2
$\begingroup$

In online learning/online convex optimization, it's often the case that you compare your algorithm against the best action in hindsight (i.e., from https://people.cs.umass.edu/~akshay/courses/cs690m/files/lec15.pdf)

$$ \operatorname{Regret}(T)=\sum_t{f_t(w_t)}-\min_u \sum_t{f_t(u)} $$

For sequence $t=1, \ldots, T$, $u$ in this case is the best action over the entire sequence. In practice, a popular approach is to use a rolling window approach (i.e., regress over past 10 periods of data). Does the $u$ mean that it's basically using a growing window? If $f_t$ changes over time, wouldn't $u$ do terribly since it's basically taking the average over all past observations?

$\endgroup$
0
$\begingroup$

The regret measure of $$\operatorname{Regret}(T)=\sum_{t=1}^T{f_t(w_t)}-\min_u \sum_{t=1}^T{f_t(u)}$$ is the difference between the cumulative loss of the player and that of the best fixed decision in hindsight. So as you have already observed, the measure implicitly assumes that there is a reasonably good decision over all iterations. The measure is thus not suitable in non-stationary environments and is sometimes called static regret.

There are two mainstream measures for tacking non-stationary environments:

  • adaptive regret: ensuring static regret in any small interval (this is exactly the measure mentioned by @Sebastiaan);
  • dynamic regret: ensuring a small regret with a changing comparator sequence (I will focus on this measure).

The definition of dynamic regret is as follows, $$\mbox{D-Regret}_T(u_1,\ldots,u_T) = \sum_{t=1}^{T} f_t(w_t) - \sum_{t=1}^{T} f_t(u_t).$$ So you can see the comparators $\{u_1,\ldots,u_T\}$ here are allowed to change over iterations instead of a fixed one in the static regret.

The dynamic regret bounds will typically involve a term measuring the non-stationarity of environments, for example, $P_T = \sum_{t=2}^{T} \| u_t - u_{t-1}\|$.

The notion of dynamic regret is also called tracking regret/ shifting regret in the early development of prediction with expert advice. For online convex optimization (OCO), the pioneering work [1] first introduces this measure, [2] establishes the lower bound and designs an optimal algorithm for OCO, [3] extends the idea to bandit convex optimization scenarios. See more introduction and related works in [2] and [3].

Reference

[1] M. Zinkevich. Online convex programming and generalized infinitesimal gradient ascent. In Proceedings of the 20th International Conference on Machine Learning (ICML), pages 928–936, 2003.

[2] Lijun Zhang, ShiyinLu, and Zhi-Hua Zhou. Adaptive Online Learning in Dynamic Environments. In Advances in Neural Information Processing Systems 31 (NeurIPS), 2018.

[3] Peng Zhao, Guanghui Wang, Lijun Zhang, Zhi-Hua Zhou. Bandit Convex Optimization in Non-stationary Environments. In Proceedings of the 22nd International Conference on Artificial Intelligence and Statistics (AISTATS), 2020.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Thanks a lot for your answer. It became clear that static regret wasn't the measure I'm after $\endgroup$ – swmfg Apr 8 at 22:24
0
$\begingroup$

I think what you are aiming at is that there might be some trend in your data which is by design overlooked by $u$, since most online learning theory assumes stationarity. In the dynamic data case there are online alternatives where you attempt to keep the regret small on every subinterval, i.e.

$$ \text{Regret}([t_1,t_2]) = \sum_{t\in[t_1,t_2]}f_t(w_t) - \min_u\sum_{t\in[t_1,t_2]}f_t(u) $$

should be small for all $[t_1,t_2]\subset[0,T]$. Two algorithms that supposedly achieve upper bounds on the regret are Fixed Share and Mixing Past Posteriors, but I know little more than their name.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Thanks for your reply. To answer my own question, u is an expanding window. A popular algorithm for static regret is Follow the Leader, which basically chooses the best policy on all data so far. This is equivalent to estimating with an expanding window. $\endgroup$ – swmfg May 29 '19 at 22:54
  • $\begingroup$ Actually, Follow-The-Leader is not very popular because of its poor performance. Even still, although FTL happens to coincide with an expanding window estimator, they are not the same. Online Learning does not work with estimators, and the concepts should not be simply compared. In all cases online learning works with 'expanding windows' because the framework works by looking at suffered losses, and it makes no sense to 'forget' past losses. $\endgroup$ – Sebastiaan Jun 3 '19 at 8:26

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.