acceptodds
Under review as a conference paper at ICLR 2027

Modular Decomposition of Learned Controlled Dynamics from One Fit

Abstract

Learned models of controlled systems are often trained and deployed as monolithic predictors, even when the underlying physical system is organized into interacting subsystems that are built, operated, and maintained separately. Decomposing such a predictor usually requires choosing module boundaries and retraining a new model for each candidate structure. We study coupling-additive predictors, in which the prediction is a sum of contributions, one for each pair of a sensor group and a control channel; the class contains bilinear Koopman models with grouped dictionaries and any network with a linear head over per-group features. We introduce MODEC, which evaluates every candidate decomposition of a ridge-fitted predictor in this class from that single fit: the weights that restricted retraining would produce for any set of removed couplings, and the resulting increase in training loss, follow in closed form from the cached fit, and a covariance matrix of the fitted coupling contributions scores all candidate boundaries at once. On independent held-out episodes, a finite-sample bound on the cost of a restriction holds simultaneously over all candidate boundaries, with their number entering only through a logarithm. A selected decomposition runs as separate modules that reproduce the restricted predictor exactly, and a finite-horizon recursion bounds how its deviation from the original model accumulates over multi-step rollouts. On the Tennessee Eastman process and the IEEE 39-bus system, the scores rank the exact retraining cost of candidate modules with Spearman correlation 0.91 and 0.89, and the selected decompositions remove 67% and 84% of the removable couplings at 10% and 38% relative increases in one-step NRMSE; on a six-reactor cascade, MODEC recovers the planted subsystem structure at every coupling strength, with a restriction cost that becomes negligible when the subsystems are decoupled.

Primary area:
learning on time series and dynamical systems
Keywords:
Modular decomposition, Koopman operator learning, Learning dynamical systems

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_rtZSOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies inverse problems with auxiliary calibration channels that only observe shared nuisance variables. It shows the benefit of calibration is governed by paired nuisance geometry rather than individual channel quality, derives minimax rates and critical sample complexity, and reveals that calibration timing drastically changes sample requirements for dynamical parabolic systems. It further proposes target-oriented resource allocation distinct from nuisance-optimal allocation.

Critical strengths

Existing modular dynamical‑system learning methods either impose modular structure during training, or require expensive retraining for every candidate partition. MODEC works as a post‑hoc tool: given one monolithic fitted model, it scores all candidate module boundaries by re‑using cached ridge statistics. This avoids repetitive full retraining, which is a major practical pain‑point for networked physical‑system modelling.

Critical weaknesses
  1. MODEC is limited to coupling-additive models, where predictions are linear within predefined coupling blocks. It cannot directly handle arbitrary nonlinear interactions in general neural networks, particularly cross-group interactions not explicitly modeled. The paper should discuss how model misspecification affects decomposition quality and suggest practical ways to handle deviations from the coupling-additive assumption.

  2. The greedy merging strategy may fail to find a globally Pareto-optimal decomposition. The paper should discuss potential failure cases and alternatives, such as local search or mixed-integer optimization.

  3. MODEC relies on predefined feature blocks and candidate partitions, so it cannot recover true subsystem boundaries absent from the candidate set. The paper should clarify that MODEC ranks existing partitions rather than discovering structures from scratch.

Questions for authors
  1. The method requires a pre‑specified family of candidate cuts/partitions. Could the authors clarify how practitioners should define deployment groups and candidate cut families in real‑world scenarios where true physical subsystem boundaries are unknown?

  2. MODEC is derived for coupling‑additive predictors. What happens when the real‑world dynamics contain interaction terms that cannot be represented within the coupling‑additive model class?

  3. The paper uses agglomerative greedy merging over the reweight score for searching decompositions. While it works well across the three presented benchmarks, greedy algorithms are prone to local optima. For larger problem scales with huge partition space (similar to IEEE‑39‑bus), what alternative search strategies would you recommend if greedy search gets stuck in poor local optima?

3 comments
AuthorsOpenReview ↗

Response on scope, search, and deployment groups

Thank you for the questions on model scope, deployment groups, and search; the revision addresses each.

Representation and misspecification. MODEC fits a linear head over declared feature blocks, whose maps can be nonlinear and may read several groups. A learned encoder can be used if frozen during the ridge fit. Our IEEE model includes 20 machine-pair blocks. A module retains a pair term when it can access both machines and the relevant control.

