Online Convex Optimization with Unbounded Memory

Online convex optimization (OCO) is a widely used framework in online learning. In each round, the learner chooses a decision in a convex set and an adversary chooses a convex loss function, and then the learner suffers the loss associated with their current decision. However, in many applications the learner's loss depends not only on the current decision but on the entire history of decisions until that point. The OCO framework and its existing generalizations do not capture this, and they can only be applied to many settings of interest after a long series of approximation arguments. They also leave open the question of whether the dependence on memory is tight because there are no non-trivial lower bounds. In this work we introduce a generalization of the OCO framework,"Online Convex Optimization with Unbounded Memory", that captures long-term dependence on past decisions. We introduce the notion of $p$-effective memory capacity, $H_p$, that quantifies the maximum influence of past decisions on present losses. We prove an $O(\sqrt{H_p T})$ upper bound on the policy regret and a matching (worst-case) lower bound. As a special case, we prove the first non-trivial lower bound for OCO with finite memory \citep{anavaHM2015online}, which could be of independent interest, and also improve existing upper bounds. We demonstrate the broad applicability of our framework by using it to derive regret bounds, and to improve and simplify existing regret bound derivations, for a variety of online learning problems including online linear control and an online variant of performative prediction.

Paper

References (67)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 58kP3/10 · confidence 4/52023-06-12

Summary

This paper focuses on online convex optimization with memory, a topic with increasing attention recently. Traditional framework assumes that the current environment is only affected by the decisions of a limited past, while this work considers that the current environment is affected by \emph{all} previous decisions. Specifically, the authors generalize the existing framework by studying the sequences of decisions in a typical sequence space. The authors then define the notion of memory capacity with the help of a bounded weighted norm in such sequence space and achieve new policy regret bounds. The authors verify its applicability by reducing several problems into the proposed framework, including variants of online linear control and performative prediction.

Strengths

(1) The motivation is meaningful. It is definitely important in online learning literature to capture the historical impact in the sequence. (2) The solution is simple, and the proof is clear and correct.

Weaknesses

(1) My first concern comes from the novelty of the proposed framework. Although it enables an infinite memory length, the impact of history is modelled as typical linear operators, and positive results are obtained only when the operator behaves like a geometric combination of past decisions, which has been studied in some areas like linear control and reinforcement learning. Could the authors provide more special cases (in sec. 2.3), or elaborate more (in the paragraph just above def 2.3), to show that the proposed framework does give some new intuition to this problem? (2) The technical contribution seems insufficient. The results mostly follows that of the traditional proofs in (Anava et al., 2015), with the operations on a finite sequence replaced by linear operators in functional space of sequences. Could the authors highlight the technical contributions in the proofs of their theory? For example, the explanation of (theorem 3.4) is quite clear and intuitive, is there something worth highlighted in the main results of the upper bounds?

Questions

Please see the comments above.

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed 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

2 fair

Limitations

Please see the comments above.

Reviewer xqDp4/10 · confidence 3/52023-06-30

Summary

This paper introduces a problem called online convex optimization (OCO) with unbound memory, which generalizes an existing problem called OCO with memory. To address this problem, the authors propose an algorithm called FTRL and analyze its regret. Moreover, the authors demonstrate that this new problem and their algorithm have several applications.

Strengths

1) The problem of online convex optimization (OCO) with unbound memory is more general than the existing problem called OCO with memory. 2) The proposed algorithm enjoys a regret bound and has some applications. 3) It seems that the results in this paper can simplify the regret anlysis for the problem of online linear control.

Weaknesses

1) The extension from OCO with memory to OCO with unbound memory seems to be straightforward. Moreover, the algorithm for OCO with unbound memory and the corresponding analysis are very similar to existing algorithms for OCO with memory and their analysis. So, to some extent, the novelty of this paper is limited. 2) Although OCO with unbound memory seems to be more challenging, the authors only consider the history determined by fixed linear operators, i.e., $A$ and $B$ in the paper. Moreover, the authors do not explain why they only consider this case. 3) In the experiments, the authors simply set the step size of the proposed algorithm and the existing algorithms as $1/\\sqrt{T}$, instead of tuning it according to their theoretical guarantees. In this way, it is not clear whether the improvement in experiments is consistent with their theoretical guarantees.

Questions

1) Please explain the reasons for only considering the history determined by fixed linear operators. 2) It would be better if the authors could conduct experiments by setting the parameters of algorithms according to their theoretical guarantees.

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

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

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

There does not exist a potential negative societal impact.

Reviewer i8qJ6/10 · confidence 3/52023-07-06

Summary

The paper considers an online learning problem between a learner and adversary. The learner chooses action $x_t$ each round, and the state $h_t$ evolves according to the dynamics $h_t = Ah_{t-1} + Bx_t$. The oblivious adversary commits to a loss function $f_t$ each round. The learner suffers cumulative loss $\sum f_t(h_t)$. This paper uses FTRL to solve this problem.

Strengths

The paper studies online learning with a linear control component. The framework captures a setting where past decisions affect future loss. The paper presented theoretical results with upper and lower bounds. The paper showed that the difficulty of this problem is captured by the so-called $p$-effective memory, which essentially quantifies the memory(less) property of the operator $A$.

Weaknesses

While the problem proposed in this paper is seemingly new, it does seem to shed much new algorithmic ideas and insights.

Questions

na

Rating

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

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors addressed limitations and directions for future work.

Reviewer cA1Y6/10 · confidence 3/52023-07-08

Summary

This paper studies a generalization of online convex optimization (OCO) with memory. The setting allows the current stage cost to depend on all past decisions via a discrete-time linear dynamical system. The authors proposed a follow-the-regularized-leader algorithm that can achieve a sublinear static regret against any fixed action. They also showed a lower bound that matches the regret upper bound in the order of horizon $T$ and Lipschitz constants. The authors discussed two applications of their results to online control and online performative prediction.

Strengths

Theory for online (convex) optimization with unbounded memory is important in the field of learning for control. While this problem is intractable in general, it is good to see results that formally define the “effective memory” and study the corresponding upper/lower bounds.

Weaknesses

My major concern about this work is about the problem setting: I believe the setting proposed here is a special case of online control with adversarial disturbances (e.g., [Agarwal et al., 2019b]). Specifically, the history $h_t$ corresponds to the state and the decision $x_t$ corresponds to the control input. The only difference might be the benchmark: While online control compares against the best DAC policy, this work compares with a fixed action. But the DAC policy class can easily contain any fixed action if we add a dummy dimension with entry 1 in the (adversarial) disturbances. I hope the authors can correct me if my understanding is wrong, and a discussion in the revision may be helpful. Since the proposed problem setting can be reduced to (or maybe equivalent to) online control with adversarial disturbances, one should evaluate the results in this work by comparing with not only [Agarwal et al., 2019b], but also more recent works on online control like [Minasyan et al., 2022], [Chen et al., 2022], and [Lin et al., 2022]. To the best of my knowledge, existing results can handle much more complicated settings that involves time-varying dynamics, unknown dynamics, or even nonlinear dynamics. They considered stronger benchmarks like adaptive regret or dynamic regret. I encourage the authors to do a more detailed literature review about online control and clarify the significance of the main results. Besides, I also have a concern about the time complexity of the proposed algorithm. It seems that the time/memory complexity of constructing $\tilde{f}_t$ from past $f_1, \cdots, f_t$ grows linearly with respect to time $t$. However, existing online control algorithms like the one in [Agarwal et al., 2019b] only requires $log(T)$ time/memory for each decision. Thus, I am not sure how practical the proposed FTRL algorithm is. [Minasyan et al., 2022]: https://arxiv.org/pdf/2202.07890.pdf [Chen et al., 2022]: https://arxiv.org/pdf/2110.07807.pdf [Lin et al., 2022]: https://arxiv.org/pdf/2210.12320v1.pdf

Questions

Please see my comments in the previous section. I gave the score based on my current understanding about the similarity with online control, and I would be happy to raise the score if the authors address my concern.

Rating

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

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

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The authors discussed about some future directions in Section 5.

Reviewer 58kP2023-08-12

Thank you for the response. I understand the points that you argue, and admit that the proposed lower bound is new. But to be very honest, I am still concerning the novelty of this paper. On the techniques to derive the framework, it is not surprising to me to replace the assumption of "bounded regularity in each direction" in finite-dimension cases with the assumption of a bounded norm of coefficient matrices in infinite-dimension cases. On the novelty of the framework, I still keep my point that the current version is not enough to show its potential to give new ideas in designing algorithms. In particular, could the authors provide more special cases that the proposed framework can recover (while the existing works cannot)?

Authorsrebuttal2023-08-12

Re: special cases - In addition to dynamics that exhibit geometric decay, our framework also captures examples that exhibit a transient behavior. Even though we use a scalar decay factor in our OCO with infinite memory example for simplicity, a more general example is when the decisions are multiplied by a matrix Z with spectral norm less than one. Dynamics describing a vector rotating with decreasing norm have operator norm and spectral radius less than 1. However, by stretching the space, the dynamics would be rotating according to an ellipse, so the norm would be alternatively growing and shrinking even as the vector eventually decays to the origin – as a result the transients are nontrivial and the operator norm is larger than 1 even as the spectral radius is less than 1. This is another simple example that our proposed framework can recover.

Reviewer cA1Y2023-08-12

Thank you for the detailed response!

I want to thank the authors for the detailed response. I apologize for ignoring the difference in the infinite/finite dimension issue in my initial review. I believe allowing the state (or history) space to have infinite dimensions is a good technical contribution, and I would appreciate it if the authors can elaborate more about the significance of this contribution. 1. I’m not fully convinced about the necessity of allowing the state (or history) space to have infinite dimensions. From the current two examples, it seems like the infinite-dimension state (or history) space is a result of using a specific proof technique, while other techniques may solve the same problems in finite dimensions. 2. If I understand correctly, one can still apply the algorithm in [Agarwal et al., 2019b] to the setting of this paper, but the regret will be unbounded when the state (or history) space has infinite dimensions. Is the dependence on the dimension a fundamental limit of the algorithm in [Agarwal et al., 2019b] or due to the proof technique? 3. The computational cost of the algorithm depends on how fast one can implement an oracle call to the operator A. I’m not sure if the oracle $O_A$ can always run in a constant time especially when the dimension of the state (or history) space is infinite.

Authorsrebuttal2023-08-14

1. * Infinite dimensional objects arise from finite dimensional control problems when we use a **policy** to react to disturbances. * The linear control problem is not convex unless we lift from system state space into the (infinite dimensional) DAC policy space. This is *not* a result of a proof technique and has algorithmic implications. * The linear dynamics that define the control problem have a finite ($d$) dimensional representation. However, the presence of external disturbances means that we need to model the *system response* rather than merely the *system state*. The system response is infinite dimensional. For example, see [System Level Synthesis, Anderson et al., 2019](https://arxiv.org/abs/1904.01634). * To compute how policies (decisions) affect costs (losses) through the function $\tilde{f}_t$, we need to model the system response. So, this is an algorithmic consideration and not just the result of a proof technique. * [Agarwal et al., 2019b] consider a *truncated* system response of length $O(\log T)$ that is an approximation. This is also where they appeal to results for OCO with finite memory, which our paper should be seen as a replacement for. * More generally, other examples of infinite dimensional history/states include infinite impulse response filters in signal processing (e.g., elliptic filters). * To contrast, suppose instead of choosing policies, we were interested in a "fixed constant input" model of control. That is, the learner chooses a fixed control input in each round and the benchmark is the best fixed control input. This would indeed result in a finite dimensional history space. * The history combines "noiseless" state with action: $\bar{s}_{t+1} F \bar{s}_t + G a_t, h_t = (\bar{s}_t, a_t)$. * The linear operators are block matrices: $A = [F, 0; 0, 0]$ and $B = [G; I] / \| [G; I] \|$. (The scaling on $B$ is to ensure that it has unit norm, which we assume in our paper for convenience.) * The loss functions are $f_t(h_t) = c_t( \bar{s}\_t + \sum_{k=1}^t F^k G w_{t-k} , a_t) = c_t(s_t, a_t)$. 2. Yes, you are correct. The algorithm of [Agarwal et al., 2019b] and their proofs work for standard Euclidean norms and the corresponding matrix norms. If the state and history spaces are infinite dimensional with non-standard norms, then there is a mismatch between the problem setup and their algorithm and proofs. So, in some sense the regret is unbounded because of both the algorithm and the proof technique. The algorithm would use a regularizer that is strongly convex with respect to the Euclidean norm, resulting in suboptimal regret compared to our algorithm that uses a regularizer adapted to the appropriate norm; the proofs would compute the Euclidean norm of infinite dimensional objects. 3. We agree with your point. * See lines 981 - 988 in the appendix where we discuss this issue. We believe that it is a feature that our framework provides a clean abstraction for the user that they can use for a variety of applications. The dimensionality of $\mathcal{X}$, the choice of the operator $A$, etc. are application dependent. The user could use our framework with a lower dimensional decision space $\mathcal{X}'$ and then analyze the error that results from such an approximation. Meanwhile, our framework allows the user to define non-standard norms on the decision and history spaces. This can be a simple but powerful way of encoding prior knowledge about the application, and our framework handles the technical complications that arise from this (e.g., bounding the Lipschitz constant, operator norms, etc. - see Lemmas E.2 and E.6 (especially lines 885 - 891) for an illustration).

Reviewer cA1Y2023-08-14

Thank you for responding to my follow-up questions. I have not seen much previous work on online learning/control that emphasizes an infinite-dimensional state space, so I encourage the authors to dig deeper if this can be another spotlight of the paper. I will raise my score because the major concern I mentioned in my initial review has been addressed.

Reviewer i8qJ2023-08-18

Thanks for the response ( yes I did mean to say 'doesn't', apologies for the typo ). I think this is a technically sound paper and will keep my original score. I will also read other reviewers' comments.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC