Model-Ensemble Space-Filling Explorer
Implementation & benchmark of arXiv:2608.13360 — Robust Space-Filling Input Design via Stochastic Optimization
Source paper: Robust Space-Filling Input Design via Stochastic Optimization arXiv:2608.13360 ⓘ · analyzed Sep 1, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Train an input-generation policy or differentiable signal parameterization to produce trajectories that cover the joint input-state feature space while remaining informative for every plausible neural world model. Replace single-model experiment design by an expectation over an ensemble of models, and optimize this objective with stochastic model and trajectory samples.
Formulas
Mathematical statement
The paper models a trajectory through x_(k+1) = f_m(x_k,u_k) and y_k = g_m(x_k,u_k), with feature z_k = (u_k,x_k). Let M be an ensemble or posterior distribution of candidate world models, let psi parameterize an input policy u_k = pi_psi(x_k,xi_k), and let z_(0:T)^m(psi) be the resulting trajectory under model m sampled from p(m). Define a space-filling or information criterion Phi(psi,m) = log det(epsilon I + (1/(T+1)) sum_k w(z_k) w(z_k)^T), where w(z) is a bounded feature embedding and epsilon is a positive numerical stabilizer. The robust design objective is J(psi) = E_m,xi[Phi(psi,m)] - lambda C(psi), where C penalizes input magnitude, constraint violations, or unsafe states. Stochastic approximation uses psi_(r+1) = psi_r + eta_r times an estimated gradient of J. If the criterion and gradient variance are bounded, diminishing step sizes with sum eta_r infinite and sum eta_r squared finite provide the standard stochastic-approximation convergence regime toward a stationary point.
Implementation notes
1. Integration point: use this as a data-collection module before or between updates of a neural state-space model. Parameterize the experiment by a small MLP policy u_k = pi_psi(x_k,xi_k), or by bounded Fourier or chirp coefficients if differentiating through the environment is unavailable. Maintain an ensemble of M world models f_(theta_m) trained on the currently collected data. 2. Pseudocode: sample B model indices; roll out the policy for T steps with x_(k+1) = f_(theta_m)(x_k,u_k); compute z_k = (u_k,x_k); construct G_m = epsilon I + (1/(T+1)) sum_k w(z_k)w(z_k)^T; set the minibatch objective to the average of log det(G_m) minus lambda times the safety and action cost; backpropagate through the rollout and update psi by gradient ascent. Clip actions and assign a large penalty to unsafe states. 3. Computed from the paper's mechanism: sampling a population of candidate models, optimizing the expected criterion, and using stochastic approximation. Estimated empirically: the model distribution, gradients, feature map w, and safety penalty. 4. First cheap experiment: use the scalar nonlinear system x_(k+1) = 0.8 x_k + a x_k cubed + u_k, with an ensemble in which coefficient a varies, and compare random excitation, single-model design, and ensemble design while training small neural predictors. Measure held-out one-step and 20-step prediction error, minimum-model log determinant, and gradient variance. The quantitative prediction is that ensemble design improves the minimum-model log determinant by at least 10 percent over single-model design at equal rollout budget, while the variance of the minibatch objective estimate decreases approximately as 1/B. As ensemble spread increases, single-model performance should degrade sharply, whereas robust design should retain flatter worst-model information until input constraints become active.
Verification
Stage 1 · Toy mechanism gate: Failed ✗
Stage 2 · Mechanism transferred to benchmark: Not tested
Stage 2 · Practical benchmark result: Not run
Methodology: Toy-system gate first; the benchmark stage runs only after a pass. How verification works
Stage 1 — Mechanism check agent confidence 9/10
Built a differentiable ensemble space-filling explorer with nonlinear candidate world models, log-determinant feature coverage, nominal and ensemble optimizers, finite-difference validation, ensemble-spread sweeps, and minibatch-gradient variance tests. The gradient check passed to 2.36e-9 maximum absolute error, and gradient MSE followed approximately 1/B with B*MSE near 4.8e-8. However, robust optimization did not improve worst-model log determinant: at spread 0.18 the change was -0.026%, far below the proposed 10% gain, so the central effect was not observed.
- Agent confidence
- 9/10
- Baseline
- Nominal single-model design; worst-model Phi at spread 0.18 = 2.54541
- Idea
- Ensemble design; worst-model Phi at spread 0.18 = 2.54475, gain = -0.026%; gradient MSE: 4.87e-8 (B=1), 7.48e-10 (B=64)
Limitations: Only a small open-loop scalar toy system and one fixed initialization were tested; no neural predictor training, learned policy, safety constraints, held-out prediction error, multiple random seeds, GPU execution, or larger model ensembles were evaluated.
How to run: python3 ensemble_explorer.py
Verdict computed by deterministic test code from paired-seed statistics — not by the language model.
Artifacts
- ensemble_explorer.py 4.0 KB View Raw
- report.md 1.4 KB View
- results.json 2.6 KB View Raw