The scores evaluate restrictions within the selected representation. Appendix A.5 gives a failure case: two centered group variables and all remaining inputs are mutually independent, and each available term reads at most one of the two groups. Their product is orthogonal to the feature class, so it contributes zero population restriction cost although the dynamics depend on both groups. Other input distributions can project an omitted interaction onto represented features, and finite-data scores can be nonzero. The restriction cost should therefore be read alongside the fitted monolith's absolute prediction error.

For a restricted predictor, error relative to the true conditional mean separates into the full model class's approximation error, the best achievable cost of the boundary, and the fitting error within the restricted class. Interpreting a low boundary cost as evidence about the plant therefore requires a model-adequacy check.

In practice, we recommend checking that error against a suitably tuned richer predictor and examining residual dependence on candidate interactions. Pilot data can guide new pair features. The enlarged dictionary is then fitted before scoring its restrictions. If D2D_2 (selection fold) results motivate the representation change, fresh independent scoring episodes preserve the original evaluation protocol. We tested this on CSTR at εinter=0.10\varepsilon_{\rm inter}=0.10 by adding six interaction-only blocks (products and joint random features of wired reactor pairs, three across planted modules). The monolith's D3D_3 (test fold) NRMSE falls from 0.0165 to 0.0078. At three modules, reweight selection still recovers the planted partition; its restricted refit improves from 0.1613 to 0.1538 and retains 4,000 rather than 1,952 coefficients. Section 3 now covers the Appendix A.5 example and the error split (Lemma A.2); Remark A.3 gives this diagnostic; Appendix E.15 reports the experiment.

Search. Greedy merging produces a nested sequence and can miss the optimum at a later module count. A four-group cut-cost example (Appendix C) illustrates the failure. With edges AD = 10, AB = 9.2, BC = 9.1, and AC = 9.0, greedy first merges A and D, then returns ABD ∣\mid C with cut cost 18.1 at two modules. The alternative ABC ∣\mid D has cost 10.

The submitted exhaustive comparisons quantify the gap. Greedy attains the minimum reweight score at every TEP module count and at five of eight nontrivial IEEE counts. At six IEEE modules, greedy and exhaustive-score selection have the same cb=0.841c_b=0.841 (share of removable couplings cut), with D3D_3 NRMSE 0.8037 and 0.7995. Exact-refit local search reaches the exhaustive D2D_2 refit optimum at each tested module count from two to six on both systems; its reported search budget is allocated separately at each count.

With closed modules and the benchmark's block-local actuator rule, each module's allowed terms depend only on its containing group set BB. An additive partition objective therefore has the form C(P)=∑B∈PC(B)C(P)=\sum_{B\in P}C(B). After block-cost preparation, subset dynamic programming gives the optimum for KK modules and MM groups in O(K3M)O(K3^M) time, and the same costs define a set-partitioning integer program. In new runs, exact selection over all 115,975 closed IEEE partitions took under 0.4 s from 1023 cached block costs. Candidate-block construction remains exponential in MM. For larger systems, beam search and move-and-swap refinement can explore additional partitions with the same evaluator.

Deployment groups. The atomic units, feature supports, and control-access rules are supplied. MODEC selects the final partition among admissible unions of those units and recovers the planted reactor pairs at K=3K=3. The six-reactor atoms admit 203 closed partitions, 90 of which are at K=3K=3. A boundary inside a supplied atom requires a finer initial grouping. Practitioners can begin with the smallest feasible set of sensors or devices, specify the inputs each feature reads, and record private and shared actuator access before scoring. The declared family in our experiments includes every closed partition; the submitted MODEC algorithm visits a greedy path through it. Exhaustive enumeration supplies separate reference optima.

The simultaneous evaluation results cover selection within the family declared before D2D_2 for the stated quantities and candidate predictors fixed before that fold. Section 5 states these conditions; Section 4.2 and Appendix C give the exact alternatives.

AuthorsOpenReview ↗

Follow-up on scope, search, and deployment groups

Thank you again for your questions. Section 3 now explains how omitted interactions can affect decomposition scores and suggests checking residuals for missing interactions. In the new CSTR experiment (Appendix E.15), MODEC still recovered the planted partition after we added pairwise interaction features, and the restricted refit’s test NRMSE dropped from 0.1613 to 0.1538.

Appendix C now gives a concrete example of when greedy merging misses the optimum. It also describes exact search using cached block costs for closed modules and move-and-swap refinement for larger problems.

