Structured Neural-PI Control with End-to-End Stability and Output Tracking Guarantees

We study the optimal control of multiple-input and multiple-output dynamical systems via the design of neural network-based controllers with stability and output tracking guarantees. While neural network-based nonlinear controllers have shown superior performance in various applications, their lack of provable guarantees has restricted their adoption in high-stake real-world applications. This paper bridges the gap between neural network-based controllers and the need for stabilization guarantees. Using equilibrium-independent passivity, a property present in a wide range of physical systems, we propose neural Proportional-Integral (PI) controllers that have provable guarantees of stability and zero steady-state output tracking error. The key structure is the strict monotonicity on proportional and integral terms, which is parameterized as gradients of strictly convex neural networks (SCNN). We construct SCNN with tunable softplus-$\beta$ activations, which yields universal approximation capability and is also useful in incorporating communication constraints. In addition, the SCNNs serve as Lyapunov functions, giving us end-to-end performance guarantees. Experiments on traffic and power networks demonstrate that the proposed approach improves both transient and steady-state performances, while unstructured neural networks lead to unstable behaviors.

Paper

References (57)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer nsH54/10 · confidence 4/52023-07-01

Summary

This paper considers the problem of learning feedback controllers for multi-input, multi-output (MIMO) dynamical systems. At the core of the paper, a PI-like nonlinear control structure is proposed. Specifically, the controller consists of two NN, one which receives the set-point error (denoted as "P-part") and another one that receives its integral ("I-part") as an input. For the NNs, a special structure is proposed that ensures strict convexity of the NN. This structure serves as the basis for proving stability and zero steady-state error of the closed loop system employing this controller, irrespective of the parameters used in the NN. Hence, these guarantees can be given a-priori irrespective of the training. Training of the NN then aims at optimizing transient behavior. Finally, two numerical examples of MIMO control systems (vehicle platooning, power network) are presented.

Strengths

S1) Developing NN control architectures that are guaranteed to satisfy key control properties by design (here: stability and steady-state optimality) is very relevant. The combination of PI-like structure with strictly convex NN seems novel. S2) Leveraging passivity properties is an interesting idea for analyzing and achieving system-wide properties. This is well-known in the control community, but should also be relvant for the NeurIPS community when considering distributed learning problems. S3) I like the idea of having a system design that is stable and comes with base performance by design and without communication, and then improving performance further when communication is available.

