1 Introduction
The stochastic multiarmed bandit (MAB) problem is a prominent framework for capturing the explorationexploitation tradeoff in online decision making and experiment design. The MAB problem proceeds in discrete sequential rounds, where in each round, the player pulls one of the possible arms. In the classic stochastic MAB setting, the player immediately observes stochastic feedback from the pulled arm in the form of a ‘reward’ which can be used to improve the decisions in subsequent rounds. One of the main application areas of MABs is in online advertising. Here, the arms correspond to adverts, and the feedback would correspond to conversions, that is users buying a product after seeing an advert. However, in practice, these conversions may not necessarily happen immediately after the advert is shown, and it may not always be possible to assign the credit of a sale to a particular showing of an advert. A similar challenge is encountered in many other applications, e.g., in personalized treatment planning, where the effect of a treatment on a patient’s health may be delayed, and it may be difficult to determine which out of several past treatments caused the change in the patient’s health; or, in content design applications, where the effects of multiple changes in the website design on website traffic and footfall may be delayed and difficult to distinguish.
In this paper, we propose a new bandit model to handle online problems with such ‘delayed, aggregated and anonymous’ feedback. In our model, a player interacts with an environment of actions (or arms) in a sequential fashion. At each time step the player selects an action which leads to a reward generated at random from the underlying reward distribution. At the same time, a nonnegative random integervalued delay is also generated i.i.d. from an underlying delay distribution. Denoting this delay by and the index of the current round by , the reward generated in round will arrive at the end of the th round. At the end of each round, the player observes only the sum of all the rewards that arrive in that round. Crucially, the player does not know which of the past plays have contributed to this aggregated reward. We call this problem multiarmed bandits with delayed, aggregated anonymous feedback (MABDAAF). As in the standard MAB problem, in MABDAAF, the goal is to maximize the cumulative reward from plays of the bandit, or equivalently to minimize the regret. The regret is the total difference between the reward of the optimal action and the actions taken.
If the delays are all zero, the MABDAAF problem reduces to the standard (stochastic) MAB problem, which has been studied considerably (e.g., Thompson, 1933; Lai & Robbins, 1985; Auer et al., 2002; Bubeck & CesaBianchi, 2012). Compared to the MAB problem, the job of the player in our problem appears to be significantly more difficult since the player has to deal with (i) that some feedback from the previous pulls may be missing due to the delays, and (ii) that the feedback takes the form of the sum of an unknown number of rewards of unknown origin.
An easier problem is when the observations are delayed, but they are nonaggregated and nonanonymous: that is, the player has to only deal with challenge (i) and not (ii). Here, the player receives delayed feedback in the shape of actionreward pairs that inform the player of both the individual reward and which action generated it. This problem, which we shall call the (nonanonymous) delayed feedback bandit problem, has been studied by Joulani et al. (2013), and later followed up by Mandel et al. (2015) (for bounded delays). Remarkably, they show that compared to the standard (nondelayed) stochastic MAB setting, the regret will increase only additively by a factor that scales with the expected delay. For delay distributions with a finite expected delay, , the worst case regret scales with . Hence, the price to pay for the delay in receiving the observations is negligible. QPMD of Joulani et al. (2013) and SBD of Mandel et al. (2015) place received rewards into queues for each arm, taking one whenever a base bandit algorithm suggests playing the arm. Throughout, we take UCB1 (Auer et al., 2002) as the base algorithm in QPMD. Joulani et al. (2013) also present a direct modification of the UCB1 algorithm. All of these algorithms achieve the stated regret. None of them require any knowledge of the delay distributions, but they all rely heavily upon the nonanonymous nature of the observations.