We also clarified how to define the search space when physical boundaries are unknown. Practitioners can start with the smallest feasible set of sensors or devices and specify actuator access. MODEC then searches for module boundaries among unions of those groups.

I would like to thank the authors for their detailed responses, and I apologize for the inaccuracies in my original paper summary. The authors' responses have addressed my concerns. However, since my expertise in this research area is limited, I can only assign a confidence score of 1 to my review.

Reviewer_S2cKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents MODE, a method finding modular decomposition for physical systems (plants, power networks, ..) with interacting subsystems without re-training. The method is directly useful for maintainability and distributed execution. The method is evaluated via comprehensive experiments.

Critical strengths

The problem is interesting and the approach promising. The approach is computationally efficient, as it allows evaluating multiple candidate cuts using previously computed statistics, without repeatedly processing the training data. The method is empirically and theoretically supported as its performance is validated carefully on realistic benchmarks (Tennessee Eastman, IEEE 39-bus, reactor cascade).

Critical weaknesses

At its present form, the paper does not explain its motivation and novelty clearly enough to make them understandable. The introduction is very compressed, core concepts are introduced too quickly and many relevant components are combined at once making their relevance and relationships unclear. Also it is not specified in Related Work what are the papers own contributions. Eg the restricted update from Hessian statistics is acknowledged to be based on known pruning methodology, but the paper does not clearly explain how this is extended or made novel within MODEC. This makes the significance of the paper’s contributions difficult to evaluate.

Several variables are either introduced too quickly or used in a slightly ambiguous way.

The experiments use two benchmarks, Tennessee Eastman process and the IEEE 39-bus system, but they are not explained nor cited in the paper. While known benchmarks in their field, they are not so widely known in the ML community that could be left without any presentation.

Questions for authors

What is the papers main problem statement? Which elements of MODEC are novel versus which are direct applications of known ideas? What are the main limitations of the method?

3 comments
AuthorsOpenReview ↗

Response on problem statement, novelty, and limitations

Thank you for pointing out where the motivation and contribution need clearer explanation. The revised introduction now explicitly states the deployment decision. D1D_1, D2D_2, D3D_3 are disjoint episode folds for fitting, selection, and testing.

Problem statement. We are given a fitted predictor of controlled dynamics, groups of measurements that can be deployed together, and rules specifying the state variables and actuators each module may access. A proposed partition determines which model terms can remain in each module. MODEC evaluates the one-step prediction cost of these restrictions and computes their restricted ridge weights using statistics from the original fit.

For example, consider two units A and B with local actuators uAu_A and uBu_B. A module for A retains its intercept, features supported on A, and terms using its permitted controls, including direct-control terms. It removes terms requiring B's measurements or unavailable controls. A feature that reads both units requires access to both. This example now appears in the introduction, before the formal notation.

The method uses three calculations. A frozen restriction removes contributions and adjusts the intercept. Reweighting also allows one scalar adjustment per surviving fitted contribution; this is the search criterion. The final restricted ridge fit optimizes all coefficients in the surviving blocks.

Contribution and established ingredients. The classical ingredient is Proposition 1: for ridge regression, the partitioned-regression identity gives, for any fixed coefficient mask, the restricted refit and its training-objective increase from the cached regularized Gram matrix. Second-order pruning uses the same identity to price masks chosen for sparsity, and we claim no novelty for it. MODEC adds three things, now listed as contributions in Section 1. First, the masks come from deployment access rules rather than from sparsity, so each legal module boundary yields an exact restricted predictor that runs module by module (Proposition 3). Second, boundaries are ranked by held-out predictive cost rather than by the training-objective increase, using covariance scores of the fitted coupling contributions. Third, we give simultaneous finite-sample bounds for the frozen restriction cost and, separately, for the risk difference of the exact restricted refits. With 10% of D1D_1, the in-sample objective increase from the classical identity loses substantial rank agreement with test loss on TEP and IEEE, while the held-out criteria degrade much less.

Equations (12)-(13) compute frozen and reweight costs from the fitted coupling contributions and their covariance with the prediction residual. This residual term matters for an imperfect fitted monolith. Lemma A.1 characterizes zero-cost cuts of the population projection under Assumption 2. Appendix A.3 shows that contribution covariance alone cannot determine the full-refit cost. The coupling attributions apply to the quadratic frozen-cost game.

