jmlr jmlr2013 jmlr2013-105 knowledge-graph by maker-knowledge-mining

105 jmlr-2013-Sparsity Regret Bounds for Individual Sequences in Online Linear Regression

Source: pdf

Author: Sébastien Gerchinovitz

Abstract: We consider the problem of online linear regression on arbitrary deterministic sequences when the ambient dimension d can be much larger than the number of time rounds T . We introduce the notion of sparsity regret bound, which is a deterministic online counterpart of recent risk bounds derived in the stochastic setting under a sparsity scenario. We prove such regret bounds for an online-learning algorithm called SeqSEW and based on exponential weighting and data-driven truncation. In a second part we apply a parameter-free version of this algorithm to the stochastic setting (regression model with random design). This yields risk bounds of the same ﬂavor as in Dalalyan and Tsybakov (2012a) but which solve two questions left open therein. In particular our risk bounds are adaptive (up to a logarithmic factor) to the unknown variance of the noise if the latter is Gaussian. We also address the regression model with ﬁxed design. Keywords: sparsity, online linear regression, individual sequences, adaptive regret bounds

Reference: text

Summary: the most important sentenses genereted by tfidf model

sentIndex sentText sentNum sentScore

1 We introduce the notion of sparsity regret bound, which is a deterministic online counterpart of recent risk bounds derived in the stochastic setting under a sparsity scenario. [sent-4, score-0.599]

2 Keywords: sparsity, online linear regression, individual sequences, adaptive regret bounds 1. [sent-10, score-0.349]

3 On the theoretical side, most sparsity-related risk bounds take the form of the so-called sparsity oracle inequalities, that is, risk bounds expressed in terms of the number of non-zero coordinates of the oracle vector. [sent-24, score-0.344]

4 We ﬁrst prove that theoretical guarantees similar to sparsity oracle inequalities can be obtained in a deterministic online setting, namely, online linear regression on individual sequences. [sent-37, score-0.377]

5 The newly obtained deterministic prediction guarantees are called sparsity regret bounds. [sent-38, score-0.385]

6 In the second part of this paper, we apply our sparsity regret bounds—of deterministic nature—to the stochastic setting (regression model with random design). [sent-40, score-0.394]

7 One of our key results is that, thanks to our online tuning techniques, these deterministic bounds imply sparsity oracle inequalities that are adaptive to the unknown variance of the noise (up to logarithmic factors) when the latter is Gaussian. [sent-41, score-0.394]

8 In the next paragraphs, we introduce our main setting and motivate the notion of sparsity regret bound from an online-learning viewpoint. [sent-43, score-0.367]

9 A forecaster has to predict in a sequential fashion the values yt ∈ R of an unknown sequence of observations given some input data xt ∈ X and some base forecasters ϕ j : X → R, 1 j d, on the basis of which he outputs a prediction yt ∈ R. [sent-47, score-1.61]

10 ) In this setting the version of the sequential ridge regression forecaster studied by Azoury and Warmuth (2001) and Vovk (2001) can be tuned to have a regret ∆T,d (u) of order at most d ln T u 2 . [sent-51, score-0.657]

11 Since the regret bound d ln T is optimal in a certain sense (see, e. [sent-53, score-0.482]

12 If the forecaster knew in advance the support J(u∗ ) { j : u∗ = 0} of u∗ , he could apply the same forecaster as above but only to the s-dimensional j linear subspace u ∈ Rd : ∀ j ∈ J(u∗ ), u j = 0 . [sent-59, score-0.36]

13 The regret bound of this “oracle” would be roughly / of order s ln T and thus sublinear in T . [sent-60, score-0.482]

14 Under this sparsity scenario, a sublinear regret thus seems the fact that ε ∈ S with high-probability is only guaranteed via concentration arguments, so it is a consequence of the underlying statistical assumptions. [sent-61, score-0.333]

15 730 S PARSITY R EGRET B OUNDS FOR I NDIVIDUAL S EQUENCES possible, though, of course, the aforementioned regret bound s ln T can only be used as an ideal benchmark (since the support of u∗ is unknown). [sent-62, score-0.482]

16 We call regret bounds of j=1 the above form sparsity regret bounds. [sent-65, score-0.594]

17 2 Related Works in the Stochastic Setting The above regret bound (1) can be seen as a deterministic online counterpart of the so-called sparsity oracle inequalities introduced in the stochastic setting in the past decade. [sent-69, score-0.591]

18 They indeed propose aggregation algorithms which satisfy sparsity oracle inequalities under almost no assumption on the base forecasters (ϕ j ) j , and which can be approximated numerically at a reasonable computational cost for large values of the ambient dimension d. [sent-106, score-0.31]

19 3 Previous Works on Sparsity in the Framework of Individual Sequences To the best of our knowledge, Corollary 2 and its reﬁnements (Corollary 7 and Theorem 10) provide the ﬁrst examples of sparsity regret bounds in the sense of (1). [sent-111, score-0.361]

20 u 732 S PARSITY R EGRET B OUNDS FOR I NDIVIDUAL S EQUENCES For all s ∈ N and all U > 0, T T ∑ (yt − yt )2 − inf t=1 u u 0 1 ∑ s t=1 U yt − u · ϕ(xt ) 2 s + 1 gT,d U, ϕ ∞ , where g grows at most logarithmically in T , d, U, and ϕ ∞ . [sent-114, score-1.055]

21 Indeed, if s ≪ T , X = Rd , and ϕ j (x) = x j , then for any forecaster, there is an individual sequence (xt , yt )1 t T such that the regret of this forecaster on u ∈ Rd : u 0 s and u 1 d is bounded from below by a quantity of order s ln T . [sent-116, score-1.144]

22 Therefore, up to logarithmic factors, any algorithm satisfying a sparsity regret bound of the form (1) is minimax optimal on intersections of ℓ0 -balls (of radii s ≪ T ) and ℓ1 -balls. [sent-117, score-0.39]

23 They focus on algorithms which output sparse linear combinations, while we are interested in algorithms whose regret is small under a sparsity scenario, that is, on ℓ0 -balls of small radii. [sent-120, score-0.333]

24 In the particular case of ℓ1 -regularization under the square loss, the aforementioned works propose algorithms which predict as a sparse linear combination yt = ut · ϕ(xt ) of the base forecasts (i. [sent-125, score-0.569]

25 However they prove bounds on the ℓ1 -regularized regret of the form T ∑ t=1 (yt − ut · xt )2 + λ ut T 1 inf u∈Rd ∑ t=1 (yt − u · xt )2 + λ u 1 + ∆T,d (u) , (3) for some regret term ∆T,d (u) which is suboptimal on intersections of ℓ0 - and ℓ1 -balls as explained below. [sent-128, score-1.215]

26 This regret bound grows as a power of and not j=1 j ∞ logarithmically in d as is expected for sparsity regret bounds (recall that we are interested in the case when d ≫ T ). [sent-132, score-0.628]

27 The three other papers mentioned above do prove (some) regret bounds with a logarithmic dependence in d, but these bounds do not have the dependence in u 1 and T we are looking for. [sent-133, score-0.312]

28 If the base forecasts ϕ j (xt ) are dense in the sense that c2 ≈ d ϕ d d 2 ∞, then we have 2c2 /T , which yields a regret bound with leading constant 1 as in (3) and with ∆T,d (u) at least of order d √ c2 T ≈ ϕ ∞ dT . [sent-139, score-0.321]

29 A question left open is thus whether it is possible to design an algorithm which both ouputs sparse linear combinations (which is statistically useful and sometimes essential for computational issues) and satisﬁes a sparsity regret bound of the form (1). [sent-146, score-0.399]

30 4 PAC-Bayesian Analysis in the Framework of Individual Sequences To derive our sparsity regret bounds, we follow a PAC-Bayesian approach combined with the choice of a sparsity-favoring prior. [sent-148, score-0.333]

31 6 of Audibert (2009) to derive a non-adaptive sparsity regret bound. [sent-156, score-0.333]

32 Thanks to the standard online-to-batch conversion, our sparsity regret bounds—of deterministic nature—imply a sparsity oracle inequality of the same ﬂavor as a result of Dalalyan and Tsybakov (2012a). [sent-164, score-0.57]

33 In Section 3 we prove the aforementioned sparsity regret bounds for our algorithm SeqSEW, ﬁrst when the forecaster has access to some a priori knowledge on the observations (Sections 3. [sent-174, score-0.541]

34 The data sequence (xt , yt )t 1 at hand is deterministic and arbitrary and we look for theoretical guarantees that hold for every individual sequence. [sent-184, score-0.549]

35 The forecaster chooses a prediction yt ∈ R (possibly as a linear combination of the ϕ j (xt ), but this is not necessary). [sent-197, score-0.693]

36 Each linear forecaster u · ϕ ∑d u j ϕ j , u ∈ Rd , incurs the loss yt − u · ϕ(xt ) j=1 forecaster incurs the loss (yt − yt )2 . [sent-201, score-1.348]

37 Sparsity Regret Bounds for Individual Sequences In this section we prove sparsity regret bounds for different variants of our algorithm SeqSEW. [sent-232, score-0.361]

38 Get the input data xt and predicta as yt Rd u · ϕ(xt ) p (du) B t ; 2. [sent-255, score-0.821]

39 Get the observation yt and compute the posterior distribution pt+1 ∈ M1+ (Rd ) as t exp −η ∑ ys − u · ϕ(xs ) s=1 pt+1 (du) 2 B πτ (du) , Wt+1 where t Wt+1 Rd exp −η ∑ ys − v · ϕ(xs ) s=1 2 B πτ (dv) . [sent-256, score-0.604]

40 Then, for all B By , all η 1/(8B2 ), and all τ > 0, the algorithm SeqSEWB,η τ satisﬁes T T ∑ (yt − yt )2 inf u∈Rd t=1 ∑ t=1 yt − u · ϕ(xt ) 2 + 4 u η 0 ln 1 + d u 1 u 0τ T + τ2 ∑ ∑ ϕ2 (xt ) . [sent-267, score-1.27]

41 j=1 j Then, when used with B = By , η = T T ∑ (yt − yt ) t=1 1 , and τ = 8B2 y 2 inf u∈Rd ∑ t=1 yt − u · ϕ(xt ) 2 16B2 y , the algorithm SeqSEWB,η satisﬁes τ BΦ + 32 B2 y u 0 ln 1+ √ BΦ u 1 4 By u 0 + 16B2 . [sent-275, score-1.27]

42 y (5) Note that, if ϕ ∞ supx∈X max1 j d |ϕ j (x)| is ﬁnite, then the last corollary provides a sparsity regret bound in the sense of (1). [sent-276, score-0.414]

43 For all τ > 0, if the algorithm SeqSEWB,η is used with B By and η 1/(8B2 ), then τ T T ∑ (yt − yt )2 t=1 inf + ρ∈M1 (Rd ) Rd ∑ t=1 T inf + ρ∈M1 (Rd ) Rd ∑ t=1 2 yt − u · ϕ(xt ) B ρ(du) + 2 yt − u · ϕ(xt ) ρ(du) + K (ρ, πτ ) (6) η K (ρ, πτ ) η . [sent-287, score-1.616]

44 Therefore, truncation to [−B, B] can only improve prediction under the square loss if the observations are [−B, B]-valued, which is the case here since by assumption yt ∈ [−By , By ] ⊂ [−B, B] for all t = 1, . [sent-297, score-0.553]

45 Since B By and η 1/(8B2 ), we can apply Lemma 3 and get T T ∑ (yt − yt )2 t=1 inf + ρ∈M1 (Rd ) 2 ∑ yt − u · ϕ(xt ) ρ(du) + Rd t=1 T ∑ Rd t=1 2 yt − u · ϕ(xt ) ρu∗ ,τ (du) + K (ρ, πτ ) η K (ρu∗ ,τ , πτ ) η . [sent-315, score-1.569]

46 (8) (2) (1) In the last inequality, ρu∗ ,τ is taken as the translated of πτ at u∗ , namely, ρu∗ ,τ (du) d (3/τ) du j dπτ (u − u∗ ) du = ∏ ∗ du j=1 2 1 + |u j − u j |/τ 4 . [sent-316, score-0.333]

47 2), the ﬁrst term (1) can be rewritten as T ∑ Rd t=1 T 2 2 yt − u · ϕ(xt ) ρu∗ ,τ (du) = ∑ yt − u∗ · ϕ(xt ) d T + τ2 ∑ ∑ ϕ2 (xt ) . [sent-319, score-0.988]

48 2, the sparsity-related term 4 u 1 u 0 ln 1 + u 0τ η in the regret bound of Proposition 1 can actually be replaced with the smaller (and continous) term 4 η d ∑ ln (1 + |u j |/τ) . [sent-326, score-0.697]

49 In this section, we remove the ﬁrst requirement and prove a sparsity regret bound for a variant of the algorithm SeqSEWB,η which is adaptive to the unknown bound τ By = max1 t T |yt |; see Proposition 5 and Remark 6 below. [sent-330, score-0.443]

50 Proposition 5 For all τ > 0, the algorithm SeqSEW∗ satisﬁes τ T T ∑ (yt − yt )2 t=1 inf u∈Rd ∑ t=1 d + τ2 ∑ yt − u · ϕ(xt ) 2 + 32B2 +1 u T 0 ln 1+ u 1 u 0τ T ∑ ϕ2j (xt ) + 5B2 +1 , T j=1 t=1 where B2 +1 T max1 2 t T yt . [sent-337, score-1.764]

51 Remark 6 In view of Proposition 1, the algorithm SeqSEW∗ satisﬁes a sparsity regret bound which τ is adaptive to the unknown bound By = max1 t T |yt |. [sent-338, score-0.443]

52 Get the input data xt and predicta as yt Rd u · ϕ(xt ) Bt pt (du); 2. [sent-343, score-0.857]

53 Then, when used with τ = 1/ BΦ , the algorithm SeqSEW∗ satisﬁes τ j=1 j T ∑ (yt − yt ) t=1 2 T inf u∈Rd ∑ yt − u · ϕ(xt ) t=1 2 + 5BT +1 + 1 2 + 32B2 +1 T u 0 ln 1+ √ BΦ u u 0 1 , √ 6. [sent-357, score-1.27]

54 is Then, when used with τ = 1/ dT , the algorithm SeqSEW∗ satisﬁes τ T T ∑ (yt − yt ) 2 u∈Rd t=1 + where B2 +1 T max1 ∑ inf t=1 + 32B2 +1 T u 1+ 0 ln √ dT u u 0 1 T d 1 dT yt − u · ϕ(xt ) 2 ∑ ∑ ϕ2j (xt ) + 5B2 +1 , T j=1 t=1 2 t T yt . [sent-363, score-1.764]

55 As for the second term (2), by deﬁnition of Wt+1 , + ′ 1 Wt+1 ln ηt Wt = = 1 ln ηt 1 ln ηt exp −ηt yt − u · ϕ(xt ) 2 Bt s=1 2 Bs πτ (du) Wt Rd Rd t−1 exp −ηt ∑ ys − u · ϕ(xs ) exp −ηt yt − u · ϕ(xt ) 2 Bt pt (du) . [sent-388, score-1.724]

56 First note that this is straightforward in the particular case where yt ∈ [−Bt , Bt ]. [sent-391, score-0.494]

57 Taking the logarithms of both sides of the last inequality and dividing by ηt , we can see that the quantity on the right-hand side of (15) is bounded 2 from above by − yt − yt . [sent-393, score-1.049]

58 In the general case, we cannot assume that yt ∈ [−Bt , Bt ], since it may happen that |yt | > max1 s t−1 |ys | Bt . [sent-394, score-0.494]

59 As shown below, we can still use the exp-concavity of the square loss if we replace yt with its clipped version [yt ]Bt . [sent-395, score-0.515]

60 More precisely, setting yt,u [u · ϕ(xt )]Bt for all u ∈ Rd , the square loss appearing in the right-hand side of (15) equals yt − yt,u 2 = [yt ]Bt − yt,u = [yt ]Bt − yt,u 2 2 + yt − [yt ]Bt + yt − [yt ]Bt 2 2 + 2 yt − [yt ]Bt [yt ]Bt − yt,u + 2 yt − [yt ]Bt [yt ]Bt − yt + ct,u , (16) 7. [sent-396, score-2.985]

61 743 G ERCHINOVITZ where we set 2 yt − [yt ]Bt ct,u yt − yt,u −4Bt yt − [yt ]Bt −4Bt (Bt+1 − Bt ) , (17) where the last two inequalities follow from the property yt , yt,u ∈ [−Bt , Bt ] (by construction) and Bt+1 − Bt . [sent-399, score-2.011]

62 from the elementary8 yet useful upper bound yt − [yt ]Bt Combining (16) with the lower bound (17) yields that, for all u ∈ Rd , 2 yt − yt,u [yt ]Bt − yt,u 2 +Ct , (18) 2 where we set Ct yt − [yt ]Bt + 2 yt − [yt ]Bt [yt ]Bt − yt − 4Bt (Bt+1 − Bt ). [sent-400, score-2.538]

63 , T and using the upper bound Bt (Bt+1 − Bt ) Bt+1 − Bt2 , Equation (14) yields T T lnWT +1 lnW1 − ηT +1 η1 2 − ∑ (yt − yt )2 + 4 ∑ Bt+1 − Bt2 t=1 t=1 T = − ∑ (yt − yt )2 + 4B2 +1 . [sent-406, score-1.022]

64 T (22) t=1 Third step: Putting (13) and (22) together, we get the PAC-Bayesian inequality T T ∑ (yt − yt )2 t=1 inf + ρ∈M1 (Rd ) Rd ∑ yt − u · ϕ(xt ) t=1 2 Bt ρ(du) + K (ρ, πτ ) ηT +1 + 4B2 +1 , T which yields (11) since ηT +1 1/(8B2 +1 ) by deﬁnition. [sent-407, score-1.115]

65 , T , yt − [u · ϕ(xt )]Bt 2 yt − u · ϕ(xt ) 2 + (Bt+1 − Bt )2 8. [sent-411, score-0.988]

66 To see why this is true, it sufﬁces to rewrite [yt ]Bt in the three cases yt < −Bt , |yt | 9. [sent-412, score-0.494]

67 If BT +1 = 0, then yt = yt = 0 for all 1 t T , which immediately yields (11). [sent-415, score-0.988]

68 We can thus assume that |yt | > Bt , or just11 that yt > Bt . [sent-421, score-0.494]

69 If yt < −Bt , it sufﬁces to apply (23) with −yt and −u. [sent-439, score-0.494]

70 This yields a sparsity regret bound with extra logarithmic multiplicative factors as compared to Proposition 5, but which holds for a fully automatic algorithm; see Theorem 10 below. [sent-443, score-0.407]

71 , (xT , yT ) ∈ X × R, ∗   T ∑ (yt − yt )2 t=1 T inf u∈Rd ∑ t=1 yt − u · ϕ(xt ) 2 + 128 max yt2 1 t T + 32 max yt2 AT u 1 t T where AT 2 + log2 ln e + u 0 ln u u 1+ 1  0 ln e + T d ∑ ∑ ϕ2j (xt )  t=1 j=1 + 1 + 9 max yt2 AT , 0 1 t T T ∑t=1 ∑d ϕ2 (xt ) . [sent-469, score-1.778]

72 It upper bounds the regret of the algorithm SeqSEW∗ ∗ uniformly over all u ∈ Rd such that u 0 s and u 1 U, where the sparsity level s ∈ N and the ℓ1 -diameter U > 0 are both unknown to the forecaster. [sent-484, score-0.384]

73 The sparsity regret bounds proved for this algorithm on individual sequences imply in both settings sparsity oracle inequalities with leading constant 1. [sent-497, score-0.605]

74 2004), we deﬁne our estimator fT : X → R as the uniform average fT 1 T ∑ ft T t=1 (25) of the estimators ft : X → R sequentially built by the algorithm SeqSEW∗ as τ ft (x) Rd u · ϕ(x) Bt pt (du) . [sent-540, score-0.876]

75 ft (Xt ) in Proof sketch (of Theorem 12) By Corollary 8 and by deﬁnition of ft above and yt Figure 3, we have, almost surely, T T ∑ (Yt − ft (Xt )) 2 t=1 inf u∈Rd + 1 dT ∑ t=1 d Yt − u · ϕ(Xt ) 2 + 32 max 1 t T Yt2 √ dT u u 0 ln 1 + u 0 1 T ∑ ∑ ϕ2j (Xt ) + 5 1maxT Yt2 . [sent-560, score-1.642]

76 The above bound is of the same order (up to a ln T factor) as the sparsity oracle inequality proved in Proposition 1 of Dalalyan and Tsybakov (2012a). [sent-634, score-0.453]

77 , (xT ,YT ) ∈ X × R, where the xt are deterministic elements in X and where Yt = f (xt ) + εt , 1 t T, (29) for some i. [sent-668, score-0.36]

78 , xT } , / T where nx t : xt = x = ∑t=1 I{xt =x} , and where the estimators ft : X → R sequentially built by the algorithm SeqSEW∗ are deﬁned by τ ft (x) Rd u · ϕ(x) 753 Bt pt (du) . [sent-692, score-0.955]

79 (31) G ERCHINOVITZ In the particular case when the xt are all distinct, fT is simply deﬁned by fT (xt ) ft (xt ) for all t ∈ {1, . [sent-693, score-0.607]

80 It follows as in the random design setting from the deterministic regret bound of Corollory 8 and from Jensen’s inequality. [sent-702, score-0.332]

81 , yt−1 but only on the basis of τ the past observations yt , t ∈ {tr−1 + 1, . [sent-734, score-0.511]

82 ,tr − 1} is upper bounded by tr −1 ∑ d ∑ ϕ2j (xt ) t=tr−1 +1 j=1 tr −1 d r ∑ ∑ ϕ2j (xt ) (e2 − 1)2 , t=1 j=1 where the last inequality follows from the fact that γtr −1 2r (by deﬁnition of tr ). [sent-755, score-0.334]

83 First note that, by the r=0 T upper bound B2 r y∗ 2 and by the elementary inequality ln(1 + xy) ln ((1 + x)(1 + y)) = ln(1 + r,t T 14. [sent-769, score-0.309]

84 In the trivial cases where tr = tr−1 + 1 for some r, the sum ∑tr −1 +1 (yt − yt )2 equals 0 by convention. [sent-770, score-0.592]

85 0 First case: R = 0 Substituting (37) in (36), we conclude the proof by noting that AT T ∑t=1 ∑d ϕ2 (xt ) j=1 j + 5y∗ 2 + 1 T 1 4 · 2R−1  ln e + 4 ln e + T  d ∑ ∑ ϕ2j (xt )  . [sent-777, score-0.45]

86 T Substituting the last two inequalities in (36) and noting that y∗ 2 = max1 T 2 t T yt concludes the proof. [sent-780, score-0.549]

87 In view of Theorem 10, we just need to check that the quantity (continuously extended in s = 0)   T d U 128 max yt2 s ln e + ∑ ∑ ϕ2 (xt )  + 32 max yt2 AT s ln 1 + j 1 t T 1 t T s t=1 j=1 is non-decreasing in s ∈ R+ and in U ∈ R+ . [sent-782, score-0.482]

88 G ERCHINOVITZ From the elementary inequality 1 1+u ln(1 + u) = − ln − u 1 −1 = , 1+u 1+u which holds for all u ∈ (−1, +∞), the above derivative is nonnegative for all s > 0 so that the continuous extension s ∈ R+ → s ln (1 +U/s) is non-decreasing. [sent-786, score-0.49]

89 ft (Xt ) in Figure 3, Proof (of Theorem 12) By Corollory 8 and the deﬁnitions of ft in (26) and yt we have, almost surely, T T ∑ (Yt − ft (Xt )) t=1 2 inf u∈Rd + 1 dT ∑ t=1 2 Yt − u · ϕ(Xt ) + 32 max d 1 t T √ dT u u 0 ln 1+ u 0 Yt2 1 T ∑ ∑ ϕ2j (Xt ) + 5 1maxT Yt2 . [sent-793, score-1.642]

90 But by deﬁnition of fT and by convexity of the square loss, E f − fT 2 E L2 f (X) − 1 T ∑E T t=1 2 1 T ∑ ft (X) T t=1 f (X) − ft (X) 2 = 1 T ∑E T t=1 f (Xt ) − ft (Xt ) The last equality follows classically from the fact that, for all t = 1, . [sent-807, score-0.861]

91 1 Applying Corollory 8 we have, almost surely, T ∑ t=1 Yt − ft (xt ) 2 T ∑ inf u∈Rd + t=1 1 dT Yt − u · ϕ(xt ) 2 + 32 max 1 t T Yt2 u 1+ 0 ln √ dT u u 0 1 T d ∑ ∑ ϕ2j (xt ) + 5 1maxT Yt2 . [sent-841, score-0.588]

92 T t=1 T t=1 This is an equality if the xt are all distinct. [sent-845, score-0.327]

93 ,xT } nx f (x) − fT (x) 2 2 1 f (x) − ft (x) nx 1 ∑ T t t:xt =x nx 1 f (x) − ft (x) nx 1 ∑ T t t:xt =x 2 T = ∑ f (xt ) − ft (xt ) 2 , t=1 15. [sent-859, score-0.968]

94 16 For any u∗ ∈ Rd and τ > 0, deﬁne ρu∗ ,τ as the translated of πτ at u∗ , namely, ρu∗ ,τ d (3/τ) du j dπτ (u − u∗ ) du = ∏ ∗ du j=1 2 1 + |u j − u j |/τ 4 . [sent-879, score-0.333]

95 , yT ) and the prediction loss f − fu 2 n 2 T is replaced with the cumulative loss ∑t=1 yt − u · ϕ(xt ) . [sent-888, score-0.513]

96 762 S PARSITY R EGRET B OUNDS FOR I NDIVIDUAL S EQUENCES Lemma 22 For all u∗ ∈ Rd and τ > 0, the probability distribution ρu∗ ,τ satisﬁes T ∑ Rd t=1 T 2 yt − u · ϕ(xt ) ρu∗ ,τ (du) = ∑ yt − u∗ · ϕ(xt ) 2 t=1 T d + τ2 ∑ ∑ ϕ2 (xt ) . [sent-890, score-0.988]

97 763 G ERCHINOVITZ Proof (of Lemma 23) By deﬁnition of ρu∗ ,τ and πτ , we have K (ρu∗ ,τ , πτ ) Rd dρu∗ ,τ ln (u) ρu∗ ,τ (du) = dπτ d =4 Rd 1 + |u j |/τ ∑ ln 1 + |u j − u∗ |/τ d Rd j=1 4 1 + |u j − u∗ |/τ j 4 ρu∗ ,τ (du) ρu∗ ,τ (du) . [sent-903, score-0.43]

98 (45) j j=1 But, for all u ∈ Rd , by the triangle inequality, 1 + |u j |/τ ln ∏ 1 + |u j |/τ 1 + |u∗ |/τ + |u j − u∗ |/τ j j 1 + |u∗ |/τ 1 + |u j − u∗ |/τ , j j so that Equation (45) yields the upper bound d K (ρu∗ ,τ , πτ ) 4 ∑ ln j=1 1 + |u∗ |/τ = 4 j ∑ j:u∗ =0 j ln 1 + |u∗ |/τ . [sent-904, score-0.679]

99 j { j : u∗ = 0} and apply Jensen’s inequality to the concave function We now recall that u∗ 0 j x ∈ (−1, +∞) −→ ln(1 + x) to get ∑ j:u∗ =0 j ln 1 + |u∗ |/τ j ∗ = u u∗ 0 1 u∗ 0 ln ∑ ln 0 j:u∗ =0 j 1+ 1 + |u∗ |/τ j u∗ 1 u∗ 0 τ ∗ u 0 ln 1+ ∑ j:u∗j =0 |u∗ | j u∗ 0τ . [sent-905, score-0.92]

100 Sparsity regret bounds for individual sequences in online linear regression. [sent-1194, score-0.353]

similar papers computed by tfidf model

tfidf for this paper:

wordName wordTfidf (topN-words)

[('yt', 0.494), ('xt', 0.327), ('ft', 0.28), ('seqsew', 0.244), ('regret', 0.233), ('bt', 0.228), ('ln', 0.215), ('dalalyan', 0.198), ('rd', 0.188), ('forecaster', 0.18), ('tsybakov', 0.138), ('egret', 0.128), ('equences', 0.128), ('erchinovitz', 0.128), ('ndividual', 0.128), ('du', 0.111), ('parsity', 0.11), ('sparsity', 0.1), ('tr', 0.098), ('dt', 0.095), ('lnwt', 0.077), ('ounds', 0.076), ('inf', 0.067), ('oracle', 0.064), ('ys', 0.055), ('forecasters', 0.049), ('corollary', 0.047), ('online', 0.047), ('bunea', 0.046), ('bem', 0.045), ('subgaussian', 0.044), ('wt', 0.042), ('xs', 0.04), ('inequality', 0.04), ('seqsewb', 0.038), ('ytr', 0.038), ('aggregation', 0.038), ('proposition', 0.037), ('pt', 0.036), ('inequalities', 0.035), ('clipping', 0.035), ('bound', 0.034), ('deterministic', 0.033), ('design', 0.032), ('sg', 0.032), ('nx', 0.032), ('corollory', 0.032), ('gerchinovitz', 0.032), ('vovk', 0.032), ('issn', 0.031), ('jensen', 0.031), ('risk', 0.03), ('zt', 0.03), ('forecasts', 0.03), ('catoni', 0.03), ('regression', 0.029), ('stochastic', 0.028), ('bounds', 0.028), ('max', 0.026), ('birg', 0.026), ('ottucs', 0.026), ('bm', 0.025), ('massart', 0.025), ('base', 0.024), ('logarithmic', 0.023), ('bd', 0.023), ('sequences', 0.023), ('unknown', 0.023), ('lounici', 0.023), ('remark', 0.023), ('tuning', 0.022), ('mum', 0.022), ('individual', 0.022), ('geer', 0.022), ('audibert', 0.022), ('gram', 0.021), ('sides', 0.021), ('square', 0.021), ('get', 0.02), ('proof', 0.02), ('nondecreasing', 0.02), ('avor', 0.02), ('mirror', 0.02), ('postponed', 0.02), ('lemma', 0.02), ('concludes', 0.02), ('elementary', 0.02), ('game', 0.019), ('adaptivity', 0.019), ('kivinen', 0.019), ('prediction', 0.019), ('truncation', 0.019), ('regime', 0.019), ('adaptive', 0.019), ('truncating', 0.018), ('temperature', 0.018), ('bs', 0.018), ('surely', 0.018), ('automatic', 0.017), ('past', 0.017)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 1.0000013 105 jmlr-2013-Sparsity Regret Bounds for Individual Sequences in Online Linear Regression

Author: Sébastien Gerchinovitz

2 0.13139461 35 jmlr-2013-Distribution-Dependent Sample Complexity of Large Margin Learning

Author: Sivan Sabato, Nathan Srebro, Naftali Tishby

Abstract: We obtain a tight distribution-speciﬁc characterization of the sample complexity of large-margin classiﬁcation with L2 regularization: We introduce the margin-adapted dimension, which is a simple function of the second order statistics of the data distribution, and show distribution-speciﬁc upper and lower bounds on the sample complexity, both governed by the margin-adapted dimension of the data distribution. The upper bounds are universal, and the lower bounds hold for the rich family of sub-Gaussian distributions with independent features. We conclude that this new quantity tightly characterizes the true sample complexity of large-margin classiﬁcation. To prove the lower bound, we develop several new tools of independent interest. These include new connections between shattering and hardness of learning, new properties of shattering with linear classiﬁers, and a new lower bound on the smallest eigenvalue of a random Gram matrix generated by sub-Gaussian variables. Our results can be used to quantitatively compare large margin learning to other learning rules, and to improve the effectiveness of methods that use sample complexity bounds, such as active learning. Keywords: supervised learning, sample complexity, linear classiﬁers, distribution-dependence

3 0.11856709 114 jmlr-2013-The Rate of Convergence of AdaBoost

Author: Indraneel Mukherjee, Cynthia Rudin, Robert E. Schapire

Abstract: The AdaBoost algorithm was designed to combine many “weak” hypotheses that perform slightly better than random guessing into a “strong” hypothesis that has very low error. We study the rate at which AdaBoost iteratively converges to the minimum of the “exponential loss.” Unlike previous work, our proofs do not require a weak-learning assumption, nor do they require that minimizers of the exponential loss are ﬁnite. Our ﬁrst result shows that the exponential loss of AdaBoost’s computed parameter vector will be at most ε more than that of any parameter vector of ℓ1 -norm bounded by B in a number of rounds that is at most a polynomial in B and 1/ε. We also provide lower bounds showing that a polynomial dependence is necessary. Our second result is that within C/ε iterations, AdaBoost achieves a value of the exponential loss that is at most ε more than the best possible value, where C depends on the data set. We show that this dependence of the rate on ε is optimal up to constant factors, that is, at least Ω(1/ε) rounds are necessary to achieve within ε of the optimal exponential loss. Keywords: AdaBoost, optimization, coordinate descent, convergence rate

4 0.1151721 65 jmlr-2013-Lower Bounds and Selectivity of Weak-Consistent Policies in Stochastic Multi-Armed Bandit Problem

Author: Antoine Salomon, Jean-Yves Audibert, Issam El Alaoui

Abstract: This paper is devoted to regret lower bounds in the classical model of stochastic multi-armed bandit. A well-known result of Lai and Robbins, which has then been extended by Burnetas and Katehakis, has established the presence of a logarithmic bound for all consistent policies. We relax the notion of consistency, and exhibit a generalisation of the bound. We also study the existence of logarithmic bounds in general and in the case of Hannan consistency. Moreover, we prove that it is impossible to design an adaptive policy that would select the best of two algorithms by taking advantage of the properties of the environment. To get these results, we study variants of popular Upper Conﬁdence Bounds (UCB) policies. Keywords: stochastic bandits, regret lower bounds, consistency, selectivity, UCB policies 1. Introduction and Notations Multi-armed bandits are a classical way to illustrate the difﬁculty of decision making in the case of a dilemma between exploration and exploitation. The denomination of these models comes from an analogy with playing a slot machine with more than one arm. Each arm has a given (and unknown) reward distribution and, for a given number of rounds, the agent has to choose one of them. As the goal is to maximize the sum of rewards, each round decision consists in a trade-off between exploitation (i.e., playing the arm that has been the more lucrative so far) and exploration (i.e., testing another arm, hoping to discover an alternative that beats the current best choice). One possible application is clinical trial: a doctor wants to heal as many patients as possible, the patients arrive sequentially, and the effectiveness of each treatment is initially unknown (Thompson, 1933). Bandit problems have initially been studied by Robbins (1952), and since then they have been applied to many ﬁelds such as economics (Lamberton et al., 2004; Bergemann and Valimaki, 2008), games (Gelly and Wang, 2006), and optimisation (Kleinberg, 2005; Coquelin and Munos, 2007; Kleinberg et al., 2008; Bubeck et al., 2009). ∗. Also at Willow, CNRS/ENS/INRIA - UMR 8548. c 2013 Antoine Salomon, Jean-Yves Audibert and Issam El Alaoui. S ALOMON , AUDIBERT AND E L A LAOUI 1.1 Setting In this paper, we consider the following model. A stochastic multi-armed bandit problem is deﬁned by: • a number of rounds n, • a number of arms K ≥ 2, • an environment θ = (ν1 , · · · , νK ), where each νk (k ∈ {1, · · · , K}) is a real-valued measure that represents the distribution reward of arm k. The number of rounds n may or may not be known by the agent, but this will not affect the present study. We assume that rewards are bounded. Thus, for simplicity, each νk is a probability on [0, 1]. Environment θ is initially unknown by the agent but lies in some known set Θ. For the problem to be interesting, the agent should not have great knowledges of its environment, so that Θ should not be too small and/or only contain too trivial distributions such as Dirac measures. To make it simple, we may assume that Θ contains all environments where each reward distribution is a Dirac distribution or a Bernoulli distribution. This will be acknowledged as Θ having the Dirac/Bernoulli property. For technical reason, we may also assume that Θ is of the form Θ1 × . . . × ΘK , meaning that Θk is the set of possible reward distributions of arm k. This will be acknowledged as Θ having the product property. The game is as follows. At each round (or time step) t = 1, · · · , n, the agent has to choose an arm It in the set {1, · · · , K}. This decision is based on past actions and observations, and the agent may also randomize his choice. Once the decision is made, the agent gets and observes a reward that is drawn from νIt independently from the past. Thus a policy (or strategy) can be described by a sequence (σt )t≥1 (or (σt )1≤t≤n if the number of rounds n is known) such that each σt is a mapping from the set {1, . . . , K}t−1 × [0, 1]t−1 of past decisions and rewards into the set of arm {1, . . . , K} (or into the set of probabilities on {1, . . . , K}, in case the agent randomizes his choices). For each arm k and all time step t, let Tk (t) = ∑ts=1 ½Is =k denote the sampling time, that is, the number of times arm k was pulled from round 1 to round t, and Xk,1 , Xk,2 , . . . , Xk,Tk (t) the corresponding sequence of rewards. We denote by Pθ the distribution on the probability space such that for any k ∈ {1, . . . , K}, the random variables Xk,1 , Xk,2 , . . . , Xk,n are i.i.d. realizations of νk , and such that these K sequences of random variables are independent. Let Eθ denote the associated expectation. Let µk = xdνk (x) be the mean reward of arm k. Introduce µ∗ = maxk∈{1,...,K} µk and ﬁx an arm ∗ ∈ argmax ∗ k k∈{1,...,K} µk , that is, k has the best expected reward. The agent aims at minimizing its regret, deﬁned as the difference between the cumulative reward he would have obtained by always drawing the best arm and the cumulative reward he actually received. Its regret is thus n n Rn = ∑ Xk∗ ,t − ∑ XIt ,TIt (t) . t=1 t=1 As most of the publications on this topic, we focus on the expected regret, for which one can check that: K E θ Rn = ∑ ∆k Eθ [Tk (n)], k=1 188 (1) L OWER B OUNDS , W EAK C ONSISTENCY AND S ELECTIVITY IN BANDIT P ROBLEMS where ∆k is the optimality gap of arm k, deﬁned by ∆k = µ∗ − µk . We also deﬁne ∆ as the gap between the best arm and the second best arm, that is, ∆ := mink:∆k >0 ∆k . Other notions of regret exist in the literature. One of them is the quantity n max ∑ Xk,t − XIt ,TIt (t) , k t=1 which is mostly used in adversarial settings. Results and ideas we want to convey here are more suited to expected regret, and considering other deﬁnitions of regret would only bring some more technical intricacies. 1.2 Consistency and Regret Lower Bounds Former works have shown the existence of lower bounds on the expected regret of a large class of policies: intuitively, to perform well the agent has to explore all arms, and this requires a signiﬁcant amount of suboptimal choices. In this way, Lai and Robbins (1985) proved a lower bound of order log n in a particular parametric framework, and they also exhibited optimal policies. This work has then been extended by Burnetas and Katehakis (1996). Both papers deal with consistent policies, meaning that they only consider policies such that: ∀a > 0, ∀θ ∈ Θ, Eθ [Rn ] = o(na ). (2) Let us detail the bound of Burnetas and Katehakis, which is valid when Θ has the product property. Given an environment θ = (ν1 , · · · , νK ) and an arm k ∈ {1, . . . , K}, deﬁne: Dk (θ) := inf ˜ νk ∈Θk :E[˜ k ]>µ∗ ν ˜ KL(νk , νk ), where KL(ν, µ) denotes the Kullback-Leibler divergence of measures ν and µ. Now ﬁx a consistent policy and an environment θ ∈ Θ. If k is a suboptimal arm (i.e., µk = µ∗ ) such that 0 < Dk (θ) < ∞, then (1 − ε) log n ∀ε > 0, lim P Tk (n) ≥ = 1. n→+∞ Dk (θ) This readily implies that: lim inf n→+∞ Eθ [Tk (n)] 1 ≥ . log n Dk (θ) Thanks to Formula (1), it is then easy to deduce a lower bound of the expected regret. One contribution of this paper is to generalize the study of regret lower bounds, by considering weaker notions of consistency: α-consistency and Hannan consistency. We will deﬁne αconsistency (α ∈ [0, 1)) as a variant of Equation (2), where equality Eθ [Rn ] = o(na ) only holds for all a > α. We show that the logarithmic bound of Burnetas and Katehakis still holds, but coefﬁcient 1−α 1 Dk (θ) is turned into Dk (θ) . We also prove that the dependence of this new bound with respect to the term 1 − α is asymptotically optimal when n → +∞ (up to a constant). We will also consider the case of Hannan consistency. Indeed, any policy achieves at most an expected regret of order n: because of the equality ∑K Tk (n) = n and thanks to Equation (1), one k=1 can show that Eθ Rn ≤ n maxk ∆k . More intuitively, this comes from the fact that the average cost of pulling an arm k is a constant ∆k . As a consequence, it is natural to wonder what happens when 189 S ALOMON , AUDIBERT AND E L A LAOUI dealing with policies whose expected regret is only required to be o(n), which is equivalent to Hannan consistency. This condition is less restrictive than any of the previous notions of consistency. In this larger class of policy, we show that the lower bounds on the expected regret are no longer logarithmic, but can be much smaller. Finally, even if no logarithmic lower bound holds on the whole set Θ, we show that there necessarily exist some environments θ for which the expected regret is at least logarithmic. The latter result is actually valid without any assumptions on the considered policies, and only requires a simple property on Θ. 1.3 Selectivity As we exhibit new lower bounds, we want to know if it is possible to provide optimal policies that achieve these lower bounds, as it is the case in the classical class of consistent policies. We answer negatively to this question, and for this we solve the more general problem of selectivity. Given a set of policies, we deﬁne selectivity as the ability to perform at least as good as the policy that is best suited to the current environment θ. Still, such an ability can not be implemented. As a by-product it is not possible to design a procedure that would speciﬁcally adapt to some kinds of environments, for example by selecting a particular policy. This contribution is linked with selectivity in on-line learning problem with perfect information, commonly addressed by prediction with expert advice (see, e.g., Cesa-Bianchi et al., 1997). In this spirit, a closely related problem is the one of regret against the best strategy from a pool studied by Auer et al. (2003). The latter designed an algorithm in the context of adversarial/nonstochastic bandit whose decisions are based on a given number of recommendations (experts), which are themselves possibly the rewards received by a set of given policies. To a larger extent, model selection has been intensively studied in statistics, and is commonly solved by penalization methods (Mallows, 1973; Akaike, 1973; Schwarz, 1978). 1.4 UCB Policies Some of our results are obtained using particular Upper Conﬁdence Bound algorithms. These algorithms were introduced by Lai and Robbins (1985): they basically consist in computing an index for each arm, and selecting the arm with the greatest index. A simple and efﬁcient way to design such policies is as follows: choose each index as low as possible such that, conditional to past observations, it is an upper bound of the mean reward of the considered arm with high probability (or, say, with high conﬁdence level). This idea can be traced back to Agrawal (1995), and has been popularized by Auer et al. (2002), who notably described a policy called UCB1. In this policy, each index Bk,s,t is deﬁned by an arm k, a time step t, an integer s that indicates the number of times arm k has been pulled before round t, and is given by: ˆ Bk,s,t = Xk,s + 2 logt , s ˆ ˆ where Xk,s is the empirical mean of arm k after s pulls, that is, Xk,s = 1 ∑s Xk,u . s u=1 To summarize, UCB1 policy ﬁrst pulls each arm once and then, at each round t > K, selects an arm k that maximizes Bk,Tk (t−1),t . Note that, by means of Hoeffding’s inequality, the index Bk,Tk (t−1),t is indeed an upper bound of µk with high probability (i.e., the probability is greater than 1 − 1/t 4 ). ˆ Another way to understand this index is to interpret the empiric mean Xk,Tk (t−1) as an ”exploitation” 190 L OWER B OUNDS , W EAK C ONSISTENCY AND S ELECTIVITY IN BANDIT P ROBLEMS term, and the square root 2 logt/s as an ”exploration” term (because the latter gradually increases when arm k is not selected). Policy UCB1 achieves the logarithmic bound (up to a multiplicative constant), as it was shown that: ∀θ ∈ Θ, ∀n ≥ 3, Eθ [Tk (n)] ≤ 12 K log n log n log n ≤ 12K . and Eθ Rn ≤ 12 ∑ 2 ∆ ∆k k=1 ∆k Audibert et al. (2009) studied some variants of UCB1 policy. Among them, one consists in changing the 2 logt in the exploration term into ρ logt, where ρ > 0. This can be interpreted as a way to tune exploration: the smaller ρ is, the better the policy will perform in simple environments where information is disclosed easily (for example when all reward distributions are Dirac measures). On the contrary, ρ has to be greater to face more challenging environments (typically when reward distributions are Bernoulli laws with close parameters). This policy, that we denote UCB(ρ), was proven by Audibert et al. to achieve the logarithmic bound when ρ > 1, and the optimality was also obtained when ρ > 1 for a variant of UCB(ρ). 2 Bubeck (2010) showed in his PhD dissertation that their ideas actually enable to prove optimality 1 of UCB(ρ) for ρ > 1 . Moreover, the case ρ = 2 corresponds to a conﬁdence level of 1 (because 2 t of Hoeffding’s inequality, as above), and several studies (Lai and Robbins, 1985; Agrawal, 1995; Burnetas and Katehakis, 1996; Audibert et al., 2009; Honda and Takemura, 2010) have shown that this level is critical. We complete these works by a precise study of UCB(ρ) when ρ ≤ 1 . We prove that UCB(ρ) 2 is (1 − 2ρ)-consistent and that it is not α-consistent for any α < 1 − 2ρ (in view of the deﬁnition above, this means that the expected regret is roughly of order n1−2ρ ). Thus it does not achieve the logarithmic bound, but it performs well in simple environments, for example, environments where all reward distributions are Dirac measures. Moreover, we exhibit expected regret bounds of general UCB policies, with the 2 logt in the exploration term of UCB1 replaced by an arbitrary function. We give sufﬁcient conditions for such policies to be Hannan consistent and, as mentioned before, ﬁnd that lower bounds need not be logarithmic any more. 1.5 Outline The paper is organized as follows: in Section 2, we give bounds on the expected regret of general 1 UCB policies and of UCB (ρ) (ρ ≤ 2 ), as preliminary results. In Section 3, we focus on α-consistent policies. Then, in Section 4, we study the problem of selectivity, and we conclude in Section 5 by general results on the existence of logarithmic lower bounds. Throughout the paper ⌈x⌉ denotes the smallest integer not less than x whereas ⌊x⌋ denotes the largest integer not greater than x, ½A stands for the indicator function of event A, Ber(p) is the Bernoulli law with parameter p, and δx is the Dirac measure centred on x. 2. Preliminary Results In this section, we estimate the expected regret of the paper. UCB 191 policies. This will be useful for the rest of S ALOMON , AUDIBERT AND E L A LAOUI 2.1 Bounds on the Expected Regret of General UCB Policies We ﬁrst study general UCB policies, deﬁned by: • Draw each arm once, • then, at each round t, draw an arm It ∈ argmax Bk,Tk (t−1),t , k∈{1,...,K} ˆ where Bk,s,t is deﬁned by Bk,s,t = Xk,s + creasing. fk (t) s and where functions fk (1 ≤ k ≤ K) are in- This deﬁnition is inspired by popular UCB1 algorithm, for which fk (t) = 2 logt for all k. The following lemma estimates the performances of UCB policies in simple environments, for which reward distributions are Dirac measures. Lemma 1 Let 0 ≤ b < a ≤ 1 and n ≥ 1. For θ = (δa , δb ), the random variable T2 (n) is uniformly 1 upper bounded by ∆2 f2 (n) + 1. Consequently, the expected regret of UCB is upper bounded by 1 ∆ f 2 (n) + 1. Proof In environment θ, best arm is arm 1 and ∆ = ∆2 = a − b. Let us ﬁrst prove the upper bound of the sampling time. The assertion is true for n = 1 and n = 2: the ﬁrst two rounds consists in 1 drawing each arm once, so that T2 (n) ≤ 1 ≤ ∆2 f2 (n) + 1 for n ∈ {1, 2}. If, by contradiction, the as1 1 sertion is false, then there exists t ≥ 3 such that T2 (t) > ∆2 f2 (t) + 1 and T2 (t − 1) ≤ ∆2 f2 (t − 1) + 1. Since f2 (t) ≥ f2 (t − 1), this leads to T2 (t) > T2 (t − 1), meaning that arm 2 is drawn at round t. Therefore, we have a + f1 (t) T1 (t−1) ≤ b+ f2 (t) T2 (t−1) , hence a − b = ∆ ≤ f2 (t) T2 (t−1) , which implies 1 1 T2 (t − 1) ≤ ∆2 f2 (t) and thus T2 (t) ≤ ∆2 f2 (t) + 1. This contradicts the deﬁnition of t, and this ends the proof of the ﬁrst statement. The second statement is a direct consequence of Formula (1). Remark: throughout the paper, we will often use environments with K = 2 arms to provide bounds on expected regrets. However, we do not lose generality by doing so, because all corresponding proofs can be written almost identically to suit to any K ≥ 2, by simply assuming that the distribution of each arm k ≥ 3 is δ0 . We now give an upper bound of the expected sampling time of any arm such that ∆k > 0. This bound is valid in any environment, and not only those of the form (δa , δb ). Lemma 2 For any θ ∈ Θ and any β ∈ (0, 1), if ∆k > 0 the following upper bound holds: n Eθ [Tk (n)] ≤ u + where u = 4 fk (n) ∆2 k ∑ t=u+1 1+ logt 1 log( β ) . 192 e−2β fk (t) + e−2β fk∗ (t) , L OWER B OUNDS , W EAK C ONSISTENCY AND S ELECTIVITY IN BANDIT P ROBLEMS An upper bound of the expected regret can be deduced from this lemma thanks to Formula 1. Proof The core of the proof is a peeling argument and the use of Hoeffding’s maximal inequality (see, e.g., Cesa-Bianchi and Lugosi, 2006, section A.1.3 for details). The idea is originally taken from Audibert et al. (2009), and the following is an adaptation of the proof of an upper bound of UCB (ρ) in the case ρ > 1 which can be found in S. Bubeck’s PhD dissertation. 2 First, let us notice that the policy selects an arm k such that ∆k > 0 at time step t ≤ n only if at least one of the three following equations holds: Bk∗ ,Tk∗ (t−1),t ≤ µ∗ , (3) fk (t) , Tk (t − 1) (4) ˆ Xk,t ≥ µk + Tk (t − 1) < 4 fk (n) . ∆2 k (5) Indeed, if none of the equations is true, then: fk (n) ˆ > Xk,t + Tk (t − 1) Bk∗ ,Tk∗ (t−1),t > µ∗ = µk + ∆k ≥ µk + 2 fk (t) = Bk,Tk (t−1),t , Tk (t − 1) which implies that arm k can not be chosen at time step t. We denote respectively by ξ1,t , ξ2,t and ξ3,t the events corresponding to Equations (3), (4) and (5). We have: n ∑ ½I =k Eθ [Tk (n)] = Eθ t n n ∑ ½{I =k}∩ξ = Eθ t t=1 + Eθ 3,t ∑ ½{I =k}\ξ t 3,t . t=1 t=1 n Let us show that the sum ∑t=1 ½{It =k}∩ξ3,t is almost surely lower than u := ⌈4 fk (n)/∆2 ⌉. We assume k m−1 n by contradiction that ∑t=1 ½{It =k}∩ξ3,t > u. Then there exists m < n such that ∑t=1 ½{It =k}∩ξ3,t < m 4 fk (n)/∆2 and ∑t=1 ½{It =k}∩ξ3,t = ⌈4 fk (n)/∆2 ⌉. Therefore, for any s > m, we have: k k m m t=1 t=1 Tk (s − 1) ≥ Tk (m) = ∑ ½{It =k} ≥ ∑ ½{It =k}∩ξ3,t = 4 fk (n) 4 fk (n) ≥ , 2 ∆k ∆2 k so that ½{Is =k}∩ξ3,s = 0. But then m n ∑ ½{I =k}∩ξ t t=1 3,t = ∑ ½{It =k}∩ξ3,t = t=1 4 fk (n) ≤ u, ∆2 k which is the contradiction expected. n n We also have ∑t=1 ½{It =k}\ξ3,t = ∑t=u+1 ½{It =k}\ξ3,t : since Tk (t − 1) ≤ t − 1, event ξ3,t always happens at time step t ∈ {1, . . . , u}. And then, since event {It = k} is included in ξ1,t ∪ ξ2,t ∪ ξ3,t : n Eθ ∑ ½{It =k}\ξ3,t ≤ Eθ t=u+1 n n t=u+1 t=u+1 ∑ ½ξ1,t ∪ξ2,t ≤ ∑ Pθ (ξ1,t ) + Pθ (ξ2,t ). 193 S ALOMON , AUDIBERT AND E L A LAOUI It remains to ﬁnd upper bounds of Pθ (ξ1,t ) and Pθ (ξ2,t ). To this aim, we apply the peeling argument with a geometric grid over the time interval [1,t]: fk∗ (t) ≤ µ∗ Tk∗ (t − 1) ˆ Pθ (ξ1,t ) = Pθ Bk∗ ,Tk∗ (t−1),t ≤ µ∗ = Pθ Xk∗ ,Tk∗ (t−1) + ˆ ≤ Pθ ∃s ∈ {1, · · · ,t}, Xk∗ ,s + fk∗ (t) ≤ µ∗ s logt log(1/β) ≤ ∑ j=0 ˆ Pθ ∃s : {β j+1t < s ≤ β j t}, Xk∗ ,s + logt log(1/β) ≤ ∑ j=0 s Pθ ∃s : {β j+1t < s ≤ β j t}, logt log(1/β) ≤ ∑ j=0 ∑ j=0 ∑ (Xk ,l − µ∗ ) ≤ − ∗ s fk∗ (t) β j+1t fk∗ (t) l=1 ∑ (µ∗ − Xk ,l ) ≥ t < s ≤ β j t}, logt log(1/β) = fk∗ (t) ≤ µ∗ s j ∗ β j+1t fk∗ (t) l=1 s Pθ max ∑ (µ∗ − Xk∗ ,l ) ≥ s≤β j t l=1 β j+1t fk∗ (t) . As the range of the random variables (Xk∗ ,l )1≤l≤s is [0, 1], Hoeffding’s maximal inequality gives:   2 logt log(1/β) β j+1t fk∗ (t)  2 logt  Pθ (ξ1,t ) ≤ + 1 e−2β fk∗ (t) . ≤ ∑ exp − jt β log(1/β) j=0 Similarly, we have: logt + 1 e−2β fk (t) , log(1/β) and the result follows from the combination of previous inequalities. Pθ (ξ2,t ) ≤ 2.2 Bounds on the Expected Regret of UCB(ρ), ρ ≤ We study the performances of UCB (ρ) 1 2 1 policy, with ρ ∈ (0, 2 ]. We recall that ρ logt s . UCB (ρ) is the UCB ˆ policy deﬁned by fk (t) = ρ log(t) for all k, that is, Bk,s,t = Xk,s + Small values of ρ can be interpreted as a low level of experimentation in the balance between exploration and exploitation. 1 Precise regret bound orders of UCB(ρ) when ρ ∈ (0, 2 ] are not documented in the literature. We ﬁrst give an upper bound of expected regret in simple environments, where it is supposed to perform well. As stated in the following proposition (which is a direct consequence of Lemma 1), the order of the bound is ρ log n . ∆ 194 L OWER B OUNDS , W EAK C ONSISTENCY AND S ELECTIVITY IN BANDIT P ROBLEMS Lemma 3 Let 0 ≤ b < a ≤ 1 and n ≥ 1. For θ = (δa , δb ), the random variable T2 (n) is uniformly ρ upper bounded by ∆2 log(n) + 1. Consequently, the expected regret of UCB(ρ) is upper bounded by ρ ∆ log(n) + 1. One can show that the expected regret of UCB(ρ) is actually equivalent to ρ log n as n goes to ∆ inﬁnity. These good performances are compensated by poor results in more complex environments, as showed in the following theorem. We exhibit an expected regret upper bound which is valid for any θ ∈ Θ, and which is roughly of order n1−2ρ . We also show that this upper bound is asymptot1 ically optimal. Thus, with ρ ∈ (0, 2 ), UCB(ρ) does not perform enough exploration to achieve the logarithmic bound, as opposed to UCB(ρ) with ρ ∈ ( 1 , +∞). 2 1 Theorem 4 For any ρ ∈ (0, 2 ], any θ ∈ Θ and any β ∈ (0, 1), one has Eθ [Rn ] ≤ 4ρ log n ∑ ∆k + ∆k + 2∆k k:∆k >0 log n n1−2ρβ +1 . log(1/β) 1 − 2ρβ Moreover, if Θ has the Dirac/Bernoulli property, then for any ε > 0 there exists θ ∈ Θ such that Eθ [Rn ] lim n→+∞ n1−2ρ−ε = +∞. 1 1 The value ρ = 2 is critical, but we can deduce from the upper bound of this theorem that UCB( 2 ) is consistent in the classical sense of Lai and Robbins (1985) and of Burnetas and Katehakis (1996). log Proof We set u = 4ρ∆2 n . By Lemma 2 we get: k n Eθ [Tk (n)] ≤ u + 2 = u+2 ∑ logt + 1 e−2βρ log(t) log(1/β) ∑ logt 1 + 1 2ρβ log(1/β) t t=u+1 n t=u+1 n 1 ≤ u+2 log n +1 log(1/β) ≤ u+2 log n +1 log(1/β) 1+ ∑ ≤ u+2 log n +1 log(1/β) 1+ ≤ u+2 log n +1 . log(1/β) 1 − 2ρβ ∑ t 2ρβ t=1 n 1 t 2ρβ t=2 n−1 1 1−2ρβ n 1 t 2ρβ dt As usual, the upper bound of the expected regret follows from Formula (1). Now, let us show the lower bound. The result is obtained by considering an environment θ of the √ 1 form Ber( 1 ), δ 1 −∆ , where ∆ lies in (0, 2 ) and is such that 2ρ(1 + ∆)2 < 2ρ + ε. This notation is 2 2 obviously consistent with the deﬁnition of ∆ as an optimality gap. We set Tn := ⌈ ρ log n ⌉, and deﬁne ∆ the event ξn by: 1 1 ˆ ξn = X1,Tn < − (1 + √ )∆ . 2 ∆ 195 S ALOMON , AUDIBERT AND E L A LAOUI When event ξn occurs, one has for any t ∈ {Tn , . . . , n} ˆ X1,Tn + ρ logt Tn ˆ ≤ X1,Tn + ≤ √ ρ log n 1 1 < − (1 + √ )∆ + ∆ Tn 2 ∆ 1 − ∆, 2 so that arm 1 is chosen no more than Tn times by UCB(ρ) policy. Consequently: Eθ [T2 (n)] ≥ Pθ (ξn )(n − Tn ). We will now ﬁnd a lower bound of the probability of ξn thanks to Berry-Esseen inequality. We denote by C the corresponding constant, and by Φ the c.d.f. of the standard normal distribution. For convenience, we also deﬁne the following quantities: σ := E X1,1 − Using the fact that Φ(−x) = e− √ 2 β(x) 2πx 1 2 2 1 = , M3 := E 2 X1,1 − 1 2 3 1 = . 8 x2 with β(x) − − → 1, we have: −− x→+∞ ˆ √ X1,Tn − 1 √ 1 2 Tn ≤ −2 1 + √ ∆ Tn σ ∆ √ √ CM3 Φ −2(∆ + ∆) Tn − 3 √ σ Tn √ 2 exp −2(∆ + ∆) Tn √ √ CM3 √ √ √ β 2(∆ + ∆) Tn − 3 √ σ Tn 2 2π(∆ + ∆) Tn √ 2 ρ log n exp −2(∆ + ∆) ( ∆ + 1) √ √ CM3 √ √ √ β 2(∆ + ∆) Tn − 3 √ σ Tn 2 2π(∆ + ∆) Tn √ √ −2ρ(1+ ∆)2 exp −2(∆ + ∆)2 √ √ CM3 n √ √ √ β 2(∆ + ∆) Tn − 3 √ . Tn σ Tn 2 2π(∆ + ∆) Pθ (ξn ) = Pθ ≥ ≥ ≥ ≥ Previous calculations and Formula (1) give Eθ [Rn ] = ∆Eθ [T2 (n)] ≥ ∆Pθ (ξn )(n − Tn ), √ 1−2ρ(1+ ∆)2 so that we ﬁnally obtain a lower bound of Eθ [Rn ] of order n √log n . Therefore, nEθ [Rn ] is at least 1−2ρ−ε √ 2 √ 2 n2ρ+ε−2ρ(1+ ∆) √ of order . Since 2ρ + ε − 2ρ(1 + ∆) > 0, the numerator goes to inﬁnity, faster than log n √ log n. This concludes the proof. 196 L OWER B OUNDS , W EAK C ONSISTENCY AND S ELECTIVITY IN BANDIT P ROBLEMS 3. Bounds on the Class α-consistent Policies In this section, our aim is to ﬁnd how the classical results of Lai and Robbins (1985) and of Burnetas and Katehakis (1996) can be generalised if we do not restrict the study to consistent policies. As a by-product, we will adapt their results to the present setting, which is simpler than their parametric frameworks. We recall that a policy is consistent if its expected regret is o(na ) for all a > 0 in all environments θ ∈ Θ. A natural way to relax this deﬁnition is the following. Deﬁnition 5 A policy is α-consistent if ∀a > α, ∀θ ∈ Θ, Eθ [Rn ] = o(na ). For example, we showed in the previous section that, for any ρ ∈ (0, 1 ], UCB(ρ) is (1−2ρ)-consistent 2 and not α-consistent if α < 1 − 2ρ. Note that the relevant range of α in this deﬁnition is [0, 1): the case α = 0 corresponds to the standard deﬁnition of consistency (so that throughout the paper the term ”consistent” also means ”0-consistent”), and any value α ≥ 1 is pointless as any policy is then α-consistent. Indeed, the expected regret of any policy is at most of order n. This also lead us to wonder what happens if we only require the expected regret to be o(n): ∀θ ∈ Θ, Eθ [Rn ] = o(n). This requirement corresponds to the deﬁnition of Hannan consistency. The class of Hannan consistent policies includes consistent policies and α-consistent policies for any α ∈ [0, 1). Some results about this class will be obtained in Section 5. We focus on regret lower bounds on α-consistent policies. We ﬁrst show that the main result of Burnetas and Katehakis can be extended in the following way. Theorem 6 Assume that Θ has the product property. Fix an α-consistent policy and θ ∈ Θ. If ∆k > 0 and if 0 < Dk (θ) < ∞, then ∀ε > 0, lim Pθ Tk (n) ≥ (1 − ε) n→+∞ (1 − α) log n = 1. Dk (θ) Consequently lim inf n→+∞ 1−α Eθ [Tk (n)] ≥ . log n Dk (θ) Remind that the lower bound of the expected regret is then deduced from Formula (1), and that coefﬁcient Dk (θ) is deﬁned by: Dk (θ) := inf ˜ νk ∈Θk :E[˜ k ]>µ∗ ν ˜ KL(νk , νk ), where KL(ν, µ) denotes the Kullback-Leibler divergence of measures ν and µ. Note that, as opposed to Burnetas and Katehakis (1996), there is no optimal policy in general (i.e., a policy that would achieve the lower bound in all environment θ). This can be explained intuitively as follows. If by contradiction there existed such a policy, its expected regret would be of order log n and consequently it would be (0-)consistent. Then the lower bounds in the case of 197 S ALOMON , AUDIBERT AND E L A LAOUI 1−α 0-consistency would necessarily hold. This can not happen if α > 0 because Dk (θ) < Dk1 . (θ) Nevertheless, this argument is not rigorous because the fact that the regret would be of order log n is only valid for environments θ such that 0 < Dk (θ) < ∞. The non-existence of optimal policies is implied by a stronger result of the next section (yet, only if α > 0.2). Proof We adapt Proposition 1 in Burnetas and Katehakis (1996) and its proof. Let us denote θ = (ν1 , . . . , νK ). We ﬁx ε > 0, and we want to show that: lim Pθ n→+∞ Set δ > 0 and δ′ > α such that ˜ that E[νk ] > µ∗ and 1−δ′ 1+δ Tk (n) (1 − ε)(1 − α) < log n Dk (θ) = 0. ˜ > (1 − ε)(1 − α). By deﬁnition of Dk (θ), there exists νk such ˜ Dk (θ) < KL(νk , νk ) < (1 + δ)Dk (θ). ˜ ˜ ˜ Let us set θ = (ν1 , . . . , νk−1 , νk , νk+1 , . . . , νK ). Environment θ lies in Θ by the product property and δ = KL(ν , ν ) and arm k is its best arm. Deﬁne I k ˜k ′ Aδ := n Tk (n) 1 − δ′ < δ log n I ′′ δ , Cn := log LTk (n) ≤ 1 − δ′′ log n , where δ′′ is such that α < δ′′ < δ′ and Lt is deﬁned by log Lt = ∑ts=1 log δ′ δ′ δ′′ δ′ dνk ˜ d νk (Xk,s ) . δ′′ Now, we show that Pθ (An ) = Pθ (An ∩Cn ) + Pθ (An \Cn ) − − → 0. −− n→+∞ On the one hand, one has: ′′ ′′ ′ ′′ ′ δ δ Pθ (Aδ ∩Cn ) ≤ n1−δ Pθ (Aδ ∩Cn ) ˜ n n ′′ ′ (6) ′′ ≤ n1−δ Pθ (Aδ ) = n1−δ Pθ n − Tk (n) > n − ˜ ˜ n 1 − δ′ Iδ log n ′′ ≤ n1−δ Eθ [n − Tk (n)] ˜ (7) ′ n − 1−δ log n Iδ ′′ = n−δ Eθ ∑K Tℓ (n) − Tk (n) ˜ l=1 ′ n − 1−δ Iδ log n n ′′ ≤ ∑ℓ=k n−δ Eθ [Tℓ (n)] ˜ ′ 1 − 1−δ Iδ log n n − − → 0. −− (8) n→+∞ ′ Equation (6) results from a partition of Aδ into events {Tk (n) = a}, 0 ≤ a < n ′′ 1−δ′ Iδ log n . Each event ′′ δ {Tk (n) = a} ∩ Cn equals {Tk (n) = a} ∩ ∏a dνk (Xk,s ) ≤ n1−δ and is measurable with respect s=1 d νk ˜ to Xk,1 , . . . , Xk,a and to Xℓ,1 , . . . , Xℓ,n (ℓ = k). Thus, ½{Tk (n)=a}∩Cn ′′ can be written as a function f of δ the latter r.v. and we have: ′′ δ Pθ {Tk (n) = a} ∩Cn = f (xk,s )1≤s≤a , (xℓ,s )ℓ=k,1≤s≤n ∏ ℓ=k 1≤s≤n ≤ f (xk,s )1≤s≤a , (xℓ,s )ℓ=k,1≤s≤n ∏ ℓ=k 1≤s≤n ′′ ′′ δ = n1−δ Pθ {Tk (n) = a} ∩Cn ˜ 198 . dνℓ (xℓ,s ) ∏ dνk (xk,s ) 1≤s≤a ′′ dνℓ (xℓ,s )n1−δ ∏ 1≤s≤a ˜ d νk (xk,s ) L OWER B OUNDS , W EAK C ONSISTENCY AND S ELECTIVITY IN BANDIT P ROBLEMS Equation (7) is a consequence of Markov’s inequality, and the limit in (8) is a consequence of α-consistency. ′ On the other hand, we set bn := 1−δ log n, so that Iδ ′ ′′ δ Pθ (Aδ \Cn ) ≤ P n ≤ P max log L j > (1 − δ′′ ) log n j≤⌊bn ⌋ 1 1 − δ′′ max log L j > I δ bn j≤⌊bn ⌋ 1 − δ′ . This term tends to zero, as a consequence of the law of large numbers. ′ Now that Pθ (Aδ ) tends to zero, the conclusion results from n 1 − δ′ 1 − δ′ (1 − ε)(1 − α) > ≥ . δ (1 + δ)Dk (θ) Dk (θ) I The previous lower bound is asymptotically optimal with respect to its dependence in α, as claimed in the following proposition. Proposition 7 Assume that Θ has the Dirac/Bernoulli property. There exist θ ∈ Θ and a constant c > 0 such that, for any α ∈ [0, 1), there exists an α-consistent policy such that: lim inf n→+∞ Eθ [Tk (n)] ≤ c, (1 − α) log n for any k satisfying ∆k > 0. Proof In any environment of the form θ = (δa , δb ) with a = b, Lemma 3 implies the following estimate for UCB(ρ): Eθ Tk (n) ρ lim inf ≤ 2, n→+∞ log n ∆ where k = k∗ . Because 1−α ∈ (0, 1 ) and since UCB(ρ) is (1 − 2ρ)-consistent for any ρ ∈ (0, 1 ] (Theorem 4), we 2 2 2 1 obtain the result by choosing the α-consistent policy UCB( 1−α ) and by setting c = 2∆2 . 2 4. Selectivity In this section, we address the problem of selectivity. By selectivity, we mean the ability to adapt to the environment as and when rewards are observed. More precisely, a set of two (or more) policies is given. The one that performs the best depends on environment θ. We wonder if there exists an adaptive procedure that, given any environment θ, would be as good as any policy in the given set. Two major reasons motivate this study. On the one hand this question was answered by Burnetas and Katehakis within the class of consistent policies. They exhibits an asymptotically optimal policy, that is, that achieves the regret 199 S ALOMON , AUDIBERT AND E L A LAOUI lower bounds they have proven. The fact that a policy performs as best as any other one obviously solves the problem of selectivity. On the other hand, this problem has already been studied in the context of adversarial bandit by Auer et al. (2003). Their setting differs from our not only because their bandits are nonstochastic, but also because their adaptive procedure takes only into account a given number of recommendations, whereas in our setting the adaptation is supposed to come from observing rewards of the chosen arms (only one per time step). Nevertheless, one can wonder if an ”exponentially weighted forecasters” procedure like E XP 4 could be transposed to our context. The answer is negative, as stated in the following theorem. To avoid confusion, we make the notations of the regret and of sampling time more precise by adding the considered policy: under policy A , Rn and Tk (n) will be respectively denoted Rn (A ) and Tk (n, A ). ˜ Theorem 8 Let A be a consistent policy and let ρ be a real number in (0, 0.4). If Θ has the ˜ Dirac/Bernoulli property and the product property, there is no policy which can both beat A and UCB (ρ), that is: ∀A , ∃θ ∈ Θ, lim sup n→+∞ Eθ [Rn (A )] > 1. ˜ min(Eθ [Rn (A )], Eθ [Rn (UCB(ρ))]) Thus the existence of optimal policies does not hold when we extend the notion of consistency. Precisely, as UCB(ρ) is (1 − 2ρ)-consistent, we have shown that there is no optimal policy within the class of α-consistent policies, with α > 0.2. Consequently, there do not exist optimal policies in the class of Hannan consistent policies either. Moreover, Theorem 8 shows that methods that would be inspired by related literature in adversarial bandit can not apply to our framework. As we said, this impossibility may come from the fact that we can not observe at each time step the decisions and rewards of more than one algorithm. If we were able to observe a given set of policies from step to step, then it would be easy to beat them all: it would be sufﬁcient to aggregate all the observations and simply pull the arm with the greater empiric mean. The case where we only observe decisions (and not rewards) of a set of policies may be interesting, but is left outside of the scope of this paper. Proof Assume by contradiction that ∃A , ∀θ ∈ Θ, lim sup un,θ ≤ 1, n→+∞ [Rn where un,θ = min(E [R (Eθ)],E(A )](UCB(ρ))]) . ˜ θ n A θ [Rn For any θ, we have Eθ [Rn (A )] = Eθ [Rn (A )] ˜ ˜ Eθ [Rn (A )] ≤ un,θ Eθ [Rn (A )], ˜ Eθ [Rn (A )] (9) ˜ so that the fact that A is a consistent policy implies that A is also consistent. Consequently the lower bound of Theorem 6 also holds for policy A . For the rest of the proof, we focus on environments of the form θ = (δ0 , δ∆ ) with ∆ > 0. In this case, arm 2 is the best arm, so that we have to compute D1 (θ). On the one hand, we have: D1 (θ) = inf ˜ ν1 ∈Θ1 :E[˜ 1 ν ]>µ∗ ˜ KL(ν1 , ν1 ) = inf ˜ ν1 ∈Θ1 :E[˜ 1 ]>∆ ν 200 ˜ KL(δ0 , ν1 ) = inf ˜ ν1 ∈Θ1 :E[˜ 1 ]>∆ ν log 1 . ˜ ν1 (0) L OWER B OUNDS , W EAK C ONSISTENCY AND S ELECTIVITY IN BANDIT P ROBLEMS ˜ ˜ As E[ν1 ] ≤ 1 − ν1 (0), we get: D1 (θ) ≥ inf ˜ ν1 ∈Θ1 :1−˜ 1 (0)≥∆ ν log 1 ˜ ν1 (0) ≥ log 1 . 1−∆ One the other hand, we have for any ε > 0: D1 (θ) ≤ KL(δ0 , Ber(∆ + ε)) = log Consequently D1 (θ) = log 1 1−∆ 1 1−∆−ε , and the lower bound of Theorem 6 reads: lim inf n→+∞ 1 Eθ [T1 (n, A )] . ≥ 1 log n log 1−∆ Just like Equation (9), we have: Eθ [Rn (A )] ≤ un,θ Eθ [Rn (UCB(ρ))]. Moreover, Lemma 3 provides: Eθ [Rn (UCB(ρ))] ≤ 1 + ρ log n . ∆ Now, by gathering the three previous inequalities and Formula (1), we get: 1 log 1 1−∆ ≤ lim inf n→+∞ Eθ [T1 (n, A )] Eθ [Rn (A )] = lim inf n→+∞ log n ∆ log n un,θ Eθ [Rn (UCB(ρ))] un,θ ρ log n 1+ ≤ lim inf n→+∞ ∆ log n ∆ log n ∆ ρun,θ un,θ ρ ρ + lim inf 2 = 2 lim inf un,θ ≤ 2 lim sup un,θ ≤ lim inf n→+∞ ∆ n→+∞ ∆ log n ∆ n→+∞ ∆ n→+∞ ρ . ≤ ∆2 ≤ lim inf n→+∞ This means that ρ has to be lower bounded by ∆2 , 1 log( 1−∆ ) but this is greater than 0.4 if ∆ = 0.75, hence the contradiction. Note that this proof gives a simple alternative to Theorem 4 to show that UCB(ρ) is not consistent (if ρ ≤ 0.4). Indeed if it were consistent, then in environment θ = (δ0 , δ∆ ) the same contradiction between the lower bound of Theorem 6 and the upper bound of Lemma 3 would hold. 5. General Bounds In this section, we study lower bounds on the expected regret with few requirements on Θ and on the class of policies. With a simple property on Θ but without any assumption on the policy, we show that there always exist logarithmic lower bounds for some environments θ. Then, still with a 201 S ALOMON , AUDIBERT AND E L A LAOUI simple property on Θ, we show that there exists a Hannan consistent policy for which the expected regret is sub-logarithmic for some environment θ. Note that the policy that always pulls arm 1 has a 0 expected regret in environments where arm 1 has the best mean reward, and an expected regret of order n in other environments. So, for this policy, expected regret is sub-logarithmic in some environments. Nevertheless, this policy is not Hannan consistent because its expected regret is not always o(n). 5.1 The Necessity of a Logarithmic Regret in Some Environments The necessity of a logarithmic regret in some environments can be explained by a simple sketch proof. Assume that the agent knows the number of rounds n, and that he balances exploration and exploitation in the following way: he ﬁrst pulls each arm s(n) times, and then selects the arm that has obtained the best empiric mean for the rest of the game. Denote by ps(n) the probability that the best arm does not have the best empiric mean after the exploration phase (i.e., after the ﬁrst Ks(n) rounds). The expected regret is then of the form c1 (1 − ps(n) )s(n) + c2 ps(n) n. (10) Indeed, if the agent manages to match the best arm then he only suffers the pulls of suboptimal arms during the exploration phase. That represents an expected regret of order s(n). If not, the number of pulls of suboptimal arms is of order n, and so is the expected regret. Now, let us approximate ps(n) . It has the same order as the probability that the best arm gets X ∗ −µ∗ an empiric mean lower than the second best mean reward. Moreover, k ,s(n) s(n) (where σ is σ ∗ ,1 ) has approximately a standard normal distribution by the central limit theorem. the variance of Xk Therefore, we have: ps(n) ≈ Pθ (Xk∗ ,s(n) ≤ µ∗ − ∆) = Pθ ≈ ≈  σ 1 1 √ exp − 2 2π ∆ s(n) Xk∗ ,s(n) − µ∗ σ  2 ∆ s(n)  σ s(n) ≤ − ∆ s(n) σ 1 σ ∆2 s(n) √ . exp − 2σ2 2π ∆ s(n) It follows that the expected regret has to be at least logarithmic. Indeed, to ensure that the second term c2 ps(n) n of Equation (10) is sub-logarithmic, s(n) has to be greater than log n. But then ﬁrst term c1 (1 − ps(n) )s(n) is greater than log n. Actually, the necessity of a logarithmic regret can be written as a consequence of Theorem 6. n Indeed, if we assume by contradiction that lim supn→+∞ Eθ Rn = 0 for all θ (i.e., Eθ Rn = o(log n)), log the considered policy is consistent. Consequently, Theorem 6 implies that lim sup n→+∞ E θ Rn E θ Rn ≥ lim inf > 0. n→+∞ log n log n Yet, this reasoning needs Θ having the product property, and conditions of the form 0 < Dk (θ) < ∞ also have to hold. The following proposition is a rigorous version of our sketch, and it shows that the necessity of a logarithmic lower bound can be based on a simple property on Θ. 202 L OWER B OUNDS , W EAK C ONSISTENCY AND S ELECTIVITY IN BANDIT P ROBLEMS ˜ ˜ ˜ Proposition 9 Assume that there exist two environments θ = (ν1 , . . . , νK ) ∈ Θ, θ = (ν1 , . . . , νK ) ∈ Θ, and an arm k ∈ {1, . . . , K} such that 1. k has the best mean reward in environment θ, ˜ 2. k is not the winning arm in environment θ, ˜ 3. νk = νk and there exists η ∈ (0, 1) such that dνℓ ∏ d νℓ (Xℓ,1 ) ≥ η ˜ ℓ=k Pθ − a.s. ˜ (11) ˆ Then, for any policy, there exists θ ∈ Θ such that lim sup n→+∞ E θ Rn ˆ > 0. log n ˜ Let us explain the logic of the three conditions of the proposition. If νk = νk , and in case νk seems to be the reward distribution of arm k, then arm k has to be pulled often enough for the regret to be small if the environment is θ. Nevertheless, one has to explore other arms to know ˜ whether the environment is actually θ. Moreover, Inequality (11) makes sure that the distinction ˜ is tough to make: it ensures that pulling any arm ℓ = k gives a reward which is between θ and θ likely in both environments. Without such an assumption the problem may be very simple, and providing a logarithmic lower bound is hopeless. Indeed, the distinction between any pair of tricky ˜ environments (θ, θ) may be solved in only one pull of a given arm ℓ = k, that would almost surely give a reward that is possible in only one of the two environments. The third condition can be seen as an alternate version of condition 0 < Dk (θ) < ∞ in Theorem 6, though there is no logical connection with it. Finally, let us remark that one can check that any set Θ that has the Dirac/Bernoulli property satisﬁes the conditions of Proposition 9. Proof The proof consists in writing a proper version of Expression (10). To this aim we compute a lower bound of Eθ Rn , expressed as a function of Eθ Rn and of an arbitrary function g(n). ˜ ˜ ˜ In the following, ∆k denotes the optimality gap of arm k in environment θ. As event ∑ℓ=k Tℓ (n) ≤ g(n) is measurable with respect to Xℓ,1 , . . . , Xℓ,⌊g(n)⌋ (ℓ = k) and to Xk,1 , . . . , Xk,n , we also introduce the function q such that ½{∑ℓ=k Tℓ (n)≤g(n)} = q (Xℓ,s )ℓ=k, s=1..⌊g(n)⌋ , (Xk,s )s=1..n . 203 S ALOMON , AUDIBERT AND E L A LAOUI We have: ˜ ˜ ˜ Eθ Rn ≥ ∆k Eθ [Tk (n)] ≥ ∆k (n − g(n))Pθ (Tk (n) ≥ n − g(n)) ˜ ˜ (12) ˜ = ∆k (n − g(n))Pθ n − ∑ Tℓ (n) ≥ n − g(n) ˜ ℓ=k ˜ = ∆k (n − g(n))Pθ ˜ ˜ = ∆k (n − g(n)) ∑ Tℓ (n) ≤ g(n) ℓ=k q (xℓ,s )ℓ=k, s=1..⌊g(n)⌋ , (xk,s )s=1..n ˜ ˜ ∏ d νℓ (xℓ,s )∏d νk (xk,s ) ℓ=k s = 1..⌊g(n)⌋ s=1..n ˜ ≥ ∆k (n − g(n)) q (xℓ,s )ℓ=k, s=1..⌊g(n)⌋ , (xk,s )s=1..n η⌊g(n)⌋∏ dνℓ (xℓ,s )∏dνk (xk,s ) ℓ=k s = 1..⌊g(n)⌋ ˜ ≥ ∆k (n − g(n))ηg(n) q (xℓ,s )ℓ=k, s=1..⌊g(n)⌋ , (xk,s )s=1..n ∏ dνℓ (xℓ,s )∏dνk (xk,s ) ℓ=k s = 1..⌊g(n)⌋ ˜ = ∆k (n − g(n))ηg(n) Pθ (13) s=1..n s=1..n ∑ Tℓ (n) ≤ g(n) ℓ=k ˜ = ∆k (n − g(n))ηg(n) 1 − Pθ ∑ Tℓ (n) > g(n) ℓ=k ˜ ≥ ∆k (n − g(n))ηg(n) 1 − Eθ ∑ℓ=k Tℓ (n) g(n) (14) ˜ ≥ ∆k (n − g(n))ηg(n) 1 − Eθ ∑ℓ=k ∆ℓ Tℓ (n) ∆g(n) (15) E θ Rn ˜ ≥ ∆k (n − g(n))ηg(n) 1 − , ∆g(n) where the ﬁrst inequality of (12) is a consequence of Formula (1), the second inequality of (12) and inequality (14) come from Markov’s inequality, Inequality (13) is a consequence of (11), and Inequality (15) results from the fact that ∆ℓ ≥ ∆ for all ℓ. n θ −− Now, let us conclude. If Eθ Rn − − → 0, we set g(n) = 2E∆Rn , so that log n→+∞ g(n) ≤ min n − log n 2 , 2 log η for n large enough. Then, we have: √ − log n ˜ k n − g(n) ηg(n) ≥ ∆k n η 2 log η = ∆k n . ˜ ˜ E θ Rn ≥ ∆ ˜ 2 4 4 In particular, Eθ Rn ˜ −− log n − − → n→+∞ +∞, and the result follows. 204 L OWER B OUNDS , W EAK C ONSISTENCY AND S ELECTIVITY IN BANDIT P ROBLEMS 5.2 Hannan Consistency We will prove that there exists a Hannan consistent policy such that there can not be a logarithmic lower bound for every environment θ of Θ. To this aim, we make use of general UCB policies again (cf. Section 2.1). Let us ﬁrst give sufﬁcient conditions on the fk for UCB policy to be Hannan consistent. Proposition 10 Assume that fk (n) = o(n) for all k ∈ {1, . . . , K}. Assume also that there exist γ > 1 2 and N ≥ 3 such that fk (n) ≥ γ log log n for all k ∈ {1, . . . , K} and for all n ≥ N. Then UCB is Hannan consistent. Proof Fix an arm k such that ∆k > 0 and choose β ∈ (0, 1) such that 2βγ > 1. By means of Lemma 2, we have for n large enough: n Eθ [Tk (n)] ≤ u + 2 ∑ 1+ t=u+1 logt 1 log( β ) e−2βγ log logt , k where u = 4 f∆(n) . 2 k Consequently, we have: n Eθ [Tk (n)] ≤ u + 2 ∑ t=2 1 1 1 + 1 (logt)2βγ−1 2βγ (logt) log( β ) . (16) n n 1 Sums of the form ∑t=2 (logt)c with c > 0 are equivalent to (log n)c as n goes to inﬁnity. Indeed, on the one hand we have n n n 1 dx 1 ∑ (logt)c ≤ 2 (log x)c ≤ ∑ (logt)c , t=2 t=3 n 1 so that ∑t=2 (logt)c ∼ n dx 2 (log x)c . n 2 On the other hand, we have n x dx = c (log x) (log x)c n dx 2 (log x)c+1 n 1 n ∑t=2 (logt)c ∼ (log n)c n +c 2 2 dx . (log x)c+1 n dx 2 (log x)c n dx 2 (log x)c n (log n)c . As both integrals are divergent we have =o Combining the fact that constant C > 0 such that with Equation (16), we get the existence of a Eθ [Tk (n)] ≤ , so that ∼ Cn 4 fk (n) + . 2 ∆ (log n)2βγ−1 Since fk (n) = o(n) and 2βγ − 1 > 0, the latter inequality shows that Eθ [Tk (n)] = o(n). The result follows. We are now in the position to prove the main result of this section. Theorem 11 If Θ has the Dirac/Bernoulli property, there exist Hannan consistent policies for which the expected regret can not be lower bounded by a logarithmic function in all environments θ. 205 S ALOMON , AUDIBERT AND E L A LAOUI Proof If f1 (n) = f2 (n) = log log n for n ≥ 3, UCB is Hannan consistent by Proposition 10. According to Lemma 1, the expected regret is then of order log log n in environments of the form (δa , δb ), a = b. Hence the conclusion on the non-existence of logarithmic lower bounds. Thus we have obtained a lower bound of order log log n. This order is critical regarding the methods we used. Yet, we do not know if this order is optimal. Acknowledgments This work has been supported by the French National Research Agency (ANR) through the COSINUS program (ANR-08-COSI-004: EXPLO-RA project). References R. Agrawal. Sample mean based index policies with o(log n) regret for the multi-armed bandit problem. Advances in Applied Mathematics, 27:1054–1078, 1995. H. Akaike. Information theory and an extension of the maximum likelihood principle. In Second International Symposium on Information Theory, volume 1, pages 267–281. Springer Verlag, 1973. J.-Y. Audibert, R. Munos, and C. Szepesv´ ri. Exploration-exploitation tradeoff using variance estia mates in multi-armed bandits. Theoretical Computer Science, 410(19):1876–1902, 2009. P. Auer, N. Cesa-Bianchi, and P. Fischer. Finite-time analysis of the multiarmed bandit problem. Machine Learning, 47(2-3):235–256, 2002. P. Auer, N. Cesa-Bianchi, Y. Freund, and R.E. Schapire. The nonstochastic multiarmed bandit problem. SIAM Journal on Computing, 32(1):48–77, 2003. D. Bergemann and J. Valimaki. Bandit problems. In The New Palgrave Dictionary of Economics, 2nd ed. Macmillan Press, 2008. S. Bubeck. Bandits Games and Clustering Foundations. PhD thesis, Universit´ Lille 1, France, e 2010. S. Bubeck, R. Munos, G. Stoltz, and C. Szepesvari. Online optimization in X-armed bandits. In Advances in Neural Information Processing Systems 21, pages 201–208. 2009. A.N. Burnetas and M.N. Katehakis. Optimal adaptive policies for sequential allocation problems. Advances in Applied Mathematics, 17(2):122–142, 1996. N. Cesa-Bianchi and G. Lugosi. Prediction, Learning, and Games. Cambridge University Press, New York, NY, 2006. N. Cesa-Bianchi, Y. Freund, D. Haussler, D.P. Helmbold, R.E. Schapire, and M.K. Warmuth. How to use expert advice. Journal of the ACM (JACM), 44(3):427–485, 1997. 206 L OWER B OUNDS , W EAK C ONSISTENCY AND S ELECTIVITY IN BANDIT P ROBLEMS P.A. Coquelin and R. Munos. Bandit algorithms for tree search. In Uncertainty in Artiﬁcial Intelligence, 2007. S. Gelly and Y. Wang. Exploration exploitation in go: UCT for Monte-Carlo go. In Online Trading between Exploration and Exploitation Workshop, Twentieth Annual Conference on Neural Information Processing Systems (NIPS 2006), 2006. J. Honda and A. Takemura. An asymptotically optimal bandit algorithm for bounded support models. In Proceedings of the Twenty-Third Annual Conference on Learning Theory (COLT), 2010. R. Kleinberg, A. Slivkins, and E. Upfal. Multi-armed bandits in metric spaces. In Proceedings of the 40th Annual ACM Symposium on Theory of Computing, pages 681–690, 2008. R. D. Kleinberg. Nearly tight bounds for the continuum-armed bandit problem. In Advances in Neural Information Processing Systems 17, pages 697–704. 2005. T. L. Lai and H. Robbins. Asymptotically efﬁcient adaptive allocation rules. Advances in Applied Mathematics, 6:4–22, 1985. D. Lamberton, G. Pag` s, and P. Tarr` s. When can the two-armed bandit algorithm be trusted? e e Annals of Applied Probability, 14(3):1424–1454, 2004. C.L. Mallows. Some comments on cp. Technometrics, pages 661–675, 1973. H. Robbins. Some aspects of the sequential design of experiments. Bulletin of the American Mathematics Society, 58:527–535, 1952. G. Schwarz. Estimating the dimension of a model. The Annals of Statistics, 6(2):461–464, 1978. W.R. Thompson. On the likelihood that one unknown probability exceeds another in view of the evidence of two samples. Biometrika, 25(3/4):285–294, 1933. 207