While these results are encouraging, the assumption that the rewards are observed individually in a nonanonymous fashion is limiting for most practical applications with delays (e.g., recall the applications discussed earlier). How big is the price to be paid for receiving only aggregated anonymous feedback? Our main result is to prove that essentially there is no extra price to be paid provided that the value of the expected delay (or a bound on it) is available. In particular, this means that detailed knowledge of which action led to a particular delayed reward can be replaced by the much weaker requirement that the expected delay, or a bound on it, is known. Fig. 1 summarizes the relationship between the nondelayed, the delayed and the new problem by showing the leading terms of the regret. In all cases, the dominant term is . Hence, asymptotically, the delayed, aggregated anonymous feedback problem is no more difficult than the standard multiarmed bandit problem.
1.1 Our Techniques and Results
We now consider what sort of algorithm will be able to achieve the aforementioned results for the MABDAAF problem. Since the player only observes delayed, aggregated anonymous rewards, the first problem we face is how to even estimate the mean reward of individual actions. Due to the delays and anonymity, it appears that to be able to estimate the mean reward of an action, the player wants to have played it consecutively for long stretches. Indeed, if the stretches are sufficiently long compared to the mean delay, the observations received during the stretch will mostly consist of rewards of the action played in that stretch. This naturally leads to considering algorithms that
switch actions rarely and this is indeed the basis of our approach.Several popular MAB algorithms are based on choosing the action with the largest upper confidence bound (UCB) in each round (e.g., Auer et al., 2002; Cappé et al., 2013). UCBstyle algorithms tend to switch arms frequently and will only play the optimal arm for long stretches if a unique optimal arm exists. Therefore, for MABDAAF, we will consider alternative algorithms where armswitching is more tightly controlled. The design of such algorithms goes back at least to the work of Agrawal et al. (1988) where the problem of bandits with switching costs was studied. The general idea of these rarely switching algorithms is to gradually eliminate suboptimal arms by playing arms in phases and comparing each arm’s upper confidence bound to the lower confidence bound of a leading arm at the end of each phase. Generally, this sort of rarely switching algorithm switches arms only times. We base our approach on one such algorithm, the socalled Improved UCB^{1}^{1}1The adjective “Improved” indicates that the algorithm improves upon the regret bounds achieved by UCB1. The improvement replaces by in the regret bound. algorithm of Auer & Ortner (2010).
Using a rarely switching algorithm alone will not be sufficient for MABDAAF. The remaining problem, and where the bulk of our contribution lies, is to construct appropriate confidence bounds and adjust the length of the periods of playing each arm to account for the delayed, aggregated anonymous feedback. In particular, in the confidence bounds attention must be paid to fine details: it turns out that unless the variance of the observations is dealt with, there is a blowup by a multiplicative factor of . We avoid this by an improved analysis involving Freedman’s inequality (Freedman, 1975). Further, to handle the dependencies between the number of plays of each arm and the past rewards, we combine Doob’s optimal skipping theorem (Doob, 1953) and AzumaHoeffding inequalities. Using a rarely switching algorithm for MABDAAF means we must also consider the dependencies between the elimination of arms in one phase and the corruption of observations in the next phase (ie. past plays can influence both whether an arm is still active and the corruption of its next plays). We deal with this through careful algorithmic design.
Using the above, we provide an algorithm that achieves worst case regret of using only knowledge of the expected delay, . We then show that this regret can be improved by using a more careful martingale argument that exploits the fact that our algorithm is designed to remove most of the dependence between the corruption of future observations and elimination of arms. Particularly, if the delays are bounded with known bound , we can recover worst case regret of , matching that of Joulani et al. (2013). If the delays are unbounded but have known variance , we show that the problem independent regret can be reduced to .
