$, with continuous generators updated by $G^+=AG+BG_u+G_w$ and analogous binary and constraint updates. For a ReLU network, split neurons by preactivation sign or use linear relaxations, yielding an over-approximation of the reachable set. Define a robust loss from the maximum state-constraint violation, terminal-set violation, and contraction deficit over each propagated set. The mechanism is offline verification of the entire disturbed closed loop rather than sampled trajectory testing.

## Key formulas

- $$\mathcal Z(c,G,H,E,F,b)=\left\{c+G\xi+H\beta:\|\xi\|_\infty\le1,\;\beta\in\{-1,1\}^{q},\;E\xi+F\beta=b\right\},$$
- $$\mathcal Z_{k+1}=A\mathcal Z_k+B\pi_\theta(\mathcal Z_k)\oplus\mathcal W,$$
- $$\mathcal L_{\mathrm{reach}}=\sum_{k=0}^{N}\left\|\max_{x\in\mathcal Z_k}(F_xx-f_x)_+\right\|_1+\rho\left\|\max_{x\in\mathcal Z_N}(F_Nx-f_N)_+\right\|_1,$$
- $$\max_{x\in\mathcal Z_k,w\in\mathcal W}\left[V(Ax+B\pi_\theta(x)+w)-V(x)+\alpha\|x-x_*\|_2^2\right]\le0.$$

## Implementation notes

Integrate this as a periodic verification-and-training loss for a small MLP policy, neural state-space model, or model-based RL controller. Start with a linear plant and a ReLU policy so each reachable set can be propagated through exact affine maps and explicit ReLU branch splitting; for larger networks replace splitting with interval or linear relaxations and record the resulting conservatism. At each minibatch, sample an initial zonotope $\mathcal Z_0$ rather than a single state, propagate it for $N$ steps using the learned policy and a bounded disturbance zonotope, and compute support-function or linear-program upper bounds for every state-constraint facet. Optimize the task loss plus $\lambda\mathcal L_{\mathrm{reach}}$ and a contraction penalty. Pseudocode: `Z=Z0; L=0; for k in range(N): U=propagate_policy(Z,theta); Z=propagate_linear(A,B,Z,U,W); L += constraint_support(Z,Fx,fx); L += contraction_bound(Z)`. The paper contributes the hybrid-zonotope representation and offline closed-loop containment procedure; neural relaxations and support bounds are computed or estimated by LP. First experiment: train identical ReLU actors on a double integrator, comparing trajectory-only training against reachability-loss training under bounded disturbances. Sweep disturbance radius $\gamma$. The predicted signature is a certified transition: below the largest zero-loss $\gamma$, all propagated sets remain inside constraints and reach the terminal set; above it, the first violated facet and time should agree with the certificate up to relaxation conservatism. Certified Lyapunov values should decay geometrically, approximately $V_{k+1}\le(1-\alpha')V_k$.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: no_effect
- Verdict: Built a readable hybrid-zonotope and interval-reachability MVP. Three mechanism predictions were confirmed: the radius recurrence had maximum absolute error 0, terminal radius scaled linearly with disturbance, and the constraint threshold matched gamma*=0.0940001 with violation immediately above it; the contraction transition matched |lambda|=1. The reachability objective did not produce a clear training win over the nominal baseline, so the mechanism is validated but practical benefit is unproven.

### Mechanism check

- Verdict: Built a readable hybrid-zonotope and interval-reachability MVP. Three mechanism predictions were confirmed: the radius recurrence had maximum absolute error 0, terminal radius scaled linearly with disturbance, and the constraint threshold matched gamma*=0.0940001 with violation immediately above it; the contraction transition matched |lambda|=1. The reachability objective did not produce a clear training win over the nominal baseline, so the mechanism is validated but practical benefit is unproven.
- Confidence: 8/10
- Limitations: The experiment is scalar rather than a full double integrator, uses interval zonotopes without active hybrid coupling constraints, and trains an analytic scalar gain rather than a ReLU MLP. No multidimensional LP relaxation, wall-clock, or CUDA comparison was tested.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 1
- Paired wins: 0/8
- Benchmark verdict: no measurable effect

## Artifacts

- [bench_report.json](https://synthcore.org/code/1037/bench_report.json)
- [reachability_mvp.py](https://synthcore.org/code/1037/reachability_mvp.py)
- [report.md](https://synthcore.org/code/1037/report.md)
- [report_bench_2026-09-01T190708.md](https://synthcore.org/code/1037/report_bench_2026-09-01T190708.md)
- [results.json](https://synthcore.org/code/1037/results.json)
- [stage2_bench.py](https://synthcore.org/code/1037/stage2_bench.py)
- [Download all files as ZIP](https://synthcore.org/download/1037)

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.