5 0.096695639 116 jmlr-2013-Truncated Power Method for Sparse Eigenvalue Problems

Author: Xiao-Tong Yuan, Tong Zhang

Abstract: This paper considers the sparse eigenvalue problem, which is to extract dominant (largest) sparse eigenvectors with at most k non-zero components. We propose a simple yet effective solution called truncated power method that can approximately solve the underlying nonconvex optimization problem. A strong sparse recovery result is proved for the truncated power method, and this theory is our key motivation for developing the new algorithm. The proposed method is tested on applications such as sparse principal component analysis and the densest k-subgraph problem. Extensive experiments on several synthetic and real-world data sets demonstrate the competitive empirical performance of our method. Keywords: sparse eigenvalue, power method, sparse principal component analysis, densest k-subgraph

6 0.09196604 102 jmlr-2013-Sparse Matrix Inversion with Scaled Lasso

7 0.088665485 8 jmlr-2013-A Theory of Multiclass Boosting

8 0.082067765 16 jmlr-2013-Bayesian Nonparametric Hidden Semi-Markov Models

9 0.076311491 81 jmlr-2013-Optimal Discovery with Probabilistic Expert Advice: Finite Time Analysis and Macroscopic Optimality

10 0.072134316 10 jmlr-2013-Algorithms and Hardness Results for Parallel Large Margin Learning

11 0.072056584 97 jmlr-2013-Risk Bounds of Learning Processes for Lévy Processes

12 0.068714142 104 jmlr-2013-Sparse Single-Index Model

13 0.057605281 76 jmlr-2013-Nonparametric Sparsity and Regularization

14 0.056291468 94 jmlr-2013-Ranked Bandits in Metric Spaces: Learning Diverse Rankings over Large Document Collections

15 0.05358588 77 jmlr-2013-On the Convergence of Maximum Variance Unfolding

16 0.05182638 57 jmlr-2013-Kernel Bayes' Rule: Bayesian Inference with Positive Definite Kernels

17 0.049156569 34 jmlr-2013-Distance Preserving Embeddings for General n-Dimensional Manifolds

18 0.046192206 25 jmlr-2013-Communication-Efficient Algorithms for Statistical Optimization

19 0.042903107 56 jmlr-2013-Keep It Simple And Sparse: Real-Time Action Recognition

20 0.042822484 79 jmlr-2013-On the Mutual Nearest Neighbors Estimate in Regression

similar papers computed by lsi model

lsi for this paper:

topicId topicWeight

[(0, -0.225), (1, 0.112), (2, 0.062), (3, 0.18), (4, -0.103), (5, -0.218), (6, 0.085), (7, -0.11), (8, 0.017), (9, 0.03), (10, 0.001), (11, 0.056), (12, -0.074), (13, -0.007), (14, 0.162), (15, 0.046), (16, 0.049), (17, -0.182), (18, 0.321), (19, -0.029), (20, -0.081), (21, 0.099), (22, -0.052), (23, -0.052), (24, -0.086), (25, -0.014), (26, -0.094), (27, -0.015), (28, -0.074), (29, 0.169), (30, 0.083), (31, 0.087), (32, 0.022), (33, -0.157), (34, 0.108), (35, 0.057), (36, -0.01), (37, -0.032), (38, -0.02), (39, 0.125), (40, 0.013), (41, 0.051), (42, -0.036), (43, 0.06), (44, -0.058), (45, -0.002), (46, -0.038), (47, 0.01), (48, -0.023), (49, 0.08)]

similar papers list:

simIndex simValue paperId paperTitle

same-paper 1 0.98222715 105 jmlr-2013-Sparsity Regret Bounds for Individual Sequences in Online Linear Regression

Author: Sébastien Gerchinovitz

2 0.64764726 116 jmlr-2013-Truncated Power Method for Sparse Eigenvalue Problems

Author: Xiao-Tong Yuan, Tong Zhang

3 0.51308739 35 jmlr-2013-Distribution-Dependent Sample Complexity of Large Margin Learning

Author: Sivan Sabato, Nathan Srebro, Naftali Tishby

4 0.42916983 57 jmlr-2013-Kernel Bayes' Rule: Bayesian Inference with Positive Definite Kernels

Author: Kenji Fukumizu, Le Song, Arthur Gretton

Abstract: A kernel method for realizing Bayes’ rule is proposed, based on representations of probabilities in reproducing kernel Hilbert spaces. Probabilities are uniquely characterized by the mean of the canonical map to the RKHS. The prior and conditional probabilities are expressed in terms of RKHS functions of an empirical sample: no explicit parametric model is needed for these quantities. The posterior is likewise an RKHS mean of a weighted sample. The estimator for the expectation of a function of the posterior is derived, and rates of consistency are shown. Some representative applications of the kernel Bayes’ rule are presented, including Bayesian computation without likelihood and ﬁltering with a nonparametric state-space model. Keywords: kernel method, Bayes’ rule, reproducing kernel Hilbert space

5 0.42078644 114 jmlr-2013-The Rate of Convergence of AdaBoost

Author: Indraneel Mukherjee, Cynthia Rudin, Robert E. Schapire

6 0.41655678 65 jmlr-2013-Lower Bounds and Selectivity of Weak-Consistent Policies in Stochastic Multi-Armed Bandit Problem

7 0.40808505 81 jmlr-2013-Optimal Discovery with Probabilistic Expert Advice: Finite Time Analysis and Macroscopic Optimality

8 0.40275171 16 jmlr-2013-Bayesian Nonparametric Hidden Semi-Markov Models

9 0.35909739 102 jmlr-2013-Sparse Matrix Inversion with Scaled Lasso

10 0.3059411 94 jmlr-2013-Ranked Bandits in Metric Spaces: Learning Diverse Rankings over Large Document Collections

11 0.30101767 25 jmlr-2013-Communication-Efficient Algorithms for Statistical Optimization

12 0.29613146 10 jmlr-2013-Algorithms and Hardness Results for Parallel Large Margin Learning

13 0.29249004 104 jmlr-2013-Sparse Single-Index Model

14 0.28099337 76 jmlr-2013-Nonparametric Sparsity and Regularization

15 0.26652882 97 jmlr-2013-Risk Bounds of Learning Processes for Lévy Processes

16 0.25788328 51 jmlr-2013-Greedy Sparsity-Constrained Optimization

17 0.25508884 106 jmlr-2013-Stationary-Sparse Causality Network Learning

18 0.2494517 9 jmlr-2013-A Widely Applicable Bayesian Information Criterion

19 0.24633074 77 jmlr-2013-On the Convergence of Maximum Variance Unfolding

20 0.24194108 8 jmlr-2013-A Theory of Multiclass Boosting

similar papers computed by lda model

lda for this paper:

topicId topicWeight

[(0, 0.021), (5, 0.138), (6, 0.027), (10, 0.037), (20, 0.018), (23, 0.064), (26, 0.283), (68, 0.035), (70, 0.078), (75, 0.022), (85, 0.099), (87, 0.047), (89, 0.014), (95, 0.01)]

similar papers list:

simIndex simValue paperId paperTitle

1 0.76594532 88 jmlr-2013-Perturbative Corrections for Approximate Inference in Gaussian Latent Variable Models

Author: Manfred Opper, Ulrich Paquet, Ole Winther

Abstract: Expectation Propagation (EP) provides a framework for approximate inference. When the model under consideration is over a latent Gaussian ﬁeld, with the approximation being Gaussian, we show how these approximations can systematically be corrected. A perturbative expansion is made of the exact but intractable correction, and can be applied to the model’s partition function and other moments of interest. The correction is expressed over the higher-order cumulants which are neglected by EP’s local matching of moments. Through the expansion, we see that EP is correct to ﬁrst order. By considering higher orders, corrections of increasing polynomial complexity can be applied to the approximation. The second order provides a correction in quadratic time, which we apply to an array of Gaussian process and Ising models. The corrections generalize to arbitrarily complex approximating families, which we illustrate on tree-structured Ising model approximations. Furthermore, they provide a polynomial-time assessment of the approximation error. We also provide both theoretical and practical insights on the exactness of the EP solution. Keywords: expectation consistent inference, expectation propagation, perturbation correction, Wick expansions, Ising model, Gaussian process

same-paper 2 0.72530168 105 jmlr-2013-Sparsity Regret Bounds for Individual Sequences in Online Linear Regression

Author: Sébastien Gerchinovitz

3 0.53721625 102 jmlr-2013-Sparse Matrix Inversion with Scaled Lasso

Author: Tingni Sun, Cun-Hui Zhang

Abstract: We propose a new method of learning a sparse nonnegative-deﬁnite target matrix. Our primary example of the target matrix is the inverse of a population covariance or correlation matrix. The algorithm ﬁrst estimates each column of the target matrix by the scaled Lasso and then adjusts the matrix estimator to be symmetric. The penalty level of the scaled Lasso for each column is completely determined by data via convex minimization, without using cross-validation. We prove that this scaled Lasso method guarantees the fastest proven rate of convergence in the spectrum norm under conditions of weaker form than those in the existing analyses of other ℓ1 regularized algorithms, and has faster guaranteed rate of convergence when the ratio of the ℓ1 and spectrum norms of the target inverse matrix diverges to inﬁnity. A simulation study demonstrates the computational feasibility and superb performance of the proposed method. Our analysis also provides new performance bounds for the Lasso and scaled Lasso to guarantee higher concentration of the error at a smaller threshold level than previous analyses, and to allow the use of the union bound in column-by-column applications of the scaled Lasso without an adjustment of the penalty level. In addition, the least squares estimation after the scaled Lasso selection is considered and proven to guarantee performance bounds similar to that of the scaled Lasso. Keywords: precision matrix, concentration matrix, inverse matrix, graphical model, scaled Lasso, linear regression, spectrum norm

4 0.52474743 6 jmlr-2013-A Plug-in Approach to Neyman-Pearson Classification

Author: Xin Tong

Abstract: The Neyman-Pearson (NP) paradigm in binary classiﬁcation treats type I and type II errors with different priorities. It seeks classiﬁers that minimize type II error, subject to a type I error constraint under a user speciﬁed level α. In this paper, plug-in classiﬁers are developed under the NP paradigm. Based on the fundamental Neyman-Pearson Lemma, we propose two related plug-in classiﬁers which amount to thresholding respectively the class conditional density ratio and the regression function. These two classiﬁers handle different sampling schemes. This work focuses on theoretical properties of the proposed classiﬁers; in particular, we derive oracle inequalities that can be viewed as ﬁnite sample versions of risk bounds. NP classiﬁcation can be used to address anomaly detection problems, where asymmetry in errors is an intrinsic property. As opposed to a common practice in anomaly detection that consists of thresholding normal class density, our approach does not assume a speciﬁc form for anomaly distributions. Such consideration is particularly necessary when the anomaly class density is far from uniformly distributed. Keywords: plug-in approach, Neyman-Pearson paradigm, nonparametric statistics, oracle inequality, anomaly detection

5 0.5238331 25 jmlr-2013-Communication-Efficient Algorithms for Statistical Optimization

Author: Yuchen Zhang, John C. Duchi, Martin J. Wainwright

Abstract: We analyze two communication-efﬁcient algorithms for distributed optimization in statistical settings involving large-scale data sets. The ﬁrst algorithm is a standard averaging method that distributes the N data samples evenly to m machines, performs separate minimization on each subset, and then averages the estimates. We provide a sharp analysis of this average mixture algorithm, showing that under a reasonable set of conditions, the combined parameter achieves √ mean-squared error (MSE) that decays as O (N −1 + (N/m)−2 ). Whenever m ≤ N, this guarantee matches the best possible rate achievable by a centralized algorithm having access to all N samples. The second algorithm is a novel method, based on an appropriate form of bootstrap subsampling. Requiring only a single round of communication, it has mean-squared error that decays as O (N −1 + (N/m)−3 ), and so is more robust to the amount of parallelization. In addition, we show that a stochastic gradient-based method attains mean-squared error decaying as O (N −1 + (N/m)−3/2 ), easing computation at the expense of a potentially slower MSE rate. We also provide an experimental evaluation of our methods, investigating their performance both on simulated data and on a large-scale regression problem from the internet search domain. In particular, we show that our methods can be used to efﬁciently solve an advertisement prediction problem from the Chinese SoSo Search Engine, which involves logistic regression with N ≈ 2.4 × 108 samples and d ≈ 740,000 covariates. Keywords: distributed learning, stochastic optimization, averaging, subsampling

6 0.51849568 35 jmlr-2013-Distribution-Dependent Sample Complexity of Large Margin Learning

7 0.51794255 14 jmlr-2013-Asymptotic Results on Adaptive False Discovery Rate Controlling Procedures Based on Kernel Estimators

8 0.51781642 39 jmlr-2013-Efficient Active Learning of Halfspaces: An Aggressive Approach

9 0.51219356 4 jmlr-2013-A Max-Norm Constrained Minimization Approach to 1-Bit Matrix Completion

10 0.50993985 26 jmlr-2013-Conjugate Relation between Loss Functions and Uncertainty Sets in Classification Problems

11 0.50753313 65 jmlr-2013-Lower Bounds and Selectivity of Weak-Consistent Policies in Stochastic Multi-Armed Bandit Problem

12 0.50585186 9 jmlr-2013-A Widely Applicable Bayesian Information Criterion

13 0.49850887 10 jmlr-2013-Algorithms and Hardness Results for Parallel Large Margin Learning

14 0.49711314 114 jmlr-2013-The Rate of Convergence of AdaBoost

15 0.49443701 30 jmlr-2013-Counterfactual Reasoning and Learning Systems: The Example of Computational Advertising

16 0.49245676 72 jmlr-2013-Multi-Stage Multi-Task Feature Learning

17 0.4922691 97 jmlr-2013-Risk Bounds of Learning Processes for Lévy Processes

18 0.49125859 68 jmlr-2013-Machine Learning with Operational Costs

19 0.4911975 2 jmlr-2013-A Binary-Classification-Based Metric between Time-Series Distributions and Its Use in Statistical and Learning Problems

20 0.49111879 20 jmlr-2013-CODA: High Dimensional Copula Discriminant Analysis