1.2 Related Work
We have already discussed several of the most relevant works to our own. However, there has also been other work looking at different flavors of the bandit problem with delayed (nonanonymous) feedback. For example, Neu et al. (2010) and CesaBianchi et al. (2016) consider nonstochastic bandits with fixed constant delays; Dudik et al. (2011) look at stochastic contextual bandits with a constant delay and Desautels et al. (2014) consider Gaussian Process bandits with a bounded stochastic delay. The general observation that delay causes an additive regret penalty in stochastic bandits and a multiplicative one in adversarial bandits is made in Joulani et al. (2013). The empirical performance of armed stochastic bandit algorithms in delayed settings was investigated in Chapelle & Li (2011). A further related problem is the ‘batched bandit’ problem studied by Perchet et al. (2016). Here the player must fix a set of time points at which to collect feedback on all plays leading up to that point. Vernade et al. (2017) consider delayed Bernoulli bandits where some observations could also be censored (e.g., no conversion is ever actually observed if the delay exceeds some threshold) but require complete knowledge of the delay distribution. Crucially, here and in all the aforementioned works, the feedback is always assumed to take the form of armreward pairs and knowledge of the assignment of rewards to arms underpins the suggested algorithms, rendering them unsuitable for MABDAAF. To the best of our knowledge, ours is the first work to develop algorithms to deal with delayed, aggregated anonymous feedback in the bandit setting.
1.3 Organization
The reminder of this paper is organized as follows: In the next section (Section 2) we give the formal problem definition. We present our algorithm in Section 3. In Section 4, we discuss the performance of our algorithm under various delay assumptions; known expectation, bounded support with known bound and expectation, and known variance and expectation. This is followed by a numerical illustration of our results in Section 5. We conclude in Section 6.
2 Problem Definition
There are actions or arms in the set . Each action is associated with a reward distribution and a delay distribution . The reward distribution is supported in and the delay distribution is supported on . We denote by the mean of , and define to be the reward gap, that is the expected loss of reward each time action is chosen instead of an optimal action. Let
be an infinite array of random variables defined on the probability space
which are mutually independent. Further, follows the distribution and follows the distribution . The meaning of these random variables is that if the player plays action at time , a payoff of will be added to the aggregated feedback that the player receives at the end of the th play. Formally, if denotes the action chosen by the player at time , then the observation received at the end of the th play isFor the remainder, we will consider i.i.d. delays across arms. We also assume discrete delay distributions, although most results hold for continuous delays by redefining the event as in . In our analysis, we will sum over stochastic index sets. For a stochastic index set and random variables we denote such sums as .
Regret definition
In most bandit problems, the regret is the cumulative loss due to not playing an optimal action. In the case of delayed feedback, there are several possible ways to define the regret. One option is to consider only the loss of the rewards received before horizon (as in Vernade et al. (2017)). However, we will not use this definition. Instead, as in Joulani et al. (2013), we consider the loss of all generated rewards and define the (pseudo)regret by
This includes the rewards received after the horizon and does not penalize large delays as long as an optimal action is taken. This definition is natural since, in practice, the player should eventually receive all outstanding reward.
Lai & Robbins (1985) showed that the regret of any algorithm for the standard MAB problem must satisfy,
(1) 
where is the KLdivergence between the reward distributions of arm and an optimal arm. Theorem 4 of Vernade et al. (2017) shows that the lower bound in (1) also holds for delayed feedback bandits with no censoring and their alternative definition of regret. We therefore suspect (1) should hold for MABDAAF. However, due to the specific problem structure, finding a lower bound for MABDAAF is nontrivial and remains an open problem.
Assumptions on delay distribution
For our algorithm for MABDAAF, we need some assumptions on the delay distribution. We assume that the expected delay, , is bounded and known. This quantity is used in the algorithm.
Assumption 1
The expected delay is bounded and known to the algorithm.
We then show that under some further mild assumptions on the delay, we can obtain better algorithms with even more efficient regret guarantees. We consider two settings: delay distributions with bounded support, and bounded variance.
Assumption 2 (Bounded support)
There exists some constant known to the algorithm such that the support of the delay distribution is bounded by .
Assumption 3 (Bounded variance)
The variance, , of the delay is bounded and known to the algorithm.
In fact the known expected value and known variance assumption can be replaced by a ‘known upper bound’ on the expected value and variance respectively. However, for simplicity, in the remaining, we use and directly. The next sections provide algorithms and regret analysis for different combinations of the above assumptions.
3 Our Algorithm
Our algorithm is a phasebased elimination algorithm based on the Improved UCB algorithm by Auer & Ortner (2010). The general structure is as follows. In each phase, each arm is played multiple times consecutively. At the end of the phase, the observations received are used to update mean estimates, and any arm with an estimated mean below the best estimated mean by a gap larger than a ‘separation gap tolerance’ is eliminated. This separation tolerance is decreased exponentially over phases, so that it is very small in later phases, eliminating all but the best arm(s) with high probability. An alternative formulation of the algorithm is that at the end of a phase, any arm with an upper confidence bound lower than the best lower confidence bound is eliminated. These confidence bounds are computed so that with high probability they are more (less) than the true mean, but within the separation gap tolerance. The phase lengths are then carefully chosen to ensure that the confidence bounds hold. Here we assume that the horizon is known, but we expect that this can be relaxed as in Auer & Ortner (2010).
Algorithm overview
Our algorithm, ODAAF, is given in Algorithm 1. It operates in phases . Define to be the set of active arms in phase . The algorithm takes parameter which defines the number of samples of each active arm required by the end of phase .
In Step 1 of phase of the algorithm, each active arm is played repeatedly for steps. We record all timesteps where arm was played in the first phases (excluding bridge periods) in the set . The active arms are played in any arbitrary but fixed order. In Step 2, the observations from timesteps in are averaged to obtain a new estimate of . Arm is eliminated if is further than from .
A further nuance in the algorithm structure is the ‘bridge period’ (see Figure 2). The algorithm picks an active arm to play in this bridge period for
steps. The observations received during the bridge period are discarded, and not used for computing confidence intervals. The significance of the bridge period is that it breaks the dependence between confidence intervals calculated in phase
and the delayed payoffs seeping into phase . Without the bridge period this dependence would impair the validity of our confidence intervals. However, we suspect that, in practice, it may be possible to remove it.Choice of
A key element of our algorithm design is the careful choice of . Since determines the number of times each active (possibly suboptimal) arm is played, it clearly has an impact on the regret. Furthermore,
needs to be chosen so that the confidence bounds on the estimation error hold with given probability. The main challenge is developing these confidence bounds from delayed, aggregated anonymous feedback. Handling this form of feedback involves a credit assignment problem of deciding which samples can be used for a given arm’s mean estimation, since each sample is an aggregate of rewards from multiple previously played arms. This credit assignment problem would be hopeless in a passive learning setting without further information on how the samples were generated. Our algorithm utilizes the power of active learning to design the phases in such a way that the feedback can be effectively ‘decensored’ without losing too many samples.
A naive approach to defining the confidence bounds for delays bounded by a constant would be to observe that,
since all rewards are in . Then we could use Hoeffding’s inequality to bound (see Appendix F) and select
for some constants . This corresponds to worst case regret of . For and large , this is significantly worse than that of Joulani et al. (2013). In Section 4, we show that, surprisingly, it is possible to recover the same rate of regret as Joulani et al. (2013), but this requires a significantly more nuanced argument to get tighter confidence bounds and smaller . In the next section, we describe this improved choice of for every phase and its implications on the regret, for each of the three cases mentioned previously: (i) Known and bounded expected delay (Assumption 1), (ii) Bounded delay with known bound and expected value (Assumptions 1 and 2), (iii) Delay with known and bounded variance and expectation (Assumptions 1 and 3).
4 Regret Analysis
In this section, we specify the choice of parameters and provide regret guarantees for Algorithm 1 for each of the three previously mentioned cases.
4.1 Known and Bounded Expected Delay
First, we consider the setting with the weakest assumption on delay distribution: we only assume that the expected delay, , is bounded and known. No assumption on the support or variance of the delay distribution is made. The regret analysis for this setting will not use the bridge period, so Step 4 of the algorithm could be omitted in this case.
Choice of
Estimation of error bounds
We bound the error between and by . In order to do this we first bound the corruption of the observations received during timesteps due to delays.
Fix a phase and arm . Then the observations in the period are composed of two types of rewards: a subset of rewards from plays of arm in this period, and delayed rewards from some of the plays before this period. The expected value of observations from this period would be but for the rewards entering and leaving this period due to delay. Since the reward is bounded by , a simple observation is that expected discrepancy between the sum of observations in this period and the quantity is bounded by the expected delay ,
(3) 
Summing this over phases gives a bound
(4) 
Note that given the choice of in (2), the above is smaller than , when large enough constants are used. Using this, along with concentration inequalities and the choice of from (2), we can obtain the following high probability bound. A detailed proof is provided in Appendix B.1.
Regret bounds
Using Lemma 1, we derive the following regret bounds in the current setting.
Proof: Given Lemma 1, the proof of Theorem 2 closely follows the analysis of the Improved UCB algorithm of Auer & Ortner (2010). Lemma 1 and the elimination condition in Algorithm 1 ensure that, with high probability, any suboptimal arm will be eliminated by phase , thus incurring regret at most We then substitute in from (2), and sum over all suboptimal arms. A detailed proof is in Appendix B.2. As in Auer & Ortner (2010), we avoid a union bound over all arms (which would result in an extra ) by (i) reasoning about the regret of each arm individually, and (ii) bounding the regret resulting from erroneously eliminating the optimal arm by carefully controlling the probability it is eliminated in each phase.
Considering the worstcase values of (roughly ), we obtain the following problem independent bound.
4.2 Delay with Bounded Support
If the delay is bounded by some constant and a single arm is played repeatedly for long enough, we can restrict the number of arms corrupting the observation at a given time . In fact, if each arm is played consecutively for more than rounds, then at any time , the observation will be composed of the rewards from at most two arms: the current arm , and previous arm . Further, from the elimination condition, with high probability, arm will have been eliminated if it is clearly suboptimal. We can then recursively use the confidence bounds for arms and from the previous phase to bound . Below, we formalize this intuition to obtain a tighter bound on for every arm and phase , when each active arm is played a specified number of times per phase.
Choice of
Here, we define,
(6)  
for some large enough constants (see Appendix C, Equation (18) for the exact values). This choice of means that for large , we essentially revert back to the choice of from (2) for the unbounded case, and we gain nothing by using the bound on the delay. However, if is not large, the choice of in (6) is smaller than (2) since the second term now scales with rather than .
Estimation of error bounds
In this setting, by the elimination condition and bounded delays, the expectation of each reward entering will be within of , with high probability. Then, using knowledge of the upper bound of the support of , we can obtain a tighter bound and get an error bound similar to Lemma 1 with the smaller value of in (6). We prove the following proposition. Since , this is considerably tighter than (3).
Proposition 4
Assume for phases . Define as the event that all arms satisfy error bounds . Then, for every arm ,
Proof: (Sketch). Consider a fixed arm . The expected value of the sum of observations for would be were it not for some rewards entering and leaving this period due to the delays. Because of the i.i.d. assumption on the delay, in expectation, the number of rewards leaving the period is roughly the same as the number of rewards entering this period, i.e., . (Conditioning on does not effect this due to the bridge period). Since , the reward coming into the period can only be from the previous arm . All rewards leaving the period are from arm . Therefore the expected difference between rewards entering and leaving the period is . Then, if is close to , the total reward leaving the period is compensated by total reward entering. Due to the bridge period, even when is the first arm played in phase , , so it was not eliminated in phase . By the elimination condition in Algorithm 1, if the error bounds are satisfied for all arms in , then . This gives the result.
Repeatedly using Proposition 4 we get,
since . Then, observe that is small. This bound is an improvement of a factor of compared to (4). For the regret analysis, we derive a high probability version of the above result. Using this, and the choice of from (6), for large enough constants, we derive the following lemma. A detailed proof is given in Appendix C.1.
Regret bounds
We now give regret bounds for this case.
Theorem 6
Proof: (Sketch). Given Lemma 5, the proof is similar to that of Theorem 2. The full proof is in Section C.2.
Then, if , we get the following problem independent regret bound which matches that of Joulani et al. (2013).
4.3 Delay with Bounded Variance
If the delay is unbounded but well behaved in the sense that we know (a bound on) the variance, then we can obtain similar regret bounds to the bounded delay case. Intuitively, delays from the previous phase will only corrupt observations in the current phase if their delays exceed the length of the bridge period. We control this by using the bound on the variance to bound the tails of the delay distributions.
Choice of
Regret bounds
We get the following instance specific and problem independent regret bound in this case.
Theorem 8
Remark
If , then the delay penalty can be reduced to (see Appendix D).
Thus, it is sufficient to know a bound on variance to obtain regret bounds similar to those in bounded delay case. Note that this approach is not possible just using knowledge of the expected delay since we cannot guarantee that the reward entering phase is from an arm active in phase .
5 Experimental Results
We compared the performance of our algorithm (under different assumptions) to QPMD (Joulani et al., 2013) in various experimental settings. In these experiments, our aim was to investigate the effect of the delay on the performance of the algorithms. In order to focus on this, we used a simple setup of two arms with Bernoulli rewards and . In every experiment, we ran each algorithm to horizon and used UCB1 (Auer et al., 2002) as the base algorithm in QPMD. The regret was averaged over replications. For ease of reading, we define ODAAF to be our algorithm using only knowledge of the expected delay, with defined as in 2 and run without a bridge period, and ODAAFB and ODAAFV to be the versions of Algorithm 1 that use a bridge period and information on the bounded support and the finite variance of the delay to define as in 6 and 7 respectively.
We tested the algorithms with different delay distributions. In the first case, we considered bounded delay distributions whereas in the second case, the delays were unbounded. In Fig. 2(a), we plotted the ratios of the regret of ODAAF and ODAAFB (with knowledge of , the delay bound) to the regret of QPMD. We see that in all cases the ratios converge to a constant. This shows that the regret of our algorithm is essentially of the same order as that of QPMD. Our algorithm predetermines the number of times to play each active arm per phase (the randomness appears in whether an arm is active), so the jumps in the regret are it changing arm. This occurs at the same points in all replications.
Fig. 2(b) shows a similar story for unbounded delays with mean (where
denotes the the half normal distribution). The ratios of the regret of ODAAF and ODAAFV (with knowledge of the delay variance) to the regret of QPMD again converge to constants. Note that in this case, these constants, and the location of the jumps, vary with the delay distribution and
. When the variance of the delay is small, it can be seen that using the variance information leads to improved performance. However, for exponential delays where , the large variance causes to be large and so the suboptimal arm is played more, increasing the regret. In this case ODAAFV had only just eliminated the suboptimal arm at time .6 Conclusion
We have studied an extension of the multiarmed bandit problem to bandits with delayed, aggregated anonymous feedback. Here, a sum of observations is received after some stochastic delay and we do not learn which arms contributed to each observation. In this more difficult setting, we have proven that, surprisingly, it is possible to develop an algorithm that performs comparably to those for the simpler delayed feedback bandits problem, where the assignment of rewards to plays is known. Particularly, using only knowledge of the expected delay, our algorithm matches the worst case regret of Joulani et al. (2013) up to a logarithmic factor. This logarithmic factors can be removed using an improved analysis and slightly more information about the delay; if the delay is bounded, we achieve the same worst case regret as Joulani et al. (2013), and for unbounded delays with known finite variance, we have an extra additive term. We supported these claims experimentally. Note that while our algorithm matches the order of regret of QPMD, the constants are worse. Hence, it is an open problem to find algorithms with better constants.
Acknowledgments
CPB would like to thank the EPSRC funded EP/L015692/1 STORi centre for doctoral training and Sparx. We would like to thank the reviewers for their helpful comments.
References
 Agrawal et al. (1988) Agrawal, R., Hedge, M., and Teneketzis, D. Asymptotically efficient adaptive allocation rules for the multiarmed bandit problem with switching cost. IEEE Transactions on Automatic Control, 33(10):899–906, 1988.
 Auer & Ortner (2010) Auer, P. and Ortner, R. UCB revisited: Improved regret bounds for the stochastic multiarmed bandit problem. Periodica Mathematica Hungarica, 61(12):55–65, 2010.
 Auer et al. (2002) Auer, P., CesaBianchi, N., and Fischer, P. Finitetime analysis of the multiarmed bandit problem. Journal of Machine Learning Research, 47(23):235–256, 2002.
 Bubeck & CesaBianchi (2012) Bubeck, S. and CesaBianchi, N. Regret Analysis of Stochastic and Nonstochastic Multiarmed Bandit Problems. Foundations and Trends in Machine Learning. 2012.
 Cappé et al. (2013) Cappé, O., Garivier, A., Maillard, O.A., Munos, R., and Stoltz, G. Kullback–Leibler upper confidence bounds for optimal sequential allocation. The Annals of Statistics, 41(3):1516–1541, 2013.
 CesaBianchi et al. (2016) CesaBianchi, N., Gentile, C., Mansour, Y., and Minora, A. Delay and cooperation in nonstochastic bandits. In Conference on Learning Theory, pp. 605–622, 2016.

Chapelle & Li (2011)
Chapelle, O. and Li, L.
An empirical evaluation of Thompson sampling.
In Advances in Neural Information Processing Systems, pp. 2249–2257, 2011.  Desautels et al. (2014) Desautels, T., Krause, A., and Burdick, J. Parallelizing explorationexploitation tradeoffs in Gaussian process bandit optimization. Journal of Machine Learning Research, 15(1):3873–3923, 2014.
 Doob (1953) Doob, J. L. Stochastic processes. John Wiley & Sons, 1953.

Dudik et al. (2011)
Dudik, M., Hsu, D., Kale, S., Karampatziakis, N., Langford, J., Reyzin, L., and
Zhang, T.
Efficient optimal learning for contextual bandits.
In
Conference on Uncertainty in Artificial Intelligence
, pp. 169–178, 2011.  Freedman (1975) Freedman, D. A. On tail probabilities for martingales. The Annals of Probability, 3(1):100–118, 1975.
 Joulani et al. (2013) Joulani, P., György, A., and Szepesvári, C. Online learning under delayed feedback. In International Conference on Machine Learning, pp. 1453–1461, 2013.
 Lai & Robbins (1985) Lai, T. and Robbins, H. Asymptotically efficient adaptive allocation rules. Advances in Applied Mathematics, 6(1):4–22, 1985.

Mandel et al. (2015)
Mandel, T., Liu, Y.E., Brunskill, E., and Popovic, Z.
The queue method: Handling delay, heuristics, prior data, and evaluation in bandits.
In AAAI, pp. 2849–2856, 2015. 
Neu et al. (2010)
Neu, G., Antos, A., György, A., and Szepesvári, C.
Online Markov decision processes under bandit feedback.
In Advances in Neural Information Processing Systems, pp. 1804–1812, 2010.  Perchet et al. (2016) Perchet, V., Rigollet, P., Chassang, S., and Snowberg, E. Batched bandit problems. The Annals of Statistics, 44(2):660–681, 2016.

Szita & Szepesvári (2011)
Szita, I. and Szepesvári, C.
Agnostic KWIK learning and efficient approximate reinforcement learning.
In Conference on Learning Theory, pp. 739–772, July 2011.  Thompson (1933) Thompson, W. On the likelihood that one unknown probability exceeds another in view of the evidence of two samples. Biometrika, 25(3–4):285–294, 1933.
 Vernade et al. (2017) Vernade, C., Cappé, O., and Perchet, V. Stochastic bandit models for delayed conversions. In Conference on Uncertainty in Artificial Intelligence, 2017.
Appendix
Appendix A Preliminaries
a.1 Table of Notation
For ease of reading, we define here key notation that will be used in this Appendix.
:  The horizon.  
:  The gap between the mean of the optimal arm and the mean of arm , .  
:  The approximation to at round of the ODAAF algorithm, .  
:  The number of samples of an active arm ODAAF needs by the end of round .  
:  The number of times each arm is played in phase , .  
:  The bound on the delay in the case of bounded delay.  
:  The first round of the ODAAF algorithm where .  
:  The random variable representing the round arm is eliminated in.  
:  The set of all time point where arm is played up to (and including) round .  
:  The reward received at time (from any possible past plays of the algorithm).  
:  The reward generated by playing arm at time .  
:  The delay associated with playing arm at time .  
:  The expected delay (assuming i.i.d. delays).  
:  The variance of the delay (assuming i.i.d. delays).  
:  The estimated reward of arm in phase . See Algorithm 1 for the definition.  
:  The start point of the th phase. See Section A.2 for more details.  
:  The end point of the th phase. See Section A.2 for more details.  
:  The start point of phase of playing arm . See Section A.2 for more details.  
:  The end point of phase of playing arm . See Section A.2 for more details.  
:  The set of active arms in round of the ODAAF algorithm.  
:  The contribution of the reward generated at time in certain intervals relating to phase to the corruption. See (11) for the exact definitions.  
:  The smallest algebra containing all information up to time , see (8) for a definition. 
a.2 Beginning and End of Phases
We formalize here some notation that will be used throughout the analysis to denote the start and end points of each phase. Define the random variables and for each phase to be the start and end points of the phase. Then let , denote the start and end points of playing arm in phase . See Figure 4 for details. By convention, let if arm is not active in phase , if the algorithm never reaches phase and let for all . It is important to point out that are deterministic so at the end of any phase , once we have eliminated suboptimal arms, we also know which arms are in and consequently the start and end points of phase . Furthermore, since we play arms in a given order, we also know the specific rounds when we start and finish playing each active arm in phase . Hence, at any time step in phase , and for all active arms will be known. More formally, define the filtration where
(8) 
and . This means the joint events like for all , .
a.3 Useful Results
For our analysis, we will need Freedman’s version of Bernstein’s inequality for the righttail of martingales with bounded increments:
Theorem 10 (Freedman’s version of Bernstein’s inequality; Theorem 1.6 of Freedman (1975))
Let be a realvalued martingale with respect to the filtration with increments : and , for . Assume that the difference sequence is uniformly bounded on the right: almost surely for . Define the predictable variation process for . Then, for all , ,
This result implies that if for some deterministic constant, , holds almost surely, then holds for any .
We will also make use of the following technical lemma which combines the HoeffdingAzuma inequality and Doob’s optional skipping theorem (Theorem 2.3 in Chapter VII of Doob (1953))):
Lemma 11
Fix the positive integers and let . Let be a filtration, be a sequence of valued random variables such that for , is measurable, is measurable, and . Further, assume that with probability one. Then, for any ,
(9) 
Proof: This lemma appeared in a slightly more general form (where is allowed) as Lemma A.1 in the paper by Szita & Szepesvári (2011) so we refer the reader to the proof there.
Appendix B Results for Known and Bounded Expected Delay
b.1 High Probability Bounds
For arm and phase , assume . For notational simplicity we will use in the following for any event . If for a particular experiment then . Then for any phase and time , define,
(11) 
and note that since if arm is not active in phase , we have the equalities and . Define the filtration by and
(12) 
Then, we use the decomposition,
(13)  
where,
Recall that the filtration is defined by , and we have defined if arm is eliminated before phase and if the algorithm stops before reaching phase .
Outline of proof
We will bound each term of the above decomposition in (13) in turn, however first we need to prove several intermediary results. For term II., we will use Freedman’s inequality so we first need Lemma 12 to show that is a martingale difference and Lemma 13 to bound the variance of the sum of the ’s. Similarly, for term III., in Lemma 14, we show that is a martingale difference and bound its variance in Lemma 15. In Lemma 16, we consider term IV. and bound the conditional expectations of . Finally, in Lemma 17, we bound term I. using Lemma 11. We then combine the bounds on all terms together to conclude the proof.
Lemma 12
Let
Comments
There are no comments yet.