Theorem 1 controls the frozen cost simultaneously over a declared candidate family and upper-bounds the best reduced-class excess risk relative to the intercept-adjusted monolith. Corollary B.1 separately bounds the risk difference for a restricted predictor fixed before D2D_2, relative to the fitted monolith. These results allow selection on D2D_2 within the declared family under the stated assumptions. The concentration inequality and rollout recursion are established tools applied to these quantities.

Limitations and presentation. Section 8 now lists the main limitations. The feature dictionary and atomic deployment groups determine which interactions and boundaries can be represented. Adding interaction features requires a new initial fit. Greedy merging explores a heuristic path, but exact selection over cached block costs is fast at our scales (under 0.4 s for all 115,975 IEEE partitions). The statistical results concern one-step prediction under independent evaluation episodes. The rollout envelope compares the restricted model with the fitted monolith over a finite horizon while both remain in the declared operating region. Cached exact-refit evaluation is affordable at our tested scales. On IEEE, with the candidate family prepared, post-fit statistics and selection take an estimated 22.18 s for batched reweight selection and 45.98 s for exact-refit validation selection, summed from measured components, excluding final deployment fitting. The original unbatched implementation in Appendix E.11 was slower than exact-refit evaluation.

We added a notation table to the appendix and describe the benchmarks in Section 7. TEP is a simulated chemical plant; our experiment uses 22 continuous measurements, 11 manipulated variables, and seven deployment groups. IEEE 39-bus is a ten-generator power-system benchmark simulated with ANDES. Section 7 and Appendix E.1 now cite Downs and Vogel (1993), Athay et al. (1979), and ANDES.

AuthorsOpenReview ↗

Follow-up on problem statement, novelty, and limitations

Thank you again for your feedback. The revised introduction uses a two-unit example to make the problem concrete. Section 1 now identifies the restricted ridge refit formula as a classical identity and clarifies MODEC’s contribution. MODEC uses deployment rules to define admissible cuts, scores them on held-out data, and produces executable module predictors. Section 5 clarifies the separate guarantees for the frozen restriction and the deployed refit.

The conclusion now explains how the chosen features and initial groups limit the structures MODEC can represent. We also added a notation table and descriptions and citations for the TEP and IEEE 39-bus benchmarks. We’d be happy to clarify any remaining points.

Thank you for your extensive work on providing the clarifications and all revision efforts. While the replies address several issues I rised, my concerns about the clarity of the motivation and novelty remain, so I'm keeping my original rating.

Reviewer_ueQdOpenReview ↗
Rating
2
Confidence
1
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

This paper proposes MODEC, a method for decomposing a fitted controlled-dynamics predictor into modules from a single ridge-regression fit. It derives exact restricted-refitting formulas from cached statistics, develops covariance-based scores for candidate decompositions with a finite-sample guarantee. Experiments on some benchmarks demonstrate promising trade-offs between modularity and prediction accuracy.

Critical strengths
  • The paper addresses a relevant problem.
  • The exact restricted-refitting identities and simultaneous held-out bound may be useful.
  • Experiments on three benchmarks offer encouraging evidence of practical utility.
Critical weaknesses
  • Limited comparison with established methods: The empirical evaluation relies mainly on generic clustering heuristics and local search. Comparisons against recent methods are missing.
  • The writing seems hard to read, maybe due to heavily assisted by AI tools. The proofs of some theorems seem to contain ideas, not the full proofs.
Questions for authors

NA

2 comments
AuthorsOpenReview ↗

Response on comparisons, readability, and proofs

Thank you for identifying the comparison and presentation concerns. We ran both comparisons on all three benchmarks and added them to the revision (Section 7.5, Appendices E.12-E.14). D1D_1, D2D_2, D3D_3 are disjoint episode folds for fitting, selection, and testing.

Comparisons. All selectors share the trained predictor, feature dictionary, legal module boundaries, and final restricted ridge estimator. Every selection is frozen before it is evaluated on D3D_3.

The criterion study tests what MODEC's held-out selection layer adds to the established quadratic restriction calculation, the exact training-objective increase of Equation (6) on D1D_1. We compare it with three D2D_2 criteria built from the same cached statistics: the frozen cost, MODEC's reweight cost, and the validation loss of the exact restricted refit, which defines our refit-selected reference. Each is minimized exactly over all closed partitions (203 CSTR, 877 TEP, 115,975 IEEE). On CSTR, all four select the same partition at every module count and recover the planted modules at three. The table gives the median Spearman correlation with the D3D_3 loss of the deployed refit within predeclared strata of equal module count and similar coupling budget, using all of D1D_1 or 10% of it (range over three seeds).

