NeurIPS 2020

The Power of Predictions in Online Control

Review 1

Summary and Contributions: Update: Thanks for the thoughtful response. The suggested longer and more systematic review of related approaches in MPC will definitely improve this work. --- This paper studies the LQR problem with known dynamics and predictions of disturbances over a fixed lookahead, for both stochastic and adversarial disturbances. For both cases, the authors derive the form of the optimal policy using the predictions, and show that the regret decays exponentially in the prediction horizon, and furthermore that a simple MPC-like strategy is nearly optimal.

Strengths: This paper makes very nice connections between the online learning literature and model predictive control. The main ideas are clearly explained, and the claims are backed up by theory.

Weaknesses: The clarity around the MPC problem setting could be improved a bit, as could the explanation of its relation to related work. See the sections below.

Correctness: The results seem to be correct.

Clarity: Overall, this paper is well written. It would be good to better clarify the considered MPC setting as distinct from general MPC frameworks, which might consider planning horizons independent from the prediction horizon and which have a stronger focus on safety and feasibility.

Relation to Prior Work: There is a solid effort at placing these results in the context of related work, but it could be improved a bit. The MPC literature does not generally treat disturbances in the same manner as this work, instead placing a stronger focus on robustness (see e.g. and references within) and constraint satisfaction. The incorporation of learning to MPC can be motivated by the difficulty of characterizing safe terminal sets/costs as in [25,26], but is also often motivated by unknown dynamics, e.g. A clearer treatment of this related literature would improve the paper, especially articulating similarities and differences between learning costs, disturbances, and dynamics (e.g. reference tracking can be modeled with either costs or disturbances).

Reproducibility: Yes

Additional Feedback:

Review 2

Summary and Contributions: This work consider the task of controlling a linear quadratic system subject to perturbations (stochastic or otherwise), when at each time step the next few perturbations are made available. The performance measures considered here is regret wrt the best posthoc control sequence (best U given all perturbations; capturing the statistical challenge). The proposed algorithm has a regret that scales inverse exponentially with the length of the lookahead for perturbations. In addition, the paper shows that a natural strategy (MPC) has a performance ratio that goes to one given a lookahead of log T, against the best online algorithm. While the latter is purely an computational question, its import is due to the fact that for adversarial w's, the optimal offline strategy is not clearly (it is unknown) efficiently computable.

Strengths: + The paper studies a problem that is practically relevant. An online high-level trajectory planning algorithm often outputs some states in to the future for the low-level controller to track. + The algorithmic solution (MPC) is nice even as the optimal online strategy for the stochastic case is efficiently computable. This is so because the suggested algorithm can use a black-box LQR solver.

Weaknesses: + The reliance on exact prediction (as noted in the paper) is unfortunate, and casts doubts on the robustness of the approach.

Correctness: The claims made, while not thoroughly checked, seem plausible.

Clarity: The paper is generally well written.

Relation to Prior Work: As far as the reviewer is aware, rigorous bounds for this setting are unavailable in prior art.

Reproducibility: Yes

Additional Feedback: Thanks for the discussion -- the score is retained.

Review 3

Summary and Contributions: -- The paper studies the problem of linear quadratic regulator (LQR) control in settings where the system knows (exactly) the future disturbances for a window of time. The paper discusses two types of disturbances, namely, stochastic and adversarial disturbance. The authors provide lower bounds on the dynamic regret in both cases. They also analyze the performance of the model predictive control (MPC) algorithm and based on that, derive upper bounds on the dynamic regret. ---- UPDATE ---- I want to thank the authors for their comprehensive response to my comments. In particular, adding the points made regarding the motivation for having access to exact disturbances and adding the results with inexact disturbances will make the paper stronger.

Strengths: -- Model predictive control is a widely used algorithm in practice which relies on having access to a reliable model of the system in a finite horizon ahead. The paper’s analysis of the dependence of the regret on the length of this horizon is indeed interesting to the control and learning communities. -- The authors focus on global optimal policies which may be adaptive (as opposed to static policies) and are more challenging to analyze. In addition to dynamic regret, they also use the notion of performance ratio that is more realistic as it considers the same amount of information for an algorithm and the optimal algorithm. -- In contrast to a bulk of existing work, the authors do not restrict the optimal controller to be a linear function. -- The theoretical analysis of the paper seems sound.

Weaknesses: -- I think that the paper needs more motivation to justify accessibility to exact disturbances. The disturbance in the LQR model is capturing the unknown part of the model, and hence, exact access to that even for a limited window needs a deeper discussion. In fact, the unavailability of such exact model in many realistic applications has advanced the robust MPC framework. -- The paper lacks empirical evaluations. While this is not atypical for papers focusing on theoretical regret analysis, having empirical evaluations can reinforce the results.

Correctness: -- The methodology and the theoretical claims of the paper seem correct.

Clarity: -- Overall, the writing is satisfactory. -- The statement of theorems and corollaries are somewhat informal. These statements need to be formal, accompanied with descriptive and technical sentences, and (almost) self-contained. -- I suggest the authors to add more explanation to the steps of the proofs in the supplementary material. -- Titled paragraphs should only capture one paragraph. If there is more than one paragraph, subsections would make the structure clearer.

Relation to Prior Work: -- The authors have mentioned the relevant literature and clearly distinguished their work.

Reproducibility: Yes

Additional Feedback: -- “MPC” in the abstract has not yet been defined. -- Line 46: Please rephrase “dynamic regret minimizing policies” -- Algorithm 1: No need to have a second “Input” within the loop -- Line 145-146: “\phi” is not defined explicitly in Algorithm 1 -- Line 168-172: I suggest the authors to give more intuition -- Line 203-204: Be more specific than “later in the paper” -- Line 211: Be more specific about which part of the analysis is novel -- Line 219: Why “\tilde{Q}_f = P”? Isn’t “\tilde{Q}_f” given? -- It would help to add a table summarizing the results. If more space is needed, some of the results may be taken to the supplementary material. -- Line 270: Could you give some intuition why the same order of regret is attainable for the adversarial case? -- When describing the notation, please describe the difference between the notations used for the order (“O” vs. “\Theta” vs. “\Omega”). -- In the bibliography, the year of some references have been repeated. -- Line 470: Avoid saying “turns out to be correct” -- Proof of Theorem 4.4 does not discuss the result for Example 4.3. -- Line 594: Where is “\Omega” used in the proof?

Review 4

Summary and Contributions: This paper investigates the effect of predictions in online Linear Quadratic Regulator control with stochastic and adversarial disturbances. Upper bounds for the optimal cost and minimum dynamic regret is computed given access to predictions. The model uses general process disturbances with only stabilizability assumption.

Strengths: The paper enjoys the merit of rigorous derivations and theoretical analysis of the proposed results. The problem statement and description of the model is clearly written and the theoretical results are rigorously established. Performance bounds are and relative benefit of additional predictions are calculated.

Weaknesses: I failed to understand the motivation and originality of the work. In particular, the strict improvement in performance for access to predictions of future disturbances is not quite clear to me, specially in the context of exponential decay of benefit given added predictions. The literature for stochastic model predictive control (SMPC) is quite vast and the claim of originality for general additive disturbance and adversarial setting needs to be supported by more concrete citations and possible comparisons with similar works, e.g., general additive disturbance: Daniel E. Quevedo, Debasish Chatterjee, Stochastic predictive control, International Journal of Robust and Nonlinear Control, 10.1002/rnc.4722, 29, 15, (4985-4986), (2019), Joel A. Paulson, Edward A. Buehler, Richard D. Braatz & Ali Mesbah (2020) Stochastic model predictive control with joint chance constraints, International Journal of Control, 93:1, 126-139, DOI: 10.1080/00207179.2017.1323351; Adversarial setting: Optimal Attack against Autoregressive Models by Manipulating the Environment, Yiding Chen and Xiaojin Zhu, arXiv: 1902.00202, 2019.

Correctness: I have not gone through every step of the proofs but overall the methods seem correct.

Clarity: The authors show effort in clearly explaining the model and the results. However, the merit and practicality of a system having access to predicted disturbances is not clear to me.

Relation to Prior Work: The authors made an effort to cite relevant background work. However, I believe the merit and originality of the paper can be enhanced by comparing works that are closer in spirit. I have mentioned a few in my previous comment and I am sure there are more such prior work that need to be cited.

Reproducibility: Yes

Additional Feedback: