NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:1084
Title:Beyond Online Balanced Descent: An Optimal Algorithm for Smoothed Online Optimization

Reviewer 1


		
I have a few questions and would appreciate if the authors can answer them in the rebuttal and / or the final version of the paper : 1. The paper deals with the movement cost ||x_t - x_{t-1} ||^2 as compared to ||x_t - x_{t-1} || (which one might expect in standard applications). What happens for the latter case ? Do the upper bounds / lower bounds and the new algorithm extend to that setting ? 2. Are there any lower bounds or impossibility results for regret when L is large (in the definition of L-constrained dynamic regret, Theorem 6) ? 3. Why is the \gamma parameter not tunable in the lower bound of Theorem 2 for OBD. 4. Suppose f_t’s are G lipschitz. What can be said then ? It seems that the current lower bound will not work if the hitting losses are forced to have bounded gradients. Is online balanced descent optimal then ? 5. Is the regret still bounded when f_t is stochastic and not known before making the decision x_t. It seems that the competitive ratio is hard to bound, but regret may still be controllable. Writing Style: The paper is overall well written. There are some minor typos (e.g. lemma-1 in appendix; line 281, etc), but I am sure they will be corrected in the final version of the paper. ============ Post Author Feedback ============ I want to thank the authors for giving me a detailed response to my review and, sharing connections of their work to the existing literature. I am increasing my score by 1 point (from 7 to 8) to acknowledge my satisfaction. Best !

Reviewer 2


		
--------- post rebuttal comments: After carefully reading the comments from the other two reviewers and the responses provided by the authors on all the comments, I am willing to upgrade my score from 6 to 7. My concern about the paper being "highly specialized" remains, but it will be of interest to some within NeurIPS. --------- a) summary of the content: In this paper, the authors consider an online convex optimization variant where an online learner plays a series of rounds against an adaptive adversary. In each round $t$, the adversary picks a convex cost function $f_t$ and reveals it to the learner. The learner chooses a point $x_t$ and incurs a hitting cost $f_t(x_t)$ and a movement cost $c(x_{t-1},x_t)$ associated with the change of point over the previous round. The objective of the learner is to minimize the sum over $T$ rounds of per-round hitting costs and movement costs, and design online algorithms that perform well against the adversary, as measured by competitive ratios. Concentrating on the case of the $\ell_2$-squared norm $c(x_{t-1},x_t)=\frac{1}{2} \|x_t-x_{t-1} \|_2^2$ and $m$-strongly convex functions $f_t$, the authors show a lower bound of $\Omega(m^{-1/2})$ on any online algorithms, show that an existing algorithm called Online Balanced Decent (OBD) has a lower bound of $\Omega(m^{-2/3})$, and propose two optimal variants, Greedy OBD (G-OBD) and Regularized OBD (R-OBD) with an $O(m^{-1/2})$ competitive ratio. They also show that assuming that under additional smoothness for $f_t$, R-OBD can simultaneously achieve a constant, dimension-free competitive ratio and sublinear ($L$-constrained dynamic) regret. b) strengths and weaknesses of the submission. * originality: This is a highly specialized contribution building up novel results on two main fronts: The derivation of the lower bound on the competitive ratio of any online algorithm and the introduction of two variants of an existing algorithm so as to meet this lower bound. Most of the proofs and techniques are natural and not surprising. In my view the main contribution is the introduction of the regularized version which brings a different, and arguably more modern interpretation, about the conditions under which these online algorithms perform well in these adversarial settings. * quality: The technical content of the paper is sound and rigorous * clarity: The paper is in general very well-written, and should be easy to follow for expert readers. * significance: As mentioned above this is a very specialized paper likely to interest some experts in the online convex optimization communities. Although narrow in scope, it contains interesting theoretical results advancing the state of the art in dealing with these specific problems. * minor details/comments: - p.1, line 6-7: I would rewrite the sentence to simply express that the lower bound is $\Omega(m^{-1/2})$ \- p.3, line 141: cost an algorithm => cost of an algorithm \- p.4, Algorithm 1, step 3: mention somewhere that this is the projection operator (not every reader will be familiar with this notation \- p.5, Theorem 2: remind the reader that the $\gamma$ in the statement is the parameter of OBD as defined in Algorithm 1 \- p.8, line 314: why surprisingly?

Reviewer 3


		
The paper derives novel bounds for Smoothed Online Convex Optimization (SOCO) and a particular state of the art algorithm on Online Balanced Descent (OBD), one in , showing its suboptimality, for the setup with (m-)strongly convex hitting costs and the squared l2-norm as movement costs. It also introduces two novel variants of OBD: G-OBD and R-OBD, and shows their optimality in terms of the strong convexity parameter. The paper takes big steps towards understanding Smoothed Online Convex Optimization (SOCO) by addressing and solving various unresolved problems in it: - It provides the first non-trivial lower bounds on SOCO with (m-)strongly convex hitting costs and the squared l2-norm as movement costs. The bound grows as \Omega(m^{-1/2}) as m goes to 0. - It introduces two variants of a state of the art algorithm, Online Balanced Descent (OBD): G-OBD and R-OBD. R-OBD matches the exact lower bound on the above setup, and thus optimal in terms of the strong convexity parameter. G-OBD, on the other hand, has slightly less competitive ratio, O(m^{-1/2}), but on a broader class of problems. - It proves a Omega(m^{-2/3}) lower bound on the competitive ratio of Online Balanced Descent (OBD), a state of the art algorithm in OBD. This thus shows its suboptimality in terms of the strong convexity parameter. - The presented bounds imply that R-OBD can achieve dimension-free competitive ratio and sublinear regret simultaneously for problems with m-strongly convex hitting costs and the squared l2-norm as movement costs. (For linear hitting and movement costs, this was known to be impossible.) The obtained results are strong contributions that improve our understanding of the topic and potentially help solving further problems in the area. The subtleties of the results and the problems are essential, which automatically renders the paper hard to approach; however the authors did a really good job in highlighting the main details and guiding the reader through several layers of the problem. The summary of the main ideas were also very efficient in conveying the main challenges and ideas. Overall, it was an enjoyable read. Minor remarks: line 179: "costs costs" -> "costs" Lemma 1: please quantify \beta. Lemma 4: this is a well-known property of convex conjugates