SystemD1D_1 usedEq. (6), in-sampleMODEC reweight, D2D_2Refit-selected, D2D_2
TEP100%0.9740.5321.000
TEP10%0.07 to 0.270.18 to 0.410.998 to 1.000
IEEE 39-bus100%0.9440.9070.992
IEEE 39-bus10%0.55 to 0.660.71 to 0.770.991 to 0.994

Held-out evaluation helps most when the fit is data-limited. At the tested 10% fraction of D1D_1 (20 IEEE episodes, fewer samples than features), the in-sample criterion loses much of its rank agreement, and MODEC's reweight criterion has lower D3D_3 NRMSE in 20 of 24 IEEE seed and module-count pairs (2 ties, 2 losses) and 9 of 15 TEP pairs (6 ties, no losses). The refit-selected reference is lower in all 24 IEEE and 12 of 15 TEP pairs, with ties otherwise, and ranks best with all of D1D_1; its D2D_2 loss estimates the quantity that D3D_3 measures. The revision reports it as a comparison; MODEC's procedure is unchanged. With all of D1D_1, at the exact coupling budget of the reweight choice, the reweight score matches the best criterion at five of eight IEEE module counts and is within 0.0023 elsewhere. On IEEE, with the candidate family prepared, post-fit statistics and selection take an estimated 22.18 s for batched reweight selection and 45.98 s for exact-refit validation selection, summed from measured components, excluding final deployment fitting.

Community baseline. We implemented input-output community decomposition (Tang and Daoutidis, 2018; Appendix E.14) on linearizations of the simulators, following the published digraph, affinities, bipartite modularity, and recursive bisection. The documented adaptations are time units chosen to avoid negative-weight cycles, a fixed rule mapping communities to deployment atoms, and fixed-module-count modularity maxima on CSTR and TEP for matched comparison. We compare only the partition step, not the full control design. Its stopping rule returns the most fragmented partition on CSTR and IEEE (every unit alone; D3D_3 NRMSE 0.3247 and 0.9326 against monoliths 0.0165 and 0.5822), the unique partitions at those module counts, where every selector ties, whereas MODEC traces the whole accuracy-composability path. On TEP it gives six modules with 30 retained couplings and 0.4614, tying the refit-selected reference at that budget, while exhaustive reweight selection gives 0.4655; MODEC's six-module partition (0.4554) retains 31 couplings. In fixed-module-count comparisons on CSTR and TEP at equal budgets, it never improves on the refit-selected reference. It also requires a mechanistic plant model, which MODEC does not.

Readability and proofs. We reorganized the introduction around a two-unit deployment example that distinguishes the frozen restriction, scalar reweighting, and restricted ridge predictor before their formulas. Section 1 now states the known quadratic identity separately from the paper-specific evaluation problem.

The appendix gives the quadratic restriction derivation in A.1, the projection formulas and singular-covariance argument in A.2, the fitted residual correction and episode-level concentration calculation in B, and the modular execution and rollout derivation in D. We relabeled the "Proof idea" passages, which preceded the full arguments, as overviews, added a proof index to the appendix overview, and expanded the Bernstein range scaling and union-bound bookkeeping (proof of Theorem 1) and the module-wise inequalities (proof of Lemma D.1). Section 5 now identifies Theorem 1 as the frozen-cost result and Corollary B.1 as the guarantee for a deployed restricted predictor fixed before D2D_2.

We are glad to clarify any derivation.

AuthorsOpenReview ↗

Follow-up on comparisons, readability, and proofs

Thank you again for your feedback. In Section 7.5 and Appendices E.12–E.14, we added an input–output community decomposition baseline (Tang and Daoutidis, 2018) and a comparison of four partition-selection criteria across all three benchmarks. At the same 30-coupling budget on TEP, the community baseline and the exact-refit validation reference tie at a test NRMSE of 0.4614. Across all three benchmarks, the exact-refit reference had the highest median rank correlation with test loss within groups of partitions with the same module count and similar coupling budgets.

To improve readability, we added a two-unit example that walks through freezing, reweighting, and exact refitting before introducing the formulas. We also added a proof index and expanded the derivations for Theorem 1 and Lemma D.1. We’d be happy to clarify any step that remains difficult to follow.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.