Weaknesses

W1) The introduction makes some imprecise or misleading statements about the state of the art as well as technical aspects. All statements should be carefully checked and (where needed) be backed up with proper argumentation or references. Examples: * ll. 33-34: "Currently, learning-based approaches mostly parameterize nonlinear controllers as neural networks and train them to optimize performance". While NN are a popular function approximator in learning-based control, they are actually used in different flavors (model learning, controller parametrization, surrogate model, etc.). Furthermore, there are many other ML techniques that are also used, such as Gaussian processes, which are particularly popular when it comes to guarantees (examples: A Lederer, J Umlauft, S Hirche, "Uniform error bounds for Gaussian process regression with application to safe control"; C Fiedler, CW Scherer, S Trimpe, "Learning-enhanced robust controller synthesis with rigorous statistical and control-theoretic guarantees") Hence, the characterization of the area of learning-based control as described in this above sentence is too narrow in my opinion. * ll. 47-49: "...often overlooking the steady-state behavior (cost at the equilibrium state)." The statement is not supported by references. In fact, I would disagree with it, as there are works that specifically focus on the long term performance of controllers or RL policies (examples: A Doerr et al, "Optimizing long-term predictions for model-based policy search"; N Lambert et al, "Learning accurate long-term dynamics for model-based reinforcement learning") * ll. 49-51: The description of classical P- and I-control is inaccurate. While it is correct that I-control is often used in practice to make steady-state error zero, the actual behavior depends on the controller *and* the dynamical process. Depending on the process dynamics (e.g., if it has an integrator itself), also a P-controller may lead to zero steady-state error. Also, the statement that P-control is for the transient, but disregarding steady-state performance is inaccurate. P-control also affects the steady-state in general. Again, the general description of how P- and I-control are often used in practice is not wrong, but the authors should more carefully word these sentences to avoid technical inaccuracies. * The statement of the main research question that is addressed in this paper is too general to be representative of the paper's contribution: "Can we learn nonlinear controllers that guarantee transient stability and zero steady-state error for MIMO systems?" (ll. 55-56). First, the authors state themselves in the next sentence that this question can't be answered for general nonlinear system. Second, there are existing works in learning-based control that focus specifically on MIMO PI-control (see below). IMO, the research question should thus be rephrased to already clarify the scope of this work. * ll. 67-68, "This class of functions is too restrictive to use for controller design" -- Why? Appropriate reasoning or reference is missing to support the statement. W2) Related work: There is no explicit discussion of related work as a separate section/subsection, which is not a requirement of course, but could be helpful. Some references are mentioned in the introduction, but IMO the following aspects require further discussion: * The control problem and the objectives that are stated are super general in control (stability, zero steady-state error, fast transient). Hence, there will be a lot of works in learning-based control addressing similar problems. While the paper does mention references and related work, I do not find this discussion very structured and have trouble pin-pointing what exactly is the difference to prior art and thus contribution of this paper. Just saying, for example, that "guaranteeing stability and output tracking" (l. 144) is not helpful in this regard as the statement is too general and not specific to the contribution at hand. It might help to first clearly specify the problem that is addressed, and afterwards discuss the related work specifically to the problem. * Given that the consideration of MIMO systems is a main aspect of this work, there should be a review specific to this. Indeed, there are existing works in learning-based control that focus specifically on MIMO PI-control (e.g., A Doerr et al, "Model-based policy search for automatic tuning of multivariate PID controllers", I Carlucho, M De Paula, GG Acosta, "An adaptive deep reinforcement learning approach for MIMO PID control of mobile robots", and many other works on "self-tuning" for MIMO systems). Overall, I do not get a clear idea how this work compares to the state of the art. W3) Specification of controller structure: The control law u = g(y) is not precisely specified. Concretely, is g(.) a static function, or does it itself involve states and memory (i.e., is it itself dynamic)? This is critical, as dynamic controllers (such as combining a state estimator and state-feedback controller) are more powerful than static maps in general. Later on, it seems that static NN are used as components of g(.). However, the controller also involves an integrator, which would not be captured by a static function g(.) and is actually a dynamic component. If, on the other hand, dynamic output-feedback controllers are admissible, one could presumably learn much richer control behavior. IMO, it is ok to restrict the controller structure (from what I understand, to an integrator + static map), but it needs to be clearly and correctly stated. W4) I am wondering whether it is really justified to call the proposed structure a proportional + integral (PI) controller. In particular, the term that is called "proportional" isn't really proportional to the error for a general nonlinear map. I can see that what the authors propose is a generalization of well-known PI-control structure. However, I find it a bit misleading to call this a "proportional control". In fact, p(-y+ybar) is a general nonlinear function of the tracking error. Hence, it will in general not be proportional to the tracking error. Please elaborate, why you call the architecture "PI", and ideally provide references if this is a common naming. W5) I do not find the notion of "end-to-end" guarantees conclusive and well justified. The authors use this notion to denote the fact that they can provide certain guarantees a-priori before training (by combining a certain system class (passive) with the proposed controller architecture). As far as I know, the term "end-to-end" is not commonly used for this (if you have references where it is used in this way, please provide them). If a controller has theoretical guarantees, these are most of the time "by construction", so I don't understand why this is called "end-to-end". * Minor: In the introduction, it is not yet clear what the authors mean when mentioning "end-to-end guarantees" on l. 80. It should be explained when introduced for the first time. W6) The presentation of experiments in the main paper is not self-contained and, IMO, insufficient. There needs to be a basic description of the problem in the main paper, which can be understood without consulting the supplementary material. But this is currently missing. Because of this, one cannot judge how challenging the control problem actually is, and whether sophisticated nonlinear control is actually needed. Furthermore, the verification of assumptions is critical because a main claim of the paper are the theoretical guarantees. This should be in the main paper. As these are essential aspects of the contribution, the experimental section isn't self contained. If the authors cannot fit all essential components, then a conference might be the wrong format for this manuscript. W7) Experimental results are not fully conclusive; they compare against a standard neural network, but more classical controller structures as baseline are missing in the main paper. Thus, I find the results only partially convincing for showing the benefit of the proposed Neural-PI controller. * It remains unclear whether the chosen problems require sophisticated nonlinear control such as the proposed Neural-PI controller. First, the problem description does not convey this (see above). Second, in the supplementary material, the authors report the results also of a standard linear PI, whose parameters have been optimized similarly to the optimization of the NNs. Apparently, the linear PI achieves almost the same performance as the Neural-PI for vehicle platooning, and very close performance for the power grid. The results suggests that at least for these systems, there is no significant benefit of the proposed and more involved Neural-PI structure. Hence, the result are not conclusive regarding the proposed architecture in this regards. * Baseline comparisons (i.e., the linear PI from the supplementary, and maybe also a model-based controller such as LQR) should be included in the main paper to judge the performance of the proposed Neural-PI. * I find it surprising that the DenseNN cannot stabilize the vehicle platooning system and actually has some significant deviation at the end of the horizon (Fig. 5(b), 15 sec). Such deviation should also be visible in the objective function, so I would expect that the NN (if trained enough) can accommodate for this. * A discussion of assumptions and theoretical guarantees is missing in the main paper. In ll. 177/178, it is claimed that in the experimental section, it will be discussed how EIP (the main assumption) can be verified if the system is unknown. While I agree that this discussion is very important with regards of the claimed contribution of this paper, I could not find this in Sec. 6. * The above points hold for both examples (platooning, power grid). Minor: * l 192: As usually, p is a gain (e.g., a real number), it would be good to emphasize when introducing it that, here, it is meant as a function of the tracking error.

Questions

Q1) Why is it justified to call the structure in (4) a PI-controller, especially, why is it justified to call the first component "proportional"? Q2) What are the 2..3 papers in existing literature that are closest to your approach? And what main limitation of these does you approach address? Q3) In the experiments, the linear PI performs rather well. Why is the more involved Neural-PI controller even needed there? Q4) Do you have an example, where the Neural-PI controller clearly outperforms linear approaches? Q5) Why did you decide to not include the experimental results of the linear PI in the main paper?

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

Limitations (such as no consideration of disturbance/noise) are mentioned briefly in the conclusion, but not discussed in detail. I would have liked to see a discussion on: what happens when there is noise/disturbance in the sensing, the process, and/or the data; do the authors expect that their controllers can be transferred to the real world, or what hurdles do they see? I don't see any particular potential negative societal impact.

Reviewer F58a5/10 · confidence 3/52023-07-03

Summary

This paper proposes a method for optimal control of dynamical systems using a neural network-based controller that guarantees stability and output tracking. The key structure of the proposed neural proportional-integral (PI) controller is based on the strict monotonicity of the proportional and integral terms, which are parameterized as the gradient of a strictly convex neural network. The paper also introduces tunable activations for universal approximation capabilities and communication constraints. Experiments on traffic and power networks show significant improvements in both transient and steady-state performance. The contribution of this paper is mainly in the field of control theory and its application.

Strengths

The strength of this paper is that it presents a method for optimal control using a neural network-based controller with provable stability guarantees. Using a strictly monotonic function parameterized by a strictly convex neural network can improve control performance and output tracking. Experiments on traffic and power networks show significant improvements in both transient and steady-state performance.

Weaknesses

The multiple equilibria problem has not been fully addressed in real-world applications. This limitation limits the applicability of the proposed method to systems with general nonlinear systems.

Questions

In real applications, how to judge the passivity of a large-scale nonlinear system is a possible problem to extend the proposed approach.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

The paper does not explicitly discuss the scalability of the proposed method to larger, more complex systems. This paper does not provide a comparison with other state-of-the-art control methods, which helps to better understand the advantages and disadvantages of the proposed method.

Reviewer F58a2023-08-19

The authors' responses are reasonable. I recommend accepting the paper.

Authorsrebuttal2023-08-19

We are grateful to the reviewer for the response and encouraging comments.

Reviewer aJSd6/10 · confidence 5/52023-07-04

Summary

This paper presents an approach to control multi-input, multi-output dynamical systems, introducing a neural network-based PI controller offering provable stability and output tracking guarantees. The neural Proportional-Integral controller is based on strict monotonicity parameterized by strictly convex neural networks (SCNN). The strict monotonicity of the network gives the convergence guarantee by construction. Testing on traffic and power networks showed improved transient and steady-state performances compared to unstructured networks.

Strengths

It is interesting to improve PI control using neural networks. The proposed framework not only keeps the structure (which is well-validated across decades) and robust tracking performance of PI controllers, but also allows optimizing the tracking performance by learning the parameters in g and r functions. The use of monotonic neural networks is also critical. The experiment results demonstrate its advantages over regular fully connected neural networks.

Weaknesses

- The authors must have known that carefully tuned PI controllers already work very well. Why didn’t they compare in the experiments? It is common that controllers built upon regular fully connected layers will diverge as time progresses, but this is not a problem in simple regular PI controllers. How well the proposed method is, compared to a carefully tuned classical PI controller? - Even if the proposed framework is better than traditional PI control in the transient phase, how can the authors guarantee its performance in situations outside the training data distribution? In real world applications where data collection can be expensive, if we only have 1000 data points, would the authors choose to train a neural network (this could be dangerous if only have limited data) or tune a classical PI controller using their extensive domain knowledge of the system dynamics (and use the 1000 data points for tuning & experiment). - In summary, this study is interesting and novel, but its importance / practical contribution to the control field might be minor.

Questions

See weaknesses.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

3 good

Contribution

2 fair

Limitations

See weaknesses.

Reviewer eDxT7/10 · confidence 4/52023-07-06

Summary

The paper "Structured Neural-PI Control with End-to-End Stability and Output Tracking Guarantees" presents a well-motivated and technically sound approach to address the lack of provable guarantees in neural network-based controllers. The combination of structured neural-PI control, equilibrium-independent passivity, and SCNNs with tunable softplus-β activations demonstrates promising results in terms of stability, output tracking, and improved performance.

Strengths

The paper is clear, well written and accurate in the mathematical control details. The method leverages a specific kind of NN and the approach is clean and seems to hold great potential, at least to an applicable systems class.

Weaknesses

The proposed approach relies on the assumption of equilibrium-independent passivity, which may limit its applicability to a specific class of physical systems. The paper should explicitly discuss the range of systems to which the approach is applicable and highlight any potential limitations or challenges in extending it to other domains.

Questions

1. In Definition 2, is it critical to demand zero tracking error? That can be augmented easily with a prefilter or an integrator. Even though the controller has an integrator, the system may be of higher type). 2. In 3.1 it is stated that it is desired to optimize the transient response before the system settles on the tracking point. What are the specifications? Maximum overshot? Settling time? Are there minimum requirements?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

Not applicable

Reviewer Cp4m6/10 · confidence 4/52023-07-09

Summary

This paper propose a neural-PI controller based on the newly designed strictly convex neural network. Thanks to the EIP condition of the considered systems, the neural-PI controller naturally satisfies the stability and output tracking guarantee with the special structure of the SCNN. The efficacy of the proposed method is demonstrated in several physical models.

Strengths

[1] The proposed neural-PI controller satisfies the stability guarantee by construction of the SCNN instead of adding regularization term to the loss function, which is a hugh improvement to the existing learning controller commuity. [2] The proposed method does not require the knowledge of the equilibria of the intrisic dynamics but can still provide stability guarantee of the unknown equilibria. [3] The experiments take communication cost into consideration, which is realistic and instructive. [4] The presentation of the paper is good, the workflow in the figures are clear and helpful.

Weaknesses

[1] The current framework is restricted in the system with EIP condition. [2] The authors did not compare the proposed method with the existing methods such as RL, QP and MPC. [3] The training policy is naïve, the learner should solve the controlled ODE to find the controlled trajectory and then update the parameters.

Questions

[1] Since the objective functional incurs the integral of the controlled ODE, can this method scale to the high-dimensional task? [2] Can this method outperform the existing methods (RL such as PPO or MPC) in terms of training speed or test performance?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

[1] The current neural method can only handle deterministic systems with EIP condition, which cannot adapt to the real-world scenarios where noise is unavoidable. [2] The authors did not provide numerical comparison with the existing methods.

Reviewer aJSd2023-08-10

Thanks for the Response

I would like to thank the authors for their detailed response. I don't have further questions, and will change to accept.

Authorsrebuttal2023-08-15

Thank you for your response and for changing the scores to accept. We are happy that we’ve addressed the questions.

Reviewer eDxT2023-08-10

Thanks for the response

I would like to the authors for the response. I have no further questions.

Authorsrebuttal2023-08-15

Thank you for your response and for changing the scores to accept. We are happy that we’ve addressed the concerns.

Reviewer nsH52023-08-16

I have read the rebuttal, I have no specific follow-up questions. In the rebuttal, the authors have well addressed some of my previous points (e.g., clarification on the contribution). Other aspects will require significant rewriting of the manuscript to an extent that (in my opinion) should require a re-review. Nonetheless, I changed my score reject -> borderline reject, as the weaknesses are less severe overall after the rebuttal.

Authorsrebuttal2023-08-17

We are grateful to the reviewer for detailed comments and for updating the score. Because we are not allowed to update the main paper and the supplementary material in this round, we briefly describe the changes we intend to make if the paper is accepted. All the clarification and revisions shown in the rebuttal will be in the revised paper, including (1) revisions and clarifications of statements in (W1); (2) adding a section of related work to emphasize the contribution of this work compared with the state of the art; (3) clarification of the controller structure as an integrator + static map and the notion of Neural-PI; (4) clarification of the notion "end-to-end" that is equivalent to "by construction"; (5) supplementing a basic description of the problem, the verification of assumptions, and the performance of the linear PI control in the main paper; (6) supplementing the experiments related to the impact of trained trajectories, noise, and disturbances to the supplementary material; (7) adding the discussions on the limitation and applicability in real-world applications to the conclusion. If the reviewer has any more specific questions, we are happy to address them.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC