Online Ranking by Projecting

Size: px
Start display at page:

Download "Online Ranking by Projecting"

Transcription

1 LETTER Communicaed by Edwad Haingon Online Ranking by Pojecing Koby Camme Yoam Singe School of Compue Science and Engineeing, Hebew Univesiy, Jeusalem 91904, Isael We discuss he poblem of anking insances. In ou famewok, each insance is associaed wih a ank o a aing, which is an inege in 1 o k. Ou goal is o find a ank-pedicion ule ha assigns each insance a ank ha is as close as possible o he insance s ue ank. We discuss a goup of closely elaed online algoihms, analyze hei pefomance in he misake-bound model, and pove hei coecness. We descibe wo ses of expeimens, wih synheic daa and wih he EachMovie daa se fo collaboaive fileing. In he expeimens we pefomed, ou algoihms oupefom online algoihms fo egession and classificaion applied o anking. 1 Inoducion The anking poblem we discuss in his aicle shaes common popeies wih boh classificaion and egession poblems. As in classificaion poblems, he goal is o assign one of k possible labels o a new insance. Simila o egession poblems, he se of k labels is sucued, as hee is a oal ode elaion beween he labels. We efe o he labels as anks and wihou loss of genealiy assume ha he anks consiue he se {1, 2,...,k}. Seings in which i is naual o ank o ae insances ahe han classify ae common in asks such as infomaion eieval and collaboaive fileing. We use he lae as ou unning example. In collaboaive fileing, he goal is o pedic a use s aing on new iems such as books o movies given he use s pas aings of he simila iems. The goal is o deemine whehe a movie fan will like a new movie and o wha degee, which is expessed as a ank. An example fo possible aings migh be, un-o-see, vey-good, good, only-if-you-mus, and do-no-bohe. While he diffeen aings cay meaningful semanics, fom a leaning-heoeic poin of view, we model he aings as a oally odeed se (whose size is five in he example above). The inees in odeing o anking of objecs is by no means new and is sill he souce of ongoing eseach in many fields, such as mahemaical economics, social science, and compue science. Fo an oveview of ank- Neual Compuaion 17, (2005) c 2004 Massachuses Insiue of Technology

2 146 K. Camme and Y. Singe ing poblems fom a leaning-heoeic poin of view see Cohen, Schapie, and Singe (1999). One of he main esuls undescoed in his aicle is a complexiy gap beween classificaion leaning and anking leaning. To sidesep he inheen inacabiliy poblems of anking leaning, seveal appoaches have been suggesed. One possible appoach is o cas a anking poblem as a egession poblem. Anohe is o educe a oal ode ino a se of pefeences ove pais (Feund, Iye, Schapie, & Singe, 2003; Hebich, Gaepel, & Obemaye, 2000). The fis case imposes a meic on he se of anking ules ha migh no be ealisic, while he second appoach is ime-consuming since i equies inceasing he sample size fom n o O(n 2 ). In his lee, we conside an alenaive appoach ha diecly mainains a oally odeed se via pojecions. Ou saing poin is simila o ha of Hebich e al. (2000) in he sense ha we pojec each insance ino he eal numbes. Howeve, ou wok hen deviaes and opeaes diecly on ankings by associaing each anking wih disinc subineval of he eal numbes and adaping he suppo of each subineval while leaning. The aicle is oganized as follows. In he nex secion, we descibe a simple and efficien online algoihm ha manipulaes concuenly he diecion ono which we pojec he insances and he division ino subinevals. In secion 3, we pove he coecness of he algoihm and analyze is pefomance in he misake-bound model on vaious assumpions. Secion 4 conains he descipion and analysis of a nom-opimized vesion of he basic anking algoihm. We hen shif ou aenion o a muliplicaive algoihm ha is descibed and analyzed in secion 5. We povide empiical validaion of he meis of he vaious anking algoihms we devise in secion 6. This secion descibes expeimens compaing he anking algoihms o online classificaion and egession algoihms. Finally, we conclude wih a bief discussion and menion a few open poblems. Befoe moving on o he coe of he aicle, we poin o a few closely elaed aicles. Fis, a peliminay vesion of his lee appeaed a Neual Infomaion Pocessing Sysems, 2001, unde he ile Panking wih Ranking (Camme & Singe, 2001a). This aicle exends he confeence vesion in numeous diecions by poviding complee analysis as well as hee new algoihms ha wee no discussed in he confeence vesion. Second, in a ecen wok by Shashua and Levin (2002), an SVM-based algoihm fo insance anking was descibed. The algoihms of Shashua and Levin shae numeous popeies wih he algoihms pesened in his aicle. Howeve, i is designed fo bach seings, while ou focus is on online algoihms. Las, we would like o noe ha Haingon (2003) demonsaed empiically an impoved genealizaion pefomance of ou algoihm using an aveaging echnique, while peseving he ode of he hesholds. 2 The Basic PRank Algoihm This lee focuses on online algoihms fo anking insances. We ae given a sequence ( x 1, y 1 ),...,( x, y ),...of insance-ank pais. Fo conceeness,

3 Online Ranking by Pojecing 147 we assume ha each insance x is in R n, and is coesponding ank y is an elemen fom finie se Y wih a oal ode elaion. We assume wihou loss of genealiy ha Y ={1, 2,...,k} wih > as he ode elaion. The oal ode ove he se Y induces a paial ode ove he insances in he following naual sense. We say ha x is pefeed ove x s if y > y s.we also say ha x and x s ae no compaable if neihe y > y s no y < y s.we denoe his case simply as y = y s. Noe ha he induced paial ode is of a unique fom in ha he insances fom k equivalence classes ha ae oally odeed. 1 We sess ha alhough he elemens of Y ae denoed by ineges, we do no use he fac ha he ineges also belong o a meic space. We assume only ha he se Y is discee and is elemens ae oally odeed. A anking ule H is a mapping fom he insance space o he se of ank values, H: R n Y. The family of anking ules we discuss in his aicle employs a veco w R n and a se of k hesholds b 1 b k 1 b k =. Fo convenience, we denoe by b = (b 1,...,b k 1 ) he veco of hesholds, excluding b k, which is fixed o. Given a new insance x, he anking ule fis compues he inne poduc beween w and x. The pediced ank is hen defined o be he index of he fis (smalles) heshold b fo which w x < b. Such a anking ule divides he space ino paallel, equally anked egions: all he insances ha saisfy b 1 < w x < b ae assigned he same ank. Fomally, given a anking ule defined by w and b, he pediced ank of an insance x is H( x) = min {1,...,k} { : w x b < 0}. Noe ha he above minimum is always well defined since we se b k =. The analysis ha we use in his lee is based on he misake-bound model fo online leaning. The algoihms we descibe wok in ounds. On ound, he leaning algoihm ges an insance x. Given x, he algoihm oupus a ank, ŷ = min { : w x b < 0}. I hen eceives he coec ank y and updaes is anking ule by modifying w and b. We say ha ou algoihm made a anking misake if ŷ y and wish o make he pediced ank as close (in a sense descibed lae in his aicle) as possible o he ue ank. Fomally, he goal of he leaning algoihm is o minimize he anking loss, which is defined o be he numbe of hesholds beween he ue ank and he pediced ank. Using he epesenaion of anks as ineges in {1,...,k}, he anking loss afe T ounds is equal o he cumulaive diffeence beween he pediced ank values and ue ank values, T =1 ŷ y. The algoihm we descibe updaes is anking ule only on ounds on which he pediced ank value was incoec. Such algoihms ae called consevaive. We now descibe he updae ule of he algoihm, which is moivaed by he pecepon algoihm fo classificaion; hence, we call i he PRank algoihm, shohand fo pecepon anking. Fo simpliciy, we omi he index of he ound when efeing o an inpu insance ank pai ( x, y) and he anking ule w and b. Since b 1 b 2 b k 1 b k, he pediced 1 Fo a discussion of his ype of paial odes see Kemeny and Snell (1962).

4 148 K. Camme and Y. Singe ank is coec if w x > b fo = 1,...,y 1and w x < b fo = y,...,k 1. We epesen he above inequaliies by expanding he ank y ino k 1 viual vaiables y 1,...,y k 1. We se y =+1fo he case w x > b and y = 1fo w x < b. Pu anohe way, a ank value y induces he veco (y 1,...,y k 1 ) = (+1,...,+1, 1,..., 1) whee he maximal index fo which y =+1isy 1. Thus, he pedicion of a anking ule is coec if y ( w x b )>0fo all. If he algoihm makes a misake by anking x as ŷ insead of y, hen hee is a leas one heshold, indexed, fo which he value of w x is on he wong side of b, ha is, y ( w x b ) 0. To coec he misake, we need o move he values of w x and b owad each ohe. We do so by modifying only he values of he b s fo which y ( w x b ) 0 and eplace hem wih b y. We also eplace he value of w wih w + ( y ) x whee he sum is aken ove he indices fo which hee was a pedicion eo, ha is, y ( w x b ) 0. An illusaion of he updae ule is given in Figue 1. In he example, we used he se Y ={1,...,5}. (Noe ha b 5 = is omied fom all he plos in Figue 1.) The coec ank of he insance is y = 4, and hus he value of w x should fall in he fouh ineval, beween b 3 and b 4. Howeve, in he illusaion, he value of w x fell below b 1 and he pediced ank is ŷ = 1. The heshold values b 1, b 2, and b 3 ae a souce of he eo since he value of b 1, b 2, b 3 is highe han w x. To compensae fo he misake, he algoihm deceases b 1, b 2, and b 3 by a uni value and eplaces hem wih b 1 1, b 2 1, and b 3 1. I also modifies w o be w + 3 x since y = 3. :y ( w x b ) 0 Thus, he inne poduc w x inceases by 3 x 2. This updae is illusaed a he middle plo of Figue 1. The pedicion ule afe he updae is illusaed on he igh-hand side of Figue 1. Noe ha afe he updae, he pediced ank of x is ŷ = 3, which is close o he ue ank y = 4. The pseudocode of algoihm is given in Figue 2. To conclude his secion, we noe ha PRank can be saighfowadly combined wih Mece kenels (Vapnik, 1998) and voing echniques (Feund & Schapie, 1999) ofen used fo impoving he pefomance of magin classifies in bach and online seings (Cisianini & Shawe-Taylo, 2000). To do so, we assume access o an inne-poduc space X equipped wih a kenel opeao K: X X R. We now need o eplace any innepoduc opeaion he algoihm pefoms wih an implici inne-poduc opeaion defined via he kenel opeao. We hus keep inne-poduc opeaos ahe han explici vecos. Fo insance, he updae of PRank becomes K( w +1, ) K( w, ) + ( τ ) K( x, ),

5 Online Ranking by Pojecing 149 Pediced ank Coec ineval Coec ineval Updaed pediced ank Coec ineval Figue 1: An Illusaion of he updae ule. The anking ule pedics a ank of ŷ = 1 insead of y = 4 (op). The updae deceases he hesholds b 1, b 2, b 3 by one uni and eplaces w wih w + 3 x (cene), yielding ha he pediced ank of x afe he updae is ŷ = 3 (boom). and hus he inne poduc w T x becomes 1 τ K( xs, x ). s=1 3 Analysis Befoe we pove he misake bound of he algoihm, we fis need o show ha i mainains a consisen hypohesis. Tha is, we need o show ha PRank peseves he coec ode of he hesholds; ohewise, i migh be impossible o induce a ank pedicion ule. We pove ha he consisency of hesholds is mainained by showing inducively ha fo any anking ule ha can be deived by he algoihm along is un, ( w 1, b 1 ),...,( w T+1, b T+1 ) he se of inequaliies b b k 1 hold fo all. Clealy, since he iniializaion of he hesholds is such ha b 1 1 b1 2 b1 k 1, hen i suffices

6 150 K. Camme and Y. Singe Iniialize: Se w 1 = 0, b 1 1,...,b1 k 1 = 0, b1 k = Loop: Fo = 1, 2,...,T Receive a new insance x R n Pedic: ŷ = min { : w x b < 0} {1,...,k} Receive a new ank-value y If ŷ y updae w (ohewise se w +1 = w, : b +1 = b ): 1. Fo = 1,...,k 1: Ify Then y = 1 Else y = 1 2. Fo = 1,...,k 1: If( w x b )y 0 Then τ = y Else τ = 0 3. Updae: w +1 w + ( τ ) x Fo = 1,...,k 1: b +1 b τ Oupu: H( x) = min {1,...,k} { : w T+1 x b T+1 < 0}. Figue 2: The PRank algoihm. o show ha he claim holds inducively. Fo simpliciy of he poof below, le us wie he updae ule of PRank in an alenaive fom. Le [[π]] be 1 if he pedicae π holds and 0 ohewise. We now ewie he value of τ (fom Figue 2) as τ = y [[( w x b )y 0]]. Noe also ha he values of b ae ineges fo all and since fo all, we iniialize b1 = 0 and b +1 b { 1, 0, +1}. Lemma 1 (ode pesevaion). Le w and b be he cuen anking ule, whee b 1 b k 1, and le ( x, y ) be an insance ank pai fed o PRank on ound. Denoe by w +1 and b +1 he esuling anking ule afe he updae of PRank. Then b +1 1 b +1 k 1. Poof. In ode o show ha PRank peseves a nondeceasing ode of he hesholds, we use he definiion of he algoihm fo y. We define y =+1 fo < y and y = 1fo y. To pove ha b b+1, we ewie he heshold updae as b +1 = b y [[( w x b )y 0]]. (3.1)

7 Online Ranking by Pojecing 151 In ode o pove ha b +1 ha +1 b+1 fo all feasible, i is sufficien o show b +1 b y +1 [[( w x b +1 )y +1 0]] y [[( w x b )y 0]].(3.2) Since by ou inducive assumpion b +1 b and b, b +1 Z, we ge ha he value of b +1 b on he lef-hand side of equaion 3.2 is a nonnegaive inege. Recall also ha y = 1ify > and y = 1ohewise, and heefoe y +1 y. We now need o analyze wo cases. We fis conside he case whee y +1 y, which implies ha y +1 = 1, y = +1. In his case, he igh-hand side of equaion 3.2 is a mos zeo, and he claim ivially holds. The second case is when y +1 = y. Hee we ge ha he value of he igh-hand side of equaion 3.2 canno exceed 1. If b +1 > b, hen since boh values ae ineges, he bound of 1 on he igh-hand side of equaion 3.2 implies ha he value of b canno exceed he value of b +1. We ae hus lef wih he case whee b = b +1 and y +1 = y. Fo his case, we have ha he ems y +1 [[( w x b +1 )y +1 < 0]] and y [[( w x b )y < 0]] aain he same value. Theefoe, he igh-hand side of equaion 3.2 is zeo and b +1 = b +1. This complees he poof. +1 We now un o he misake-bound analysis of he algoihm. In ode o simplify he analysis of he algoihm, we inoduce he following noaion. Given a hypeplane w and a se of k 1 hesholds b, we denoe by v R n+k 1 he veco ha is a concaenaion of w and b, ha is, v = ( w, b). Fo beviy, we efe o he veco v as a anking ule. Given wo vecos v = ( w, b ) and v = ( w, b), we have v v = w w + b b and v 2 = w 2 + b 2. Noe ha a veco v induces a paial ode of he vecos R n. Theoem 1 (misake bound). Le ( x 1, y 1 ),...,( x T, y T ) be an inpu sequence o PRank whee x R n and y {1,...,k}. Denoe by R 2 = max x 2.If hee exiss a anking ule v = ( w, b ) wih b 1 b k 1 of a uni nom ha classifies he enie sequence coecly wih magin γ = min, {( w x b )y } > 0, he anking loss of he algoihm T =1 ŷ y is a mos (k 1) R2 + 1 γ 2. Poof. Le us examine an example ( x, y ) ha he algoihm eceived on ound. By definiion, he algoihm anked he example using he anking ule v, which is composed of w and he se of hesholds b. Similaly, v +1 is he anking ule ( w +1, b +1 ) afe ound. Theefoe, w +1 = w + ( τ ) x and b +1 = b τ fo = 1, 2,...,k 1. Le us denoe by n = ŷ y he

8 152 K. Camme and Y. Singe diffeence beween he ue ank and he pediced ank. Since τ is zeo fo all he indices such ha sign( w x b )y > 0, hen i is saighfowad o veify ha n = τ. Noe ha if hee was no a anking misake on ound, hen τ = 0 fo = 1,...,k 1, and hus also n = 0. To pove he heoem, we bound n fom above by bounding v 2 fom above and below. Fis, we deive a lowe bound on v 2 by bounding v v +1. Subsiuing he values of w +1 and b +1, we ge ha k 1 v v +1 = v v + τ ( w x b ). (3.3) =1 We fuhe bound he em on he igh-hand side by consideing wo cases coesponding o whehe τ is posiive o zeo. Using he definiion of τ fom he pseudocode in Figue 2, we fis examine he case whee ( w x b )y 0 and heefoe τ = y. Fom he assumpion ha v anks he examples coecly wih a magin of a leas γ, we ge ha τ ( w x b ) γ. The second case is when ( w x b )y > 0. In his case, we have τ = 0 and hus τ ( w x b ) = 0. Combining he wo cases and summing now ove, wege k 1 k 1 τ ( w x b ) τ γ = n γ. (3.4) =1 =1 Combining equaions 3.3 and 3.4, we ge ha v v +1 v v +n γ. Unfolding he sum, we ge ha afe T ounds, he pojecion of he veco v T+1 on v saisfies v v T+1 n γ = γ n. (3.5) Recall ha Cauchy-Schwaz inequaliy implies ha v T+1 2 v 2 ( v T+1 v ) 2. Using equaion 3.5 wih Cauchy-Schwaz inequaliy and he assumpion ha v is of a uni nom, we ge he following lowe bound: ( ) 2 v T+1 2 n γ 2. (3.6) We nex bound he nom of v T+1 fom above. As befoe, assume ha an example ( x, y ) was anked using he anking ule v, and denoe by v +1

9 Online Ranking by Pojecing 153 he anking ule a he end of ound. We now expand he values of w +1 and b +1 whose sum is he nom of v +1 and ge v +1 2 = w 2 + b τ ( w x b ) + ( τ ) 2 x 2 + (τ )2. Since τ { 1, 0, +1}, we have ha ( τ )2 (n ) 2 and (τ )2 = n, and we heefoe ge v +1 2 v τ ( w x b ) + (n ) 2 x 2 + n. (3.7) We fuhe develop he second em using he updae ule of he algoihm and ge ( w x b ) = [[( w x b )y 0]] ( ( w x b ) )y 0. (3.8) τ Plugging equaion 3.8 ino equaion 3.7 and using he bound x 2 R 2, we ge ha v +1 2 v 2 + (n ) 2 R 2 + n. Thus, he anking ule we obain afe T ounds of he algoihm is uppe bounded by v T+1 2 R 2 (n ) 2 + n. (3.9) Combining he lowe bound v T+1 2 ( n ) 2 γ 2 wih he uppe bound of equaion 3.9, we have ha ( ) 2 n γ 2 v T+1 2 R 2 (n ) 2 + n. Dividing he above equaions by γ 2 n, we finally ge n R2 [ (n ) 2 ]/[ n ] + 1 γ 2. (3.10) Since by definiion, n is a mos k 1, i implies ha (n ) 2 n (k 1) = (k 1) n.

10 154 K. Camme and Y. Singe Plugging his inequaliy ino equaion 3.10, we ge he desied bound, T ŷ y = =1 T =1 n (k 1)R2 + 1 γ 2 (k 1) R2 + 1 γ 2. We now un ou aenion o he insepaable case in which hee does no exis a posiive magin value γ. To analyze he insepaable case, we use an analysis echnique suggesed by Feund and Schapie (1999). (This poof echnique was fis infomally given by Vapnik, 1998.) To pove a misake bound fo he insepaable case, each example ( x, y ) is augmened wih a slack vaiable, denoed d. Infomally, fo each ime sep, he vaiable d designaes how much he magin assumpion is violaed by he example. Fomally, we obain he following misake bound fo he insepaable case: Theoem 2 (misake bound fo insepaable case). Le ( x 1, y 1 ),...,( x T, y T ) be an inpu sequence fo PRank whee x R n and y {1,...,k}. Denoe by R 2 = max x 2. Le v = ( w, b ) be a anking ule of a uni nom wih b 1 b k 1. Le γ>0and define d = max{0,γ min{( w x b )y }}. (3.11) Denoe by D 2 = (d ) 2. Then he anking loss of he algoihm is bounded above by T =1 ŷ y (k 1) (D + R 2 + 1) 2 γ 2. The poof of he heoem is based on he poof echnique fo heoem 1 and is given in he appendix. To conclude his secion, we descibe and biefly analyze a simplified vesion of PRank. This vesion shaes he same algoihmic skeleon as PRank and diffes only in he way i modifies is anking ule. Rahe han modifying all of he hesholds in he eo se, his vesion chooses a single heshold o modify and updae w accodingly. Since a single heshold is modified on each ound, we use he abbeviaion Si-PRank o efe o his vesion. Moe fomally, we eplace sep 3 in Figue 2 wih he following seps, 3. Define updae index: If ŷ > y Then = ŷ 1 Else = ŷ 4. Updae: w +1 w + τ x b τ b +1 s b s (s ). b +1

11 Online Ranking by Pojecing 155 I is o veify ha his choice of a single heshold o updae peseves he oveall ode of he heshold. The poof is immediae consequence of he lemma 1. In addiion, following exacly he same poof echnique of heoem 1, we ge ha he misake bound of PRank also holds fo Si-PRank. 2 Supisingly, as we see lae in secion 6, in pacice Si-PRank pefoms slighly bee han PRank. 4 A Nom-Opimized Vesion of PRank The PRank algoihm pesened in he pevious secion pefoms he same fom of updae wheneve a anking eo occus. Fuhemoe, even when he pojecion of x ono w yields he coec ank value, he value of w x migh lie vey close o one of he hesholds b. The diffeence beween he pojecion of x and he closes heshold plays a simila ole o he noion of magin in classificaion poblems and was used in heoem 1 o deive a misake bound fo PRank. In his secion, we pesen a vesion of PRank ha solves on each ound a mini-opimizaion poblem ha balances beween wo opposing equiemens. On one hand, we equie ha he new anking ule v +1 be as simila as possible o he pevious anking ule v, which encompasses all ou knowledge on pas examples. On he ohe hand, we foce he new anking ule o ank he mos ecen example x coecly and wih a lage enough magin. Fomally, assuming ha hee was a ank pedicion eo on ound, hen we equie ha he new ule ( w +1, b +1 ) would saisfy, min {( w +1 x b +1 )y }} β, whee β is a posiive consan. These wo equiemens yield he following opimizaion poblem: min w, b 1 2 ( w, b) ( w, b ) 2 subjec o: ( w x b )y β fo = 1,...,k 1. (4.1) We call his vesion he Nom-Opimized PRank algoihm, o No-PRank in sho. The pseudocode of he algoihm appeas in Figue 3. Befoe analyzing he algoihm, le us fis fuhe develop he opimizaion poblem given in equaion 4.1 by expanding is Lagangian funcion, L = 1 2 ( w, b) ( w, b k 1 ) 2 τ [( w x b )y β]. (4.2) =1 2 In fac, an alenaive bound can be given fo Si-PRank, which saes ha he numbe of ounds on which an eo occus (indicaed by a sicly posiive value of he loss) is uppe bounded by R γ 2.

12 156 K. Camme and Y. Singe Inpu: Minimal magin paamee β Iniialize: Se w 1 = 0, b 1 1,...,b1 k 1 = 0, b1 k = Loop: Fo = 1, 2,...,T Receive a new insance x R n. Pedic: ŷ = min { : w x b < 0} {1,...,k} Receive a new ank-value y If y ŷ updae w and b : (ohewise se w +1 = w, b +1 = b ): 1. Fo = 1,...,k 1: Ify Then y = 1 Else y = Updae: se ( w +1, b +1 ) R n+k 1 o be he minimize of: min w, b 1 2 ( w, b) ( w, b ) 2 subjec o: ( w x b )y β fo = 1,...,k 1. Oupu: H( x) = min {1,...,k} { : w T+1 x b T+1 < 0}. Figue 3: The Nom-Opimized PRank algoihm. Hee, τ ae nonnegaive Lagange muliplies. Taking he deivaive of equaion 4.2 wih espec o w and compaing i o zeo, we ge w L = w w x Repeaing he pocess fo b,wege ( ) τ y = 0 w = w + τ y x. (4.3) b L = b b + τ y = 0 b = b τ y. (4.4) Plugging he value of w fom equaion 4.3 and b fom equaion 4.4 ino equaion 4.2, we ge he following dual poblem: ( ) 2 minq ( τ) = 1 τ 2 x 2 τ y + 1 τ [ ( τ y w x b ) ] β (4.5) s. 0 τ = 1,...,k 1.

13 Online Ranking by Pojecing 157 Befoe poceeding o discuss he fomal popeies of he No-PRank algoihm, i is woh noing ha he new veco obained a he end of ound, w +1 is a linea combinaion of w and he cuen insance x. Theefoe, i is ahe saighfowad o use No-PRank in conjuncion wih kenel mehods by mainaining w in is dual fom as a weighed combinaion of he insances x 1,..., x. Noe also ha he opimizaion poblem of No-PRank educes o a simple opimizaion poblem wih a quadaic objecive funcion and nonnegaiviy consains and can be solved using sandad convex opimizaion ools (Fleche, 1987). Le us now discuss he fomal popeies of No-PRank. The following simple lemma saes ha No-PRank is a consevaive online algoihm as i modifies is anking ule if and only if he minimal magin equiemens ae no aained fo he cuen insance. Lemma 2 (consevaiveness). Le ( x, y ) denoe he inpu o No-PRank on ound, and le y be 1 if y and +1 ohewise. Then if y ( w x b ) β fo all = 1,...,k 1he opimum of No-PRank s opimizaion poblem is aained a τ = 0 fo all. Poof. Conside he dual fom of No-PRank s opimizaion poblem given by equaion 4.5. The objecive funcion Q ( τ) is composed of hee summands. The fis wo ae clealy nonnegaive and aain a value of zeo iff all he muliplies τ ae zeo. If y ( w x b ) β, hen he hid summand is linea in τ wih nonnegaive coefficiens. Thus, is minimum is also aained when τ is zeo fo all. Nex we show ha No-PRank peseves he ode of he hesholds along is un and hus can always seve as a valid anking ule. Lemma 3 (ode pesevaion). Le w and b be he cuen anking ule, whee b 1 b k 1, and le ( x, y ) be an insance-ank pai fed o No-PRank on ound. Denoe by w +1 and b +1 he esuling anking ule afe he updae of No-PRank. Then b +1 1 b +1 k 1. Poof. In ode o show ha No-PRank mainains a coec (monoonically inceasing) ode of he hesholds, we use he vaiables employed by he algoihm along is un. Namely, we se y =+1fo < y and y = 1fo y. To pove ha b b+1 fo all, we expand b +1 and show ha b +1 b y +1 τ +1 y τ. (4.6) We need o analyze wo diffeen seings. The fis seing we analyze is when y +1 y, which implies ha y +1 = 1and y =+1. In his case, he igh-hand side of equaion 4.6 is a mos zeo, while he lef-hand side of

14 158 K. Camme and Y. Singe he equaion is a leas zeo. Thus, he claim holds. The ohe case is when y +1 = y = y. In his case, equaion 4.6 becomes b +1 b y(τ +1 τ ). (4.7) Assume by conadicion ha he opimal value of equaion 4.5 does no saisfy equaion 4.7 fo some. We now consuc anohe feasible se fo τ ha yields a lowe value of he objecive funcion Q. Le us define τ s yɛ s = + 1 τ s = τ s + yɛ s = ohewise, τ s fo some value of ɛ>0, which is deemined below. Infomally, he value of ɛ is se o be small enough so ha he consain τ s 0 sill holds. (This is possible since τ +1 > 0.) Since τ and τ diffe only a hei and + 1 componens, we ge Q( τ ) Q( τ) = 1 2 (τ 2 + τ 2 +1 ) + τ [y ( w x b ) β] + τ +1 [y +1( w x b +1 ) β] 1 2 (τ 2 + τ 2 +1 ) τ [y ( w x b ) β] τ +1 [y +1 ( w x b +1 ) β]. Expanding τ, we now ge Q( τ ) Q( τ) = 1 2 ((τ + yɛ) 2 + (τ +1 yɛ) 2 ) + (τ + yɛ)[y ( w x b ) β] + (τ +1 yɛ)[y +1 ( w x b +1 ) β] 1 2 (τ 2 + τ +1 2 ) τ [y ( w x b ) β] τ +1 [y +1 ( w x b +1 ) β] = (yɛ) 2 + yɛ(τ τ +1 ) + yɛ[y ( w x b ) β] yɛ[y +1 ( w x b +1 ) β]. Denoing boh y and y +1 simply as y and using he fac y { 1, +1}, we ge Q( τ ) Q( τ) = (yɛ) 2 + yɛ(τ τ +1 ) ɛb + ɛb +1 = ɛ[ɛ (y(τ +1 τ ) (b +1 b ))].

15 Online Ranking by Pojecing 159 Since we assumed by conadicion ha y(τ +1 τ ) (b +1 b ) = A > 0, we can choose 0 <ɛ A/2 and ge Q( τ ) Q( τ) = ɛ(a/2 A) = ɛa/2 < 0, which conadics he assumpion ha τ equaion 4.5. is he opimal soluion of We now un o he analysis of he pefomance of No-PRank. We fis bound he sum of he weigh employed by he dual pogam as hey accumulae along he un of No-PRank τ. Based on he bound on he weighs, we hen pove a misake bounds on he numbe of ounds on which an eo occued, ha is, y ŷ. As we see sholy, he bound depends on boh he value of he aainable magin γ and he magin paamee, β, employed by he algoihm. Theoem 3 (bound on weighs). Le ( x 1, y 1 ),...,( x T, y T ) be an inpu sequence o No-PRank whee x R n and y {1,...,k}. Le v = ( w, b ) be a anking ule wih b 1 b k 1 and w 2 + (b )2 = 1. Assume ha v classifies he enie sequence coecly wih a magin value γ = min, {( w x b )y } > 0. Then he oal sum of he weighs geneaed by No-PRank is bounded by T τ 2 β γ 2, =1 whee β is a pedefined paamee of he algoihm (see Figue 3). Poof. Le us concenae on an example ( x, y ) ha he algoihm eceived on ound. By consucion, he algoihm anked he example using he anking ule v, which is composed of w and he hesholds b. Similaly, we denoe by v +1 he updaed ule ( w +1, b +1 ) afe ound, ha is, ( ) w +1 = w + y τ x and b +1 = b y τ fo = 1, 2,...,k 1. To bound τ fom above, we deive bounds on v 2 fom boh above and below. Fis, we deive a lowe bound on v 2 by bounding v v +1. Subsiuing he values of w +1 and b +1, we ge k 1 v v +1 = v v + τ y ( w x b ). =1

16 160 K. Camme and Y. Singe Using he assumpion ha v anks he daa coecly wih a magin of a leas γ, we ge ha y ( w x b ) γ : k 1 v v +1 v v + τ γ = v v + γ =1 k 1 τ. =1 Unfolding he sum, we ge ha afe T ounds, he algoihm saisfies v v T+1 γ, τ. Plugging his esul ino Cauchy-Schwaz inequaliy, v T+1 2 v 2 ( v T+1 v ) 2, and using he assumpion ha v is of a uni nom, we ge he lowe bound, ( ) 2 v T+1 2 τ γ 2. (4.8), We nex bound he nom of v T+1 fom above. As befoe, assume ha an example ( x, y ) was anked using he anking ule v, and denoe by v +1 he anking ule afe he ound. We now expand he values of w +1 and b +1 in he nom of v +1 and ge v +1 2 = w 2 + b τ y ( w x b ) + x 2 ( ) 2 + (y τ )2. τ y We add and subac he em 2β τ on he igh-hand side of he above equaion and ge v +1 2 = w 2 + b τ [y ( w x b ) β] + x 2 ( ) 2 + ( y τ ) 2 + 2β τ y = w 2 + b 2 + 2Q( τ ) + 2β τ, (4.9) τ whee we used he definiion of Q fom equaion 4.5 o obain he las equaliy.

17 Online Ranking by Pojecing 161 Noe ha τ = 0 is a feasible soluion fo he opimizaion poblem posed in equaion 4.5. The value aained by Q fo his paicula choice of τ is zeo. Since τ is he minimize Q( τ), we mus have ha Q( τ ) 0. (4.10) Subsiuing equaion 4.10 in equaion 4.9, we obain he following bound on he nom of v +1 in ems of he nom of v : v +1 2 v 2 + 2β τ. (4.11) Thus, he anking ule we obain afe T ounds of he algoihm saisfies he uppe bound: v T+1 2 2β, τ. (4.12) Combining he lowe bound of equaion 4.8 wih he uppe bound of equaion 4.12, we have ha ( ) 2 γ 2 v T+1 2 2β τ., τ Dividing boh sides by γ 2, τ, we finally ge,, τ 2β γ 2. (4.13) We now discuss some popeies of he algoihm and is coesponding loss bound. As menioned above, he algoihm updaes is anking ule only on ounds on which he magin is less han β, which eflecs a minimal magin equiemen. Infomally, β can be viewed as a minimal diffeence equiemen beween he pojecion of he example x ono w and any of he hesholds b (see Figue 4). Thus, he lage β is, he bee is he sepaaion beween he ank levels. Fomally, he algoihm aemps o enclose all he examples in subinevals of he eal numbes defined by he hesholds. As saed above, based on heoem 3, we can deive a misake bound fo No- PRank. Supisingly, he dependency on he magin paamee β cancels ou, and he end esul is a misake bound ha depends on only he geomeical popeies of he poblem: he nomalized magin. Coollay 1 (misake bound). Assume he condiions of heoem 3 hold, and denoe by R = max x. Then he numbe of ounds fo which No-PRank made

18 162 K. Camme and Y. Singe β W X β b 1 b 2 b 3 b 4 Figue 4: An illusaion of he magin equied by he No-PRank algoihm. a misake is uppe bounded by 2 R2 + 1 γ 2. Poof. To pove he coollay, we use heoem 3 and show ha wheneve an eo occued on ound, he sum of he dual weighs τ is a leas β/(r 2 + 1). Since w +1 mus saisfy he consains of equaion 4.1, we now show ha fo = 1,...,k 1, ( w +1 x b +1 )y β. Subsiuing w +1 = w + x τ y and b+1 fom equaion 4.3 and b +1 = b τ y fom equaion 4.4, we ge ha he dual vaiables τ 1,...,τ k 1 saisfy [( β w + x s τ s y s ) ] x b + τ y y. Reaanging ems in he above equaion, we ge [( β w + x s τ s y s ) ] x b + τ y y = ( w x b )y + y x 2 s τ s y s + τ (y )2. (4.14) Since a ank pedicion eo occued on ound, hee exiss such ha ( w x b )y 0. In addiion, since y { 1, +1}, we have ha y s τ s y s s τ s. Using hese wo inequaliies in equaion 4.14 in conjuncion wih

19 Online Ranking by Pojecing 163 he fac ha τ 0, we ge β 0 + x 2 (R 2 + 1) τ + τ. τ We have hus shown ha if hee was a ank pedicion eo on ound, hen β R τ. (4.15) To pove he coollay, we combine heoem 3 wih equaion Denoe by M he numbe of ounds wih pedicion eos. We now show ha on each such ound, β/(r 2 + 1) τ, and fo he es of he ounds, we simply bound he sum τ fom below by 0. We hus ge ha M β R 2 + 1, τ. Applying heoem 3, we ge M β R β γ 2 M 2 R2 + 1 γ 2. Noe ha he misake bounds of heoem 1 and coollay 1 ae idenical up o a muliplicaive (k 1)/2 faco. Fuhemoe, if we employ he ivial fac ha y ŷ k 1, we can immediaely obain a bound on y ŷ fom coollay 1 ha is a 2 faco of he bound given in heoem 1. Howeve, he bound fo No-PRank is moe efined, as in addiion o he misake bound, we can also bound he oal sum of he weighs, τ. Unfounaely, since τ may be abiaily close o zeo, he moe efined analysis does no yield a bee misake bound. One possible diecion fo impoving he misake bound iself may be obained by modifying he minimal magin equiemen fom β o β y ŷ. Anohe viable appoach is o eplace he fixed magin consain ( w +1 x b +1 )y β wih a magin equiemen ha is dependen on he diffeence beween he coec label y and he specific heshold. Specifically, we define a new se of consains ( w +1 x b +1 )y βa, whee a = y fo < y and a = + 1 y fo y. Theefoe, he fuhe he heshold fom he ineval conaining he pojecion of he insance on he hypeplane, he lage he magin we equie. We leave hese possible exensions o fuue eseach.

20 164 K. Camme and Y. Singe 5 A Muliplicaive Vesion of PRank In his secion, we give a muliplicaive vesion of PRank ha we em Mu- PRank. This vesion is analogous o he basic PRank updae, bu i modifies w and b in a muliplicaive manne. Tha is, on each ound wih ank pedicion eo, he cuen weigh veco w and hesholds b ae muliplied by facos ha depend on x. The moivaion fo his vesion is he muliplicaive updaes employed by he wok of Wamuh and colleagues on online pedicion (see, e.g., Kivinen & Wamuh, 1997). Mu-PRank mainains a nomalized anking ule, ha is, ( w, b ) 1 = 1 fo all. As in PRank, on ound, Mu-PRank compues he veco τ, which deemines he updae of w and b. I hen akes he exponen of he updaed anking ule and nomalizes i so ha he l 1 nom of ( w +1, b +1 ) will be 1. The pseudocode of he algoihm is given in Figue 5. Examining he logaihm of b on each ound, i is ahe simple o ve- Inpu: Leaning ae η Iniialize: Se w 1 i = 1 n+k 1 i = 1,...,n, b 1 1,...,b1 k 1 = 1 n+k 1, b1 k = Loop: Fo = 1, 2,...,T Receive a new insance x R n, x 1 Pedic: ŷ = min { : w x b < 0} {1,...,k} Receive a new ank-value y If y ŷ updae w and b (ohewise se w +1 = w, b +1 = b ): 1. Fo = 1,...,k 1: Ify Then y = 1 Else y = 1 2. Fo = 1,...,k 1: If( w x b )y 0 Then τ = y Else τ = 0 3. Define: n Z = w i eηx i τ k 1 + b e ητ i=1 =1 4. Updae: Fo i = 1,...,n : wi +1 w i eηx i /Z Fo = 1,...,k 1:b +1 b e ητ /Z Oupu: H( x) = min {1,...,k} { : w T+1 x b T+1 < 0} Figue 5: The muliplicaive algoihm.

21 Online Ranking by Pojecing 165 ify ha, like PRank, Mu-PRank peseves he ode of he hesholds. The addiional sep Mu-PRank employs in is updae sage is he nomalizaion of is anking ule. Howeve, since his nomalizaion is applied o all he hesholds, i clealy keeps he ode of he hesholds inac. Moe fomally, log(b ) can be wien as ηu + log(c ), whee u is an inege and C is independen of. Applying exacly he same poof echnique used in lemma 1 o u gives he following coollay: Lemma 4 (ode pesevaion). Le w and b denoe he cuen anking ule and assume ha b 1 b k 1. Then, afe ( x, y ) is fed o Mu-PRank, he new ule ( w +1, b +1 ) peseves he ode of he hesholds, b +1 1 b +1 k 1. Nex, we analyze he misake bound of Mu-PRank. Noe ha Mu-PRank employs a leaning ae paamee, η. The misake bound of Mu-PRank given in he following heoem depends on his value. If an pioi lowe bound on he magin γ is known, we can fix he value of η so as o minimize he loss bound of Mu-PRank. The esuling bound is descibed in coollay 2, which follows heoem 4. Theoem 4 (misake bound). Le ( x 1, y 1 ),...,( x T, y T ) be an inpu sequence o Mu-PRank, whee x R n, x 1, and y {1,...,k}. Assume ha hee exiss a anking ule v = ( w, b ) wih b 1 b k 1 and w, b 1 = 1, which classifies he enie sequence coecly wih magin γ = min, {( w x b )y } > 0. Then he anking loss of Mu-PRank, T =1 ŷ y is, a mos, log ( k + n 1 ) (. log + ηγ ) 2 e η(k 1) +e η(k 1) Poof. As befoe, le v = ( w, b ) and v +1 = ( w +1, b +1 ) denoe he anking ules a he beginning and end of ound, especively. To pove he heoem, we analyze he decease in he Kullback-Leible (KL) divegence (Cove & Thomas, 1991) beween v and v. The KL divegence of wo discee pobabiliy disibuions p, q is D KL ( p q) = i p i log(p i /q i ). To pove he heoem, we examine he change of he KL divegence in wo consecuive ounds: = D KL ( v v +1 ) D KL ( v v ). We bound fom above and below. We fis bound his sum fom above.

22 166 K. Camme and Y. Singe As a eminde, we use n o denoe ŷ y. Using his noaion, we ge = = n i=1 ( w w ) i log i w i n ( w i log i=1 e ηx i [ n k 1 = w i + i=1 Z b =1 k 1 + τ =1 ( b b ) log b ) k 1 ( Z + b ) log =1 =1 e ητ ] log ( Z ) k 1 η τ ( w x b ) k 1 log(z ) ηγ τ =log(z ) ηγ n, (5.1) =1 whee we used he definiion of n and τ in conjuncion wih he fac ha v achieves a magin of γ o obain he inequaliy above. We now bound log(z ). We need he following inequaliy, e ηx a + x 2a eηa + a x 2a e ηa, (5.2) which holds fo a,η > 0 and x [ a, a]. (The poof of his inequaliy is an immediae applicaion of he convexiy of he exponen funcion.) Now ecall ha Z = n i=1 w i eηx i τ k 1 + b e ητ. (5.3) =1 We bound he lef-hand side and he igh-hand side of he above sum sepaaely. Fo he lef-hand side, we bound each em e ηx i τ. Using equaion 5.2 in conjuncion wih he fac ha x i 1and τ k 1, we ge e ηx i τ k 1 + x i 2(k 1) τ Similaly, τ 1 k 1, and hus e η(k 1) + k 1 x i 2(k 1) τ e η(k 1). (5.4) e ητ k 1 + τ 2(k 1) eη(k 1) + k 1 τ 2(k 1) e η(k 1). (5.5) Using he bounds fom equaions 5.4 and 5.5 in 5.3, we ge Z i w i [ k 1 + x i 2(k 1) τ e η(k 1) + k 1 x i 2(k 1) τ e η(k 1) ]

23 Online Ranking by Pojecing b [ k 1 + τ 2(k 1) eη(k 1) + k 1 τ ] 2(k 1) e η(k 1) = i w 1 i 2 (eη(k 1) + e η(k 1) ) + b 1 2 (eη(k 1) + e η(k 1) ) + τ ( w x b ) eη(k 1) + e η(k 1). (5.6) 2(k 1) The definiion of τ implies ha τ ( w x b ) 0. (Eihe hee was a anking eo o τ = 0). In addiion, since Mu-PRank nomalizes is anking ule a he end of each ound, we know ha ( w, b ) 1 = i w i + b = 1. Using hese facs in equaion 5.6, we ge he following bound on Z : Z 1 2 (eη(k 1) + e η(k 1) ). (5.7) Using equaion 5.7 in equaion 5.1, we obain an uppe bound on : [ ] 1 log 2 (eη(k 1) + e η(k 1) ) ηγ n. Since a anking eo occued, we know ha n 1. In addiion, he agumen of he logaihm is a leas 1; we can eaange ems and wie [ ] } 1 n {log 2 (eη(k 1) + e η(k 1) ) ηγ. (5.8) Summing ove, wege { [ 1 ( log e η(k 1) + e η(k 1))] } ηγ n. (5.9) 2 To bound fom below, we unavel he sum and ge = (D KL ( v v +1 ) D KL ( v v )) = D KL ( v v T+1 ) D KL ( v v 1 ) D KL ( v v 1 ), whee he inequaliy is due o he fac ha he KL divegence is always nonnegaive. Using he value of w 1, we ge ha D KL ( v v 1 ) = log(k 1 + n). Combining he lowe bound on wih equaion 5.9, we ge log(k 1 + n) { ( log 2 e η(k 1) + e η(k 1) ) } + ηγ n.

24 168 K. Camme and Y. Singe Reaanging ems, we finally ge he desied bound: n log ( k 1 + n ) (. log + ηγ ) 2 e η(k 1) +e η(k 1) As discussed above, he bound of heoem 4 depends on he leaning ae η.ifγ o a lowe bound on γ is known, we can se η o be ( ) 1 k 1 + γ η = 2(k 1) log. k 1 γ Fo his choice of η, we ge he following coollay: Coollay 2. η = If we un Mu-PRank wih ( ) 1 k 1 + γ 2(k 1) log, k 1 γ hen unde he assumpions of heoem 4, he cumulaive anking loss obained by Mu-PRank is bounded above by n (k 1) 2 log ( n + k 1 ) γ 2. This coollay implies ha he cumulaive anking loss of Mu-PRank is invesely popoional o squae of he magin γ. Noe ha a diec compaison of his bound o he misake bound of PRank (see heoem 1) is no possible as he noion of magin hee is diffeen. (Fo PRank, we used he l 2 nom of he insances and he anking ule o define he magin, while fo Mu- PRank, we acily use he l 1 nom of he anking ule in conjuncion wih he l of he insances.) This elaion beween he bounds also holds beween addiive and muliplicaive online algoihms fo classificaion (Rosenbla, 1958; Lilesone, 1987, 1988, 1989). Puing hese diffeences in he noion of magin aside, he wo bounds exhibi he same quadaic dependency on he magin. 6 Expeimens In his secion, we descibe expeimens we pefomed ha compaed he vaians of PRank discussed in he pevious secion wih wo ohe online leaning algoihms applied o anking: a muliclass genealizaion of he Pecepon algoihm (Camme & Singe, 2001b), denoed MCP, and he Widow-Hoff algoihm (Widow & Hoff, 1960) fo online egession leaning, which we denoe by WH. Fo WH we fixed is leaning ae o a

25 Online Ranking by Pojecing 169 consan value. The hypoheses ha he vaians of PRank and he wo ohe algoihms mainain shae similaiies bu ae diffeen in hei complexiy: PRank mainains a veco w of dimension n and a veco of k 1 modifiable hesholds b, oaling n + k 1 paamees; MCP mainains k pooypes, which ae vecos of dimension n, yielding kn paamees; WH mainains a single veco w of size n. Theefoe, MCP builds he mos complex hypohesis, while WH builds he simples. We descibe wo ses of expeimens wih wo diffeen daa ses. The daa se used in he fis expeimen is synheic and was geneaed in a simila way o he daa se used by Hebich e al. (2000). We fis geneaed poins x = (x 1, x 2 ) unifomly a andom fom he uni squae [0, 1] 2. Each poin was assigned a ank y fom he se {1,...,5} accoding o he following anking ule, y = max { :10((x 1 0.5)(x 2 0.5)) + ξ>b } whee b = (, 1, 0.1, 0.25, 1) and ξ is a nomally disibued noise wih a zeo mean and a sandad deviaion of We geneaed sequences of insance ank pais, each of lengh We fed he sequences o PRank, Si- PRank, MCP, and WH. We hen obained pedicions fo each insance. We conveed he eal-valued pedicions of WH ino anks by ounding each pedicion o is closes ank value. As in Hebich e al. (2000), we used a nonhomogeneous polynomial of degee 2, K( x 1, x 2 ) = (( x 1 x 2 ) + 1) 2 as he inne poduc opeaion beween each inpu insance and he hypeplanes ha he fou addiive algoihms mainain. A each ime sep, we compued fo each algoihm he accumulaive anking loss nomalized by he insananeous sequence lengh. Fomally, he ime-aveaged loss afe T ounds is (1/T) T ŷ y. We compued he loss fo each algoihm we esed fo T = 1,...,8000. To incease he saisical significance of he esuls, we epeaed he pocess 100 imes, picking a new andom insance ank sequence of lengh 8000 each ime, and hen aveaged he insananeous losses acoss he 100 uns. The esuls ae depiced on he lef-hand side of Figue 6. The size of he symbols in he plo is lage han 95% confidence inevals fo each esul depiced in he figue. In his expeimen, he pefomance of MCP is consanly wose han he pefomance of WH and PRank. WH iniially suffes he smalles insananeous loss, bu afe abou 500 ounds, boh PRank and Si-PRank sa o oupefom MCP and WH. Evenually he anking loss ha PRank and Si-PRank suffe is significanly lowe han boh WH and MCP. In he second se of expeimens, we used he EachMovie daa se (Mc- Jones, 1997). This daa se is used fo collaboaive fileing asks and conains aings of movies povided by 61,265 people. Each peson in he daa se viewed a subse of movies fom a collecion of 1623 iles. Each viewe aed each movie ha she saw using one of six possible aings: 0, 0.2, 0.4, 0.6, 0.8, 1. We chose subses of people who viewed a significan numbe of movies, exacing fo evaluaion people who have aed a leas 100 movies. Thee wee 7542 such viewes. We chose a andom one pe-

26 170 K. Camme and Y. Singe PRank Si PRank WH MC Pecepon 0.9 Rank Loss Round Figue 6: Compaison of he ime-aveaged anking loss of PRank, Si-PRank, WH, and MCP on synheic daa. son among hese viewes and se he peson s aings o be he age ank. We used he aings of he ohe viewes as feaues. Thus, he goal is o lean o pedic he ase of a andom use wih he use s pas aings seving as a feedback and he aings of fellow viewes as feaues. The pedicion ule associaes a weigh wih each fellow viewe and heefoe can be seen as leaning coelaions beween he ases of diffeen viewes. Nex, we subaced 0.5 fom each aing; heefoe he possible aing values ae 0.5, 0.3, 0.1, 0.1, 0.3, 0.5. This linea ansfomaion enabled us o assign a value of zeo o movies ha have no been aed. We fed each pai of feaue and ank level in an online fashion. Since we chose viewes who aed a leas 100 movies, we wee able o pefom a leas 100 ounds of online pedicions and updaes. We epeaed his expeimen 500 imes, each ime choosing a andom viewe as he age ank. The esuls ae given in he lef-hand plos of Figue 7. The eo bas in he plo indicae 95% confidence levels. We epeaed he expeimen using viewes who have seen a leas 200 movies. (Thee wee 1802 such viewes.) The esuls of his expeimen ae given on he igh-hand plos of Figue 7. The plos a he op of he figue compae he addiive algoihms: PRank, Si-PRank, MCP, and WH. Along he enie un of he algoihms, PRank and Si-PRank ae significanly

27 Online Ranking by Pojecing PRank Si PRank WH MC Pecepon PRank Si PRank WH MC Pecepon Rank Loss Rank Loss Round Round Rank Loss Si PRank η 0.50 η 1.00 η 4.00 η Rank Loss Si PRank η 0.50 η 1.00 η 4.00 η Round Round Figue 7: Compaison of he ime-aveaged anking loss of he vaians of PRank, WH, and MCP on he EachMovie daa se using viewes who aed a leas 200 movies (op) and a leas 100 movies (boom). bee han WH and consisenly bee han he muliclass pecepon algoihm, alhough he las employs a hypohesis ha is subsanially bigge. Compaing he wo addiive vaians of PRank, we see ha he anking loss of Si-PRank is slighly lowe han ha of PRank. The plos a he boom of he figue compae he bes of he addiive algoihms, Si-PRank, wih he Mu-PRank un wih diffeen leaning aes (η = 0.5, 1, 4, 16). One of he goals of his expeimen is o check he dependency of he pefomance of Mu-PRank on is leaning ae. I is clea fom he wo boom plos of Figue 7 ha he pefomance of Mu-PRank is sensiive o he choice of he leaning ae. Noe, hough, ha he bes leaning ae is poblem dependen: he bes leaning ae is η = 4 fo he fis paiion of EachMovie, while η = 0.5 esuls in he smalles anking loss in he case of he second paiion. This ype of behavio is also exhibied by muliplicaive algoihms in ohe poblems such as binay classificaion (Kivinen & Wamuh, 1997). Noneheless, Mu-PRank oupefoms mos of he ohe algoihms we compaed fo a boad ange

28 172 K. Camme and Y. Singe of leaning aes. Focusing fis on he lef plo, we obseve ha afe abou 60 ounds, Si-PRank achieves he bes pefomance, while a he sa of he aining pocess, is anking loss is almos infeio o mos of he copies of Mu-PRank. Summing up, boh he addiive and muliplicaive vesions of PRank oupefom egession and classificaion algoihms when evaluaed wih espec o he anking loss. While he supeio pefomance of PRank is no supising, i povides an empiical validaion of he fomal analysis pesened in he pevious secions. 7 Conclusion In his aicle, we descibed a family of algoihms fo insance anking. The oos of he algoihms go back he pecepon algoihm (Rosenbla, 1958). One of he majo esuls in he aicle is he descipion of a new appoach fo solving anking poblems. While mos of he pevious appoaches educe he poblem of anking o a classificaion of pais, we pesened an alenaive appoach ha builds a connecion beween ank levels and subinevals of he eals. An open poblem ha aises fom boh he heoeical analysis and he empiical esuls is deciding wha vesion of PRank o use. In paicula, PRank and Si-PRank shae he same misake bound, while in ou expeimens, Si-PRank pefomed bee han PRank. All of he vesions of PRank pesened in his aicle ae online algoihms, and fo hei analysis we used he misake-bound model. An ineesing eseach diecion is he design and analysis of algoihms fo bach seings in which all of he aining examples ae given a once. As menioned in secion 1, Shashua and Levin (2002) have descibed a bach algoihm fo anking. An ineesing quesion is how he diffeen vaians elae o he algoihm of Shashua and Levin. In addiion o coninuing ou eseach on online algoihms fo anking poblems, an ineesing eseach diecion is he design and genealizaion analysis of bach algoihms fo vaious anking losses. Appendix: Technical Poofs A.1 Poof of Theoem 2. To pove he heoem, we inoduce a sligh modificaion of heoem 1 by elaxing he assumpion ha v is of a uni nom and adding he nom of v o he misake bound of he basic PRank algoihm. We hus wie he misake bound of heoem 1 as T =1 ŷ y (k 1) (R2 + 1) v 2 γ 2. Since he case D = 0 educes o seing of heoem 1, we can assume ha D > 0. We pove he heoem by ansfoming he insepaable poblem ino a sepaable one. We do so by expanding each oiginal insance x R n ino

29 Online Ranking by Pojecing 173 a veco z R n+t as follows. The fis n coodinaes of z ae se o x. The n+ coodinae of z is se o, which is a posiive eal numbe whose value is se below. The es of he coodinaes of z ae se o zeo. We similaly exend he anking ule ( w, b ) o (ū, c ) R (n+t) R k 1 as follows. We ewie he value of d fom equaion 3.11 as { } d = max γ min {( w x b y )y }, 0,γ min {( w x b <y )y }, (A.1) and define s o be he following indicao funcion: 1 d = γ min y {( w x b s )y } = 0 d = 0. (A.2) +1 d = γ min <y {( w x b )y } Noe ha if s = 1, hen d = ( w x b )y fo some y, and hus y = 1. Similaly, if s =+1, hen d = ( w x b )y fo some < y and y =+1. We se he fis n columns ū o be w, and he n + coodinae of ū is se o be s d and he es of he coodinaes of ū ae se o zeo. We now show ha (ū, c ) achieves a magin value γ on he expanded sequence. Using he definiion of s and d, we ge (ū z c )y = ( w x + s d Noe ha by consucion, z 2 R b )y = ( w x b )y + s d y = ( w x b )y + d Since he nom of ( w, b ) is 1, we have ( w x b )y + γ ( w x b )y = γ. (A.3) ū 2 + c 2 = w 2 + b 2 + ( s d ) 2 = 1 + D2 2. We now apply he bound of heoem 1 in he fom discussed above and ge ha T =1 ŷ y (k 1) (R )(1 + D2 2 ) γ 2. (A.4) Seing 2 = D R in he equaion above yields he desied bound.

30 174 K. Camme and Y. Singe I emains o show ha he pedicions of he algoihm on each elemen of he oiginal sequence and on he expanded sequences ae idenical. This pa follows exacly he same line of poof used in heoem 1 fom Feund and Schapie (1998) and is hus omied. Acknowledgmens Thanks o Sanjoy Dasgupa and Rob Schapie fo numeous discussions on anking poblems and algoihms. Thanks also o Eleaza Eskin and Ui Maoz fo caefully eading he manuscip. This wok was suppoed in pa by he IST Pogamme of he Euopean Communiy, unde he PASCAL Newok of Excellence, IST This publicaion eflecs only he auhos views. Refeences Cohen, W., Schapie, R. E., & Singe, Y. (1999). Leaning o ode hings. Jounal of Aificial Inelligence Reseach, 10, Cove, T. M., & Thomas, J. A. (1991). Elemens of infomaion heoy. New Yok: Wiley. Camme, K., & Singe, Y. (2001a). Panking wih anking. In T. G. Dieeich, S. Becke, & Z. Ghahamani (Eds.), Advances in neual infomaion pocessing sysems, 14. Cambidge, MA: MIT Pess. Camme, K., & Singe, Y. (2001b). Ulaconsevaive online algoihms fo muliclass poblems. In Poceedings of he Foueenh Annual Confeence on Compuaional Leaning Theoy. Amsedam: Spinge. Cisianini, N., & Shawe-Taylo, J. (2000). An inoducion o suppo veco machines. Cambidge: Cambidge Univesiy Pess. Fleche, R. (1987). Pacical mehods of opimizaion (2nd ed.). New Yok: Wiley. Feund, Y., Iye, R., Schapie, R. E., & Singe, Y. (2003). An efficien boosing algoihm fo combining pefeences. Jounal of Machine Leaning Reseach, 4, Feund, Y., & Schapie, R. E. (1999). Lage magin classificaion using he pecepon algoihm. Machine Leaning, 37, Haingon, E. F. (2003). Online anking/collaboaive fileing using he pecepon algoihm. In Poceedings of he Twenieh Inenaional Confeence on Machine Leaning. Washingon, DC: AIII Pess. Hebich, R., Gaepel, T., & Obemaye, K. (2000). Lage maging ank boundaies fo odinal egession. In A. Smola, B. Schölkopf, & D. Schuumans (Eds.), Advances in lage magin classifies. Cambidge, MA: MIT Pess. Kemeny, J. G., & Snell, J. L. (1962). Mahemaical models in he social sciences. Cambidge, MA: MIT Pess. Kivinen, J., & Wamuh, M. K. (1997). Exponeniaed gadien vesus gadien descen fo linea pedicos. Infomaion and Compuaion, 132(1), Lilesone, N. (1987). Leaning when ielevan aibues abound. In 28h Annual Symposium on Foundaions of Compue Science (pp ). Los Angeles: IEEE.

On Control Problem Described by Infinite System of First-Order Differential Equations

On Control Problem Described by Infinite System of First-Order Differential Equations Ausalian Jounal of Basic and Applied Sciences 5(): 736-74 ISS 99-878 On Conol Poblem Descibed by Infinie Sysem of Fis-Ode Diffeenial Equaions Gafujan Ibagimov and Abbas Badaaya J'afau Insiue fo Mahemaical

More information

Variance and Covariance Processes

Variance and Covariance Processes Vaiance and Covaiance Pocesses Pakash Balachandan Depamen of Mahemaics Duke Univesiy May 26, 2008 These noes ae based on Due s Sochasic Calculus, Revuz and Yo s Coninuous Maingales and Bownian Moion, Kaazas

More information

Lecture-V Stochastic Processes and the Basic Term-Structure Equation 1 Stochastic Processes Any variable whose value changes over time in an uncertain

Lecture-V Stochastic Processes and the Basic Term-Structure Equation 1 Stochastic Processes Any variable whose value changes over time in an uncertain Lecue-V Sochasic Pocesses and he Basic Tem-Sucue Equaion 1 Sochasic Pocesses Any vaiable whose value changes ove ime in an unceain way is called a Sochasic Pocess. Sochasic Pocesses can be classied as

More information

[ ] 0. = (2) = a q dimensional vector of observable instrumental variables that are in the information set m constituents of u

[ ] 0. = (2) = a q dimensional vector of observable instrumental variables that are in the information set m constituents of u Genealized Mehods of Momens he genealized mehod momens (GMM) appoach of Hansen (98) can be hough of a geneal pocedue fo esing economics and financial models. he GMM is especially appopiae fo models ha

More information

7 Wave Equation in Higher Dimensions

7 Wave Equation in Higher Dimensions 7 Wave Equaion in Highe Dimensions We now conside he iniial-value poblem fo he wave equaion in n dimensions, u c u x R n u(x, φ(x u (x, ψ(x whee u n i u x i x i. (7. 7. Mehod of Spheical Means Ref: Evans,

More information

Representing Knowledge. CS 188: Artificial Intelligence Fall Properties of BNs. Independence? Reachability (the Bayes Ball) Example

Representing Knowledge. CS 188: Artificial Intelligence Fall Properties of BNs. Independence? Reachability (the Bayes Ball) Example C 188: Aificial Inelligence Fall 2007 epesening Knowledge ecue 17: ayes Nes III 10/25/2007 an Klein UC ekeley Popeies of Ns Independence? ayes nes: pecify complex join disibuions using simple local condiional

More information

Combinatorial Approach to M/M/1 Queues. Using Hypergeometric Functions

Combinatorial Approach to M/M/1 Queues. Using Hypergeometric Functions Inenaional Mahemaical Foum, Vol 8, 03, no 0, 463-47 HIKARI Ld, wwwm-hikaicom Combinaoial Appoach o M/M/ Queues Using Hypegeomeic Funcions Jagdish Saan and Kamal Nain Depamen of Saisics, Univesiy of Delhi,

More information

Computer Propagation Analysis Tools

Computer Propagation Analysis Tools Compue Popagaion Analysis Tools. Compue Popagaion Analysis Tools Inoducion By now you ae pobably geing he idea ha pedicing eceived signal sengh is a eally impoan as in he design of a wieless communicaion

More information

Adaptive Regularization of Weight Vectors

Adaptive Regularization of Weight Vectors Adapive Regulaizaion of Weigh Vecos Koby Camme Depamen of Elecical Engineing he echnion Haifa, 32000 Isael koby@ee.echnion.ac.il Alex Kulesza Depamen of Compue and Infomaion Science Univesiy of Pennsylvania

More information

Support Vector Machines

Support Vector Machines Suppo Veco Machine CSL 3 ARIFICIAL INELLIGENCE SPRING 4 Suppo Veco Machine O, Kenel Machine Diciminan-baed mehod olean cla boundaie Suppo veco coni of eample cloe o bounday Kenel compue imilaiy beeen eample

More information

Low-complexity Algorithms for MIMO Multiplexing Systems

Low-complexity Algorithms for MIMO Multiplexing Systems Low-complexiy Algoihms fo MIMO Muliplexing Sysems Ouline Inoducion QRD-M M algoihm Algoihm I: : o educe he numbe of suviving pahs. Algoihm II: : o educe he numbe of candidaes fo each ansmied signal. :

More information

STUDY OF THE STRESS-STRENGTH RELIABILITY AMONG THE PARAMETERS OF GENERALIZED INVERSE WEIBULL DISTRIBUTION

STUDY OF THE STRESS-STRENGTH RELIABILITY AMONG THE PARAMETERS OF GENERALIZED INVERSE WEIBULL DISTRIBUTION Inenaional Jounal of Science, Technology & Managemen Volume No 04, Special Issue No. 0, Mach 205 ISSN (online): 2394-537 STUDY OF THE STRESS-STRENGTH RELIABILITY AMONG THE PARAMETERS OF GENERALIZED INVERSE

More information

Quantum Algorithms for Matrix Products over Semirings

Quantum Algorithms for Matrix Products over Semirings CHICAGO JOURNAL OF THEORETICAL COMPUTER SCIENCE 2017, Aicle 1, pages 1 25 hp://cjcscsuchicagoedu/ Quanum Algoihms fo Maix Poducs ove Semiings Fançois Le Gall Haumichi Nishimua Received July 24, 2015; Revised

More information

AN EVOLUTIONARY APPROACH FOR SOLVING DIFFERENTIAL EQUATIONS

AN EVOLUTIONARY APPROACH FOR SOLVING DIFFERENTIAL EQUATIONS AN EVOLUTIONARY APPROACH FOR SOLVING DIFFERENTIAL EQUATIONS M. KAMESWAR RAO AND K.P. RAVINDRAN Depamen of Mechanical Engineeing, Calicu Regional Engineeing College, Keala-67 6, INDIA. Absac:- We eploe

More information

Lecture 18: Kinetics of Phase Growth in a Two-component System: general kinetics analysis based on the dilute-solution approximation

Lecture 18: Kinetics of Phase Growth in a Two-component System: general kinetics analysis based on the dilute-solution approximation Lecue 8: Kineics of Phase Gowh in a Two-componen Sysem: geneal kineics analysis based on he dilue-soluion appoximaion Today s opics: In he las Lecues, we leaned hee diffeen ways o descibe he diffusion

More information

Sections 3.1 and 3.4 Exponential Functions (Growth and Decay)

Sections 3.1 and 3.4 Exponential Functions (Growth and Decay) Secions 3.1 and 3.4 Eponenial Funcions (Gowh and Decay) Chape 3. Secions 1 and 4 Page 1 of 5 Wha Would You Rahe Have... $1million, o double you money evey day fo 31 days saing wih 1cen? Day Cens Day Cens

More information

Extremal problems for t-partite and t-colorable hypergraphs

Extremal problems for t-partite and t-colorable hypergraphs Exemal poblems fo -paie and -coloable hypegaphs Dhuv Mubayi John Talbo June, 007 Absac Fix ineges and an -unifom hypegaph F. We pove ha he maximum numbe of edges in a -paie -unifom hypegaph on n veices

More information

An Automatic Door Sensor Using Image Processing

An Automatic Door Sensor Using Image Processing An Auomaic Doo Senso Using Image Pocessing Depamen o Elecical and Eleconic Engineeing Faculy o Engineeing Tooi Univesiy MENDEL 2004 -Insiue o Auomaion and Compue Science- in BRNO CZECH REPUBLIC 1. Inoducion

More information

Lecture 22 Electromagnetic Waves

Lecture 22 Electromagnetic Waves Lecue Elecomagneic Waves Pogam: 1. Enegy caied by he wave (Poyning veco).. Maxwell s equaions and Bounday condiions a inefaces. 3. Maeials boundaies: eflecion and efacion. Snell s Law. Quesions you should

More information

General Non-Arbitrage Model. I. Partial Differential Equation for Pricing A. Traded Underlying Security

General Non-Arbitrage Model. I. Partial Differential Equation for Pricing A. Traded Underlying Security 1 Geneal Non-Abiage Model I. Paial Diffeenial Equaion fo Picing A. aded Undelying Secuiy 1. Dynamics of he Asse Given by: a. ds = µ (S, )d + σ (S, )dz b. he asse can be eihe a sock, o a cuency, an index,

More information

Reinforcement learning

Reinforcement learning Lecue 3 Reinfocemen leaning Milos Hauskech milos@cs.pi.edu 539 Senno Squae Reinfocemen leaning We wan o lean he conol policy: : X A We see examples of x (bu oupus a ae no given) Insead of a we ge a feedback

More information

The sudden release of a large amount of energy E into a background fluid of density

The sudden release of a large amount of energy E into a background fluid of density 10 Poin explosion The sudden elease of a lage amoun of enegy E ino a backgound fluid of densiy ceaes a song explosion, chaaceized by a song shock wave (a blas wave ) emanaing fom he poin whee he enegy

More information

MEEN 617 Handout #11 MODAL ANALYSIS OF MDOF Systems with VISCOUS DAMPING

MEEN 617 Handout #11 MODAL ANALYSIS OF MDOF Systems with VISCOUS DAMPING MEEN 67 Handou # MODAL ANALYSIS OF MDOF Sysems wih VISCOS DAMPING ^ Symmeic Moion of a n-dof linea sysem is descibed by he second ode diffeenial equaions M+C+K=F whee () and F () ae n ows vecos of displacemens

More information

ENGI 4430 Advanced Calculus for Engineering Faculty of Engineering and Applied Science Problem Set 9 Solutions [Theorems of Gauss and Stokes]

ENGI 4430 Advanced Calculus for Engineering Faculty of Engineering and Applied Science Problem Set 9 Solutions [Theorems of Gauss and Stokes] ENGI 44 Avance alculus fo Engineeing Faculy of Engineeing an Applie cience Poblem e 9 oluions [Theoems of Gauss an okes]. A fla aea A is boune by he iangle whose veices ae he poins P(,, ), Q(,, ) an R(,,

More information

CS 188: Artificial Intelligence Fall Probabilistic Models

CS 188: Artificial Intelligence Fall Probabilistic Models CS 188: Aificial Inelligence Fall 2007 Lecue 15: Bayes Nes 10/18/2007 Dan Klein UC Bekeley Pobabilisic Models A pobabilisic model is a join disibuion ove a se of vaiables Given a join disibuion, we can

More information

Monochromatic Wave over One and Two Bars

Monochromatic Wave over One and Two Bars Applied Mahemaical Sciences, Vol. 8, 204, no. 6, 307-3025 HIKARI Ld, www.m-hikai.com hp://dx.doi.og/0.2988/ams.204.44245 Monochomaic Wave ove One and Two Bas L.H. Wiyano Faculy of Mahemaics and Naual Sciences,

More information

Lecture 17: Kinetics of Phase Growth in a Two-component System:

Lecture 17: Kinetics of Phase Growth in a Two-component System: Lecue 17: Kineics of Phase Gowh in a Two-componen Sysem: descipion of diffusion flux acoss he α/ ineface Today s opics Majo asks of oday s Lecue: how o deive he diffusion flux of aoms. Once an incipien

More information

Research on the Algorithm of Evaluating and Analyzing Stationary Operational Availability Based on Mission Requirement

Research on the Algorithm of Evaluating and Analyzing Stationary Operational Availability Based on Mission Requirement Reseach on he Algoihm of Evaluaing and Analyzing Saionay Opeaional Availabiliy Based on ission Requiemen Wang Naichao, Jia Zhiyu, Wang Yan, ao Yilan, Depamen of Sysem Engineeing of Engineeing Technology,

More information

Online Completion of Ill-conditioned Low-Rank Matrices

Online Completion of Ill-conditioned Low-Rank Matrices Online Compleion of Ill-condiioned Low-Rank Maices Ryan Kennedy and Camillo J. Taylo Compue and Infomaion Science Univesiy of Pennsylvania Philadelphia, PA, USA keny, cjaylo}@cis.upenn.edu Laua Balzano

More information

Orthotropic Materials

Orthotropic Materials Kapiel 2 Ohoopic Maeials 2. Elasic Sain maix Elasic sains ae elaed o sesses by Hooke's law, as saed below. The sesssain elaionship is in each maeial poin fomulaed in he local caesian coodinae sysem. ε

More information

, on the power of the transmitter P t fed to it, and on the distance R between the antenna and the observation point as. r r t

, on the power of the transmitter P t fed to it, and on the distance R between the antenna and the observation point as. r r t Lecue 6: Fiis Tansmission Equaion and Rada Range Equaion (Fiis equaion. Maximum ange of a wieless link. Rada coss secion. Rada equaion. Maximum ange of a ada. 1. Fiis ansmission equaion Fiis ansmission

More information

arxiv: v1 [math.co] 4 Apr 2019

arxiv: v1 [math.co] 4 Apr 2019 Dieced dominaion in oiened hypegaphs axiv:1904.02351v1 [mah.co] 4 Ap 2019 Yai Cao Dep. of Mahemaics Univesiy of Haifa-Oanim Tivon 36006, Isael yacao@kvgeva.og.il This pape is dedicaed o Luz Volkmann on

More information

Today - Lecture 13. Today s lecture continue with rotations, torque, Note that chapters 11, 12, 13 all involve rotations

Today - Lecture 13. Today s lecture continue with rotations, torque, Note that chapters 11, 12, 13 all involve rotations Today - Lecue 13 Today s lecue coninue wih oaions, oque, Noe ha chapes 11, 1, 13 all inole oaions slide 1 eiew Roaions Chapes 11 & 1 Viewed fom aboe (+z) Roaional, o angula elociy, gies angenial elociy

More information

On The Estimation of Two Missing Values in Randomized Complete Block Designs

On The Estimation of Two Missing Values in Randomized Complete Block Designs Mahemaical Theoy and Modeling ISSN 45804 (Pape ISSN 505 (Online Vol.6, No.7, 06 www.iise.og On The Esimaion of Two Missing Values in Randomized Complee Bloc Designs EFFANGA, EFFANGA OKON AND BASSE, E.

More information

Reichenbach and f-generated implications in fuzzy database relations

Reichenbach and f-generated implications in fuzzy database relations INTERNATIONAL JOURNAL O CIRCUITS SYSTEMS AND SIGNAL PROCESSING Volume 08 Reichenbach and f-geneaed implicaions in fuzzy daabase elaions Nedžad Dukić Dženan Gušić and Nemana Kajmoić Absac Applying a definiion

More information

336 ERIDANI kfk Lp = sup jf(y) ; f () jj j p p whee he supemum is aken ove all open balls = (a ) inr n, jj is he Lebesgue measue of in R n, () =(), f

336 ERIDANI kfk Lp = sup jf(y) ; f () jj j p p whee he supemum is aken ove all open balls = (a ) inr n, jj is he Lebesgue measue of in R n, () =(), f TAMKANG JOURNAL OF MATHEMATIS Volume 33, Numbe 4, Wine 2002 ON THE OUNDEDNESS OF A GENERALIED FRATIONAL INTEGRAL ON GENERALIED MORREY SPAES ERIDANI Absac. In his pape we exend Nakai's esul on he boundedness

More information

ON 3-DIMENSIONAL CONTACT METRIC MANIFOLDS

ON 3-DIMENSIONAL CONTACT METRIC MANIFOLDS Mem. Fac. Inegaed As and Sci., Hioshima Univ., Se. IV, Vol. 8 9-33, Dec. 00 ON 3-DIMENSIONAL CONTACT METRIC MANIFOLDS YOSHIO AGAOKA *, BYUNG HAK KIM ** AND JIN HYUK CHOI ** *Depamen of Mahemaics, Faculy

More information

The shortest path between two truths in the real domain passes through the complex domain. J. Hadamard

The shortest path between two truths in the real domain passes through the complex domain. J. Hadamard Complex Analysis R.G. Halbud R.Halbud@ucl.ac.uk Depamen of Mahemaics Univesiy College London 202 The shoes pah beween wo uhs in he eal domain passes hough he complex domain. J. Hadamad Chape The fis fundamenal

More information

Distribution Free Evolvability of Polynomial Functions over all Convex Loss Functions

Distribution Free Evolvability of Polynomial Functions over all Convex Loss Functions Disibuion Fee Evolvabiliy of Polynomial Funcions ove all Convex Loss Funcions Paul Valian UC Beeley Beeley, Califonia pvalian@gmail.com ABSTRACT We fomulae a noion of evolvabiliy fo funcions wih domain

More information

Fig. 1S. The antenna construction: (a) main geometrical parameters, (b) the wire support pillar and (c) the console link between wire and coaxial

Fig. 1S. The antenna construction: (a) main geometrical parameters, (b) the wire support pillar and (c) the console link between wire and coaxial a b c Fig. S. The anenna consucion: (a) ain geoeical paaees, (b) he wie suppo pilla and (c) he console link beween wie and coaial pobe. Fig. S. The anenna coss-secion in he y-z plane. Accoding o [], he

More information

@FMI c Kyung Moon Sa Co.

@FMI c Kyung Moon Sa Co. Annals of Fuzzy Mahemaics Infomaics Volume, No. 2, (Apil 20), pp. 9-3 ISSN 2093 930 hp://www.afmi.o.k @FMI c Kyung Moon Sa Co. hp://www.kyungmoon.com On lacunay saisical convegence in inuiionisic fuzzy

More information

KINEMATICS OF RIGID BODIES

KINEMATICS OF RIGID BODIES KINEMTICS OF RIGID ODIES In igid body kinemaics, we use he elaionships govening he displacemen, velociy and acceleaion, bu mus also accoun fo he oaional moion of he body. Descipion of he moion of igid

More information

156 There are 9 books stacked on a shelf. The thickness of each book is either 1 inch or 2

156 There are 9 books stacked on a shelf. The thickness of each book is either 1 inch or 2 156 Thee ae 9 books sacked on a shelf. The hickness of each book is eihe 1 inch o 2 F inches. The heigh of he sack of 9 books is 14 inches. Which sysem of equaions can be used o deemine x, he numbe of

More information

Risk tolerance and optimal portfolio choice

Risk tolerance and optimal portfolio choice Risk oleance and opimal pofolio choice Maek Musiela BNP Paibas London Copoae and Invesmen Join wok wih T. Zaiphopoulou (UT usin) Invesmens and fowad uiliies Pepin 6 Backwad and fowad dynamic uiliies and

More information

Kalman Filter: an instance of Bayes Filter. Kalman Filter: an instance of Bayes Filter. Kalman Filter. Linear dynamics with Gaussian noise

Kalman Filter: an instance of Bayes Filter. Kalman Filter: an instance of Bayes Filter. Kalman Filter. Linear dynamics with Gaussian noise COM47 Inoducion o Roboics and Inelligen ysems he alman File alman File: an insance of Bayes File alman File: an insance of Bayes File Linea dynamics wih Gaussian noise alman File Linea dynamics wih Gaussian

More information

Fuzzy Hv-submodules in Γ-Hv-modules Arvind Kumar Sinha 1, Manoj Kumar Dewangan 2 Department of Mathematics NIT Raipur, Chhattisgarh, India

Fuzzy Hv-submodules in Γ-Hv-modules Arvind Kumar Sinha 1, Manoj Kumar Dewangan 2 Department of Mathematics NIT Raipur, Chhattisgarh, India Inenaional Jounal of Engineeing Reseach (IJOER) [Vol-1, Issue-1, Apil.- 2015] Fuzz v-submodules in Γ-v-modules Avind Kuma Sinha 1, Manoj Kuma Dewangan 2 Depamen of Mahemaics NIT Raipu, Chhaisgah, India

More information

Research Article A Note on Multiplication and Composition Operators in Lorentz Spaces

Research Article A Note on Multiplication and Composition Operators in Lorentz Spaces Hindawi Publishing Copoaion Jounal of Funcion Spaces and Applicaions Volume 22, Aicle ID 29363, pages doi:.55/22/29363 Reseach Aicle A Noe on Muliplicaion and Composiion Opeaos in Loenz Spaces Eddy Kwessi,

More information

r P + '% 2 r v(r) End pressures P 1 (high) and P 2 (low) P 1 , which must be independent of z, so # dz dz = P 2 " P 1 = " #P L L,

r P + '% 2 r v(r) End pressures P 1 (high) and P 2 (low) P 1 , which must be independent of z, so # dz dz = P 2  P 1 =  #P L L, Lecue 36 Pipe Flow and Low-eynolds numbe hydodynamics 36.1 eading fo Lecues 34-35: PKT Chape 12. Will y fo Monday?: new daa shee and daf fomula shee fo final exam. Ou saing poin fo hydodynamics ae wo equaions:

More information

Deviation probability bounds for fractional martingales and related remarks

Deviation probability bounds for fractional martingales and related remarks Deviaion pobabiliy bounds fo facional maingales and elaed emaks Buno Sausseeau Laboaoie de Mahémaiques de Besançon CNRS, UMR 6623 16 Roue de Gay 253 Besançon cedex, Fance Absac In his pape we pove exponenial

More information

A Negative Log Likelihood Function-Based Nonlinear Neural Network Approach

A Negative Log Likelihood Function-Based Nonlinear Neural Network Approach A Negaive Log Likelihood Funcion-Based Nonlinea Neual Newok Appoach Ponip Dechpichai,* and Pamela Davy School of Mahemaics and Applied Saisics Univesiy of Wollongong, Wollongong NSW 5, AUSTRALIA * Coesponding

More information

International Journal of Pure and Applied Sciences and Technology

International Journal of Pure and Applied Sciences and Technology In. J. Pue Appl. Sci. Technol., 4 (211, pp. 23-29 Inenaional Jounal of Pue and Applied Sciences and Technology ISS 2229-617 Available online a www.ijopaasa.in eseach Pape Opizaion of he Uiliy of a Sucual

More information

The Production of Polarization

The Production of Polarization Physics 36: Waves Lecue 13 3/31/211 The Poducion of Polaizaion Today we will alk abou he poducion of polaized ligh. We aleady inoduced he concep of he polaizaion of ligh, a ansvese EM wave. To biefly eview

More information

Probabilistic Models. CS 188: Artificial Intelligence Fall