acceptodds
Under review as a conference paper at ICLR 2027

Jump-Augmented Reservoir Computing for Trajectory Forecasting and Measurement Correction

Abstract

Echo state networks (ESNs) replace gradient-based recurrent training with a fixed, randomly generated reservoir and a single linear readout. We introduce hybrid reservoirs, in which we augment the dynamics of the reservoir state to allow for discrete, event-triggered jumps that reset the reservoir state instantaneously. Deciding when and where to jump is a challenging problem that is usually application-specific. In this work, we provide a general formulation of hybrid reservoirs and showcase their potential on trajectory forecasting and measurement correction. For both applications, a well-chosen jump map lets the reservoir be reset into the right region of its state space, improving performance over a standard, jump-free ESN without sacrificing the closed-form, replay-free training that makes reservoir computing attractive in the first place.

Primary area:
learning on time series and dynamical systems
Keywords:
Reservoir Computing, Echo State Networks, Hybrid Systems

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_qd2jOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces hybrid reservoir augment the dynamics of the reservoir state to allow for discrete, event-triggered jumps that reset the reservoir state instantaneously, verified on two applications successfully.

Critical strengths

The correction via local linearization in Section 4.2.2 is a useful component.

Critical weaknesses

Experimental verification can be more solid, add SOTA comparisons e.g. with

Nóvoa, A., & Magri, L. (2025). Online model learning with data-assimilated reservoir computers. arXiv:2504.16767.

Chen, R. T. Q., Amos, B., & Nickel, M. (2021). Learning neural event functions for ordinary differential equations. ICLR 2021.

Gauthier, D. J., Bollt, E., Griffith, A., & Barbosa, W. A. S. (2021). Next generation reservoir computing. Nature Communications, 12, 5564.

Questions for authors
  1. Fig. 4 says 22 and 27 successful HRC runs, but Fig. 8 says 17/30 and 29/30. Which one is right?
  2. Eq. 9 has no WJ,bJW_J, b_J, but Table 2 tunes them. What kind of jump map did you actually use?
1 comment
AuthorsOpenReview ↗

Official Comment by Authors

We thank the reviewer for the positive assessment of the local linearization correction introduced in Section 4.2.2 and for suggesting additional comparisons with state-of-the-art methods. To facilitate the revision process, the changes in the manuscript are colored blue and the sentences referring to them are colored bold in this response. Now we address reviewers’ comments and concerns:

Additional baselines and experimental evalutation

We have implemented Next Generation Reservoir Computing (NG-RC) [Gauthier et al.,2021] as an additional baseline. We have also extended the experimental evaluation to include noisy measurements in Experiment 2. The additional NG-RC results provide a further point of comparison with an alternative reservoir computing architecture. Nevertheless, the proposed Hybrid Reservoir consistently outperforms NG-RC in terms of predictive performance. We are also exploring the implementation of the approaches proposed by Nóvoa and Magri (2025) and Chen et al. (2021). Because of the limited time, we have not completed them for the revised version, but will work towards including them in the final version if the paper is accepted. We believe that the additional baseline and the experiments with noisy measurements provide further validation of our approach and address the reviewer’s concerns regarding experimental evaluation

Q1. Fig. 4 says 22 and 27 successful HRC runs, but Fig. 8 says 17/30 and 29/30. Which one is right?

We thank the reviewer for pointing out this discrepancy. The revised manuscript corrects the typo and ensures consistency across all figures. The reported numbers have also changed following the expanded grid-search strategy suggested by Reviewer whiJ.

Q2. Eq. 9 has no WJ , bJ , but Table 2 tunes them. What kind of jump map did you actually use?

We thank the reviewer for pointing out this discrepancy. We used different connectivity and bias matrices for the jump set as stated in Table 2. We corrected the typo in Eq. (9) in the revised manuscript and added a brief clarification in the main text to ensure consistency with the jump map parameters reported in Table 2.

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

This is a paper in the field of reservoir computing. It proposes an approach based on hybrid reservoirs, where the internal reservoir state can be reset by discrete jumps based on the occurrence of an external event. From my understanding, the paper extends the formulation provided in (McCreesh et al., 2026), in such a way that jumps depend on the input signal, and the jump map is chosen depending on the task. The paper presents 2 main applications: on hybrid dynamics forecasting (impact oscillator) and trajectory forecasting correction (Lorenz and Rossler systems), showing improvements compared to conventional ESNs.

Critical strengths
  • I like the original idea behind the paper of correcting the reservoir state itself, it looks simple and cheap to implement. It keeps the closed-form training characteristic of reservoir computing and shows gains over the input-only correction.
  • Readability of the paper, which is well written. I appreciated the presentation of the hybrid system formalism for the reservoir computing community.
Critical weaknesses
  • To the best of my understanding of the paper, the comparison in the first application is not completely fair. In fact the jumps of the model proposed are triggered by the ground-truth events of impact, thereby the model is actually receiving in input information that is not given to the standard ESN baseline. In addition, in appendix A.2.1 it is reported that the hyperparameters are selected as the configuration that has the lowest closed-loop forecast error over 2000 steps, that seems to be the same forecast length that is used for the evaluation, and no validation split is mentioned. Thereby, the resulting gap with the standard ESN might be due to the setup of the experiment.

  • Significance of the work appears limited in temrs of novelty and applications. In particular, novelty appears limited, and incremental compared to the work in McCreesh et al. (2026). Moreover, the paper is not actually comparing the result of the proposed approach with their hybrid reservoir, and thereby the advantage of the proposed extension seems not analyzed. The experimental evaluation is also limited to small synthetic noise-free systems.

Questions for authors
  • Regarding the fairness of comparison with ESNs, in the impact oscillator task are the jumps actually triggered by the external true events? Could you please explore and report the performance of a standard ESN with the external event information as another input?
  • Please, clarify how the hyperparameters have been tuned in all the applications. Have you used a separate validation split for model selection? Would it be possible to enlarge the grid (which is rather limited for the reservoir based hyperparameters' values) or to employ a cheaper strategy (e.g., random or bayesian) on a larger set of values?
  • What happens with noisy or partial measurements? How would the reservoir correction behave?
  • Please compare experimentally your method with the original hybrid reservoir of McCreesh et al. (2026).
4 comments
AuthorsOpenReview ↗

Official Comment by Authors 1/2

We are grateful to the reviewer for the thorough review and valuable comments, which have helped improve the paper. To facilitate the revision process, the changes in the manuscript are colored blue and the sentences referring to them are colored bold in this response. Now we address reviewers' comments and concerns:

Q1. Regarding the fairness of comparison with ESNs, in the impact oscillator task are the jumps actually triggered by the external true events? Could you please explore and report the performance of a standard ESN with the external event information as another input?

Thanks for pointing this out. The triggering mechanism was unclear in the original manuscript, so we have rewritten the explanation in Section 4.1.1. We clarify that neither the standard ESN nor the hybrid reservoir (HRC) receives information from the true trajectory. Instead, jumps are triggered whenever the self-predicted state x^\hat{x} crosses a prescribed threshold. Thus, forecasting remains fully autonomous, without access to the true events.

Additionally, following the reviewer's suggestion, we implemented an alternative hybrid reservoir architecture, denoted as HRC-IA (Input Augmented) in the table. Rather than incorporating a dedicated jump map, this implementation encodes event information through an additional input to the reservoir. Specifically, an event indicator becomes active whenever the predicted state x^\hat{x} enters the prescribed jump set. This indicator acts as an additional input that modifies the reservoir state through the input-weight matrix, thereby introducing event-triggered changes in the reservoir dynamics. Hence, HRC-IA constitutes an alternative implementation of a Hybrid Reservoir Computing architecture, in which discrete events are encoded through input augmentation rather than an explicit jump map.

The results of this additional experiment are reported in the following table:

ModelMean total RMSEMean x RMSEMean v RMSE
RC6.53520.40949.1887
RC-continuous0.78190.7819—
HRC-AI0.29350.21770.3441
HRC0.24440.18210.2817

As can be observed, HRC-IA achieves performance relatively close to the proposed HRC, demonstrating that event information provided through input augmentation can also capture part of the benefits of hybrid reservoir dynamics. Nevertheless, the use of an explicit jump map offers greater flexibility in modeling discrete transitions, as it allows the jump dynamics to be specified independently of the random directions imposed by the input-weight matrix. Together with the improved performance observed in our experiments, this motivates retaining the explicit jump-map formulation as the proposed HRC architecture.

Q2. Please, clarify how the hyperparameters have been tuned in all the applications. Have you used a separate validation split for model selection? Would it be possible to enlarge the grid (which is rather limited for the reservoir based hyperparameters' values) or to employ a cheaper strategy (e.g., random or bayesian) on a larger set of values?

ESNs can be particularly sensitive to hyperparameter selection. To ensure a fair comparison, we independently tuned the hyperparameters of each approach using the same grid-search procedure, selecting the best-performing configuration within the prescribed search space for each method. Following the reviewer's suggestion, we have further expanded and refined the hyperparameter grid in the revised manuscript, increasing both its size and granularity to provide a more comprehensive exploration of the parameter space. This additional tuning has led to substantially improved results across the evaluated methods. We thank the reviewer for this valuable suggestion, which has helped strengthen our experimental evaluation.

Furthermore, as discussed in Section 4.1.2, our results show that jump-augmented ESNs exhibit greater robustness to hyperparameter variations. This represents an additional advantage of the proposed method, as it reduces the sensitivity of model performance to hyperparameter selection.

Q3. What happens with noisy or partial measurements? How would the reservoir correction behave?

In the revised manuscript, we have added a robustness analysis to evaluate the effect of measurement noise on the proposed reservoir correction, together with a discussion of the results (Section B.2.1) (to be continued)

AuthorsOpenReview ↗

Official Comment by Authors 2/2

(continuation)

Q4. Please compare experimentally your method with the original hybrid reservoir of McCreesh et al. (2026).

We thank the reviewer for this suggestion. McCreesh et al. (2026) introduced the concept of hybrid reservoirs, and their study is primarily theoretical, focused on identifying conditions that guarantee the Echo State Property. They did not, however, explore the versatility the framework affords in specifying the jump sets and jump maps, which is the focus of the present work. The particular hybrid reservoir considered by McCreesh et al. corresponds to one specific choice of jump set and jump map. Beyond this, we exploit the freedom in designing jump sets and jump maps in two ways. We propose a novel mechanism that incorporates state measurements through the jumps, and we provide an experimental evaluation of how different specifications affect performance, including systematic comparisons with standard ESN baselines and an analysis of hyperparameter sensitivity.

Thank you very much for the nice response and for the revised paper.

I thank the authors for the response and the revised manuscript. I notice that the added experiment with included event information as an additional input shows that the error of the ESN goes pretty close to the one of the HRC (0.29 vs 0.24). This observation makes me think that it is possible that a large part of the improvement is actually coming from the knowledge of the jump, and I would suggest the authors to comment on this in the manuscript.

Regarding the model selection, the tuning protocol now addresses my concerns, even though some values selected from the search seem to be at the edge of the explored grid.

Finally, regarding novelty, I acknowledge the response from the authors, but I also notice that actually the advantage of the proposal against the hybrid reservoir in McCreesh et al. (2026) is not shown through experimental evidences (for instance, on the bouncing ball setting in that paper, but this is just one example). So, my novelty concerns remain.

AuthorsOpenReview ↗

Authors’ Response

Event information in the forecasting performance

We thank the reviewer for this observation. As suggested, we have expanded the discussion in the revised manuscript to clarify the role of event information in the forecasting performance. In particular, by providing an event indicator as an additional input, the ESN can undergo event-triggered changes in its internal state through the input matrix, effectively becoming a hybrid reservoir. Consequently, it recovers a substantial part of the performance gains achieved by the HRC, as reflected in the reduced prediction error (0.29 vs. 0.24). This observation highlights the importance of event information for forecasting hybrid dynamics and illustrates the central contribution of hybrid reservoirs: extending conventional ESNs to allow discrete jumps in their internal states, either triggered by specific system events or induced through discontinuities in the input signal. However, the HRC provides an explicit mechanism to incorporate these transitions through its jump dynamics, rather than relying on event indicators entering through the input matrix. We have clarified this point in the revised manuscript.

Hyperparameter tunning

We thank the reviewer for this comment. In response, we have further expanded the hyperparameter search grids to ensure that the selected configurations for the baselines are not constrained by the grid boundaries, allowing for a fairer comparison across methods. Additionally, we have extended the evaluation to multiple oscillator configurations, selecting hyperparameters that provide robust performance across variations in the system parameters, such as mass and coefficient of restitution, rather than relying on a single configuration. These additional evaluations strengthen the robustness and fairness of the experimental comparison.

Novelty with respect to McCreesh et al (2026)

First, we have substantially strengthened the theoretical contribution of the manuscript by adding Theorem 1 a novel ESP result for hybrid reservoirs with infinitely many jumps (that was not present in our original manuscript). In McCreesh et al. (2026), the ESP guarantees rely on restrictive settings, essentially considering either finitely many jumps, after which the system eventually evolves continuously, or hybrid systems whose continuous and discrete dynamics satisfy the ESP independently. In contrast, our new result addresses genuinely hybrid dynamics, where neither the flow nor the jump dynamics needs to be independently contractive, provided their combined effect ensures contraction over time. This considerably broadens the class of hybrid systems for which the ESP can be established, and thus the applicability of the proposed framework.

We acknowledge that further theoretical questions remain open. In particular, our analysis does not yet cover general switching scenarios without a common dwell-time bound across executions, for which establishing the ESP may require tools such as graphical convergence of hybrid solutions [1][2]. We consider these interesting directions for future work.

Second, Application II demonstrates the broader applicability of the hybrid reservoir formalism through a novel approach to incorporating measurement information during forecasting, going beyond the applications considered in McCreesh et al. (2026).

Finally, we provide substantially more extensive experimental validation, including additional baselines, more challenging task, and a more thorough hyperparameter optimization and evaluation procedure. Together, these contributions extend the theoretical foundations, demonstrate new applications, and provide a more comprehensive experimental assessment of hybrid reservoir computing beyond the original McCreesh et al (2026).

[1] Li, Y., & Sanfelice, R. (2015). On Necessary and Sufficient Conditions for Incremental Stability of Hybrid Systems Using the Graphical Distance Between Solutions. UC Santa Cruz. http://dx.doi.org/10.1109/cdc.2015.7403093

[2] Y. Li, S. Phillips, and R. G. Sanfelice, "Basic properties and characterizations of incremental stability prioritizing flow time for a class of hybrid systems," Systems & Control Letters, vol. 90, pp. 7–15, 2016. doi: [10.1016/j.sysconle.2015.12.015] https://doi.org/10.1016/j.sysconle.2015.12.015

Reviewer_b5tLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Paper proposes hybrid reservoir computing, extending standard echo state networks with event triggered jumps in reservoir states. Two applications are considered: forecasting discontinuous dynamics and correcting autonomous predictions using intermittent measurements. Experiments on an impact oscillator, Lorenz, and Rössler systems demonstrate improvements over conventional ESNs while retaining closed form readout training.

Critical strengths

Introducing event triggered state corrections into reservoir dynamics is a useful extension of conventional ESNs. The Jacobian based correction in Eqution 11 is particularly interesting because it incorporates measurement information without retraining or replaying previous observations. Experiments across multiple random seeds demonstrate encouraging improvements in forecasting accuracy and robustness.

Critical weaknesses

Mmain concern is that the improvements depend strongly on externally specified information. In the impact oscillator experiment, the jump condition is derived from the true physical system, while the standard ESN baselines do not receive comparable event information. Thus, the comparison does not fully isolate the benefit of the proposed reservoir architecture. Theoretical justification is also incomplete. ESP is not established for indefinitely repeated corrections, and the local Jacobian update in Equation 11 does not guarantee that the corrected state lies on the reservoir's nonlinear reachable manifold. Also, the correction experiments assume noise free measurements, and robustness to measurement errors is not evaluated. There is also an apparent inconsistency between Equation 9, which uses the original reservoir weights, and Appendix Table 2, which introduces separate jump map parameters.

Questions for authors

How are jump events detected during autonomous forecasting, and would the advantage remain against an event aware ESN baseline? Can the authors clarify the discrepancy between Equation 9 and the jump map parameters in Table 2? What guarantees prevent repeated Jacobian corrections from violating the echo state property, particularly under noisy or infrequent measurements?

1 comment

We thank the reviewer for the constructive criticism and appreciate the positive comments on the Jacobian-based correction. To make the revisions easy to follow, changes in the manuscript are highlighted in blue, and sentences in this response that refer to them are marked in bold. Below, we address each of the reviewer’s comments and concerns.

Measurement error robustness analysis

In the revised manuscript, we have added a robustness analysis to evaluate the effect of measurement noise on the proposed reservoir correction, together with a discussion of the results (Section B.2.1).

Q1. How are jump events detected during autonomous forecasting, and would the advantage remain against an event aware ESN baseline?

Thanks for pointing this out. The triggering mechanism was unclear in the original manuscript, so we have rewritten the explanation in Section 4.1.1. We clarify that neither the standard ESN nor the hybrid reservoir (HRC) receives information from the true trajectory. Instead, jumps are triggered whenever the self-predicted state x^\hat{x} crosses a prescribed threshold. Thus, forecasting remains fully autonomous, without access to the true events.

Additionally, following the reviewer's suggestion, we implemented an alternative hybrid reservoir architecture, denoted as HRC-IA (Input Augmented) in the table. Rather than incorporating a dedicated jump map, this implementation encodes event information through an additional input to the reservoir. Specifically, an event indicator becomes active whenever the predicted state x^\hat{x} enters the prescribed jump set. This indicator acts as an additional input that modifies the reservoir state through the input-weight matrix, thereby introducing event-triggered changes in the reservoir dynamics. Hence, HRC-IA constitutes an alternative implementation of a Hybrid Reservoir Computing architecture, in which discrete events are encoded through input augmentation rather than an explicit jump map.

The results of this additional experiment are reported in the following table:

ModelMean total RMSEMean x RMSEMean v RMSE
RC6.53520.40949.1887
RC-continuous0.78190.7819—
HRC-IA0.29350.21770.3441
HRC0.24440.18210.2817

As can be observed, HRC-IA achieves performance relatively close to the proposed HRC, demonstrating that event information provided through input augmentation can also capture part of the benefits of hybrid reservoir dynamics. Nevertheless, the use of an explicit jump map offers greater flexibility in modeling discrete transitions, as it allows the jump dynamics to be specified independently of the random directions imposed by the input-weight matrix. Together with the improved performance observed in our experiments, this motivates retaining the explicit jump-map formulation as the proposed HRC architecture.

Q2. Can the authors clarify the discrepancy between Equation 9 and the jump map parameters in Table 2?

We thank the reviewer for pointing out this discrepancy. We used different connectivity and bias matrices for the jump set as stated in Table 2. We corrected the typo in Eq. (9) in the revised manuscript and added a brief clarification in the main text to ensure consistency with the jump map parameters reported in Table 2.

Q3. What guarantees prevent repeated Jacobian corrections from violating the echo state property, particularly under noisy or infrequent measurements?

We thank the reviewer for raising this important point. Although the Echo State Property (ESP) can be established for a finite number of Jacobian corrections (Theorem IV.1 in [1]), our current results do not guarantee its preservation under infinitely many corrections.

We believe this extension is feasible using stability results from hybrid systems theory. In particular, the increase-balanced-by-decrease condition in the Hybrid Lyapunov Theorem (Theorem 3.19 in [2]) provides a framework for establishing stability when increases in a Lyapunov function due to discrete jumps are compensated by sufficient decreases during continuous flows. Developing the corresponding formal result and proof requires additional analysis beyond the time frame of this revision. Nevertheless, we are confident that this framework can yield meaningful theoretical guarantees, which we intend to further develop and incorporate into the final version if the paper is accepted.

[1] M. McCreesh, A. Marzabal, A. Ganlath, R. Gupta, J. Sanchez, and J. Cort´es. Reservoir com- puting with hybrid systems: establishing the echo state property. In American Control Conference, pp. 2838–2843, New Orleans, LA, May 2026

[2] R. G. Sanfelice. Hybrid Feedback Control. Princeton University Press, Princeton, NJ, 2021

Comments on the paper

1 comment
AuthorsOpenReview ↗

Major Changes Following Reviewer Discussion

Following the rebuttal and the subsequent discussion with the reviewers, the manuscript has been substantially revised and strengthened. The main changes include:

Theoretical contributions (Theorem 1): We have strengthened the theoretical analysis by introducing a new Echo State Property (ESP) result for hybrid reservoirs with infinitely many jumps. Unlike previous results (McCreesh et. al 2026) that rely on restrictive assumptions on the continuous and discrete dynamics, our result establishes conditions under which their combined effect guarantees contraction, even when neither component is independently contractive.

Expanded experimental evaluation: We have considerably extended the hyperparameter search, refined the tuning procedure, and evaluated the methods across multiple configurations of the impact oscillator to assess robustness to changes in system parameters. We have also included next generation-RC as an additional baseline and added experiments analyzing the robustness of reservoir corrections to measurement noise.

Clarification of the event-triggering mechanism: We have clarified that events during autonomous forecasting are detected from the model's own predicted state, rather than from ground-truth trajectories.

Overall, we believe these revisions significantly strengthen the theoretical contributions, experimental validation, and clarity of the manuscript, while directly addressing the main concerns raised during the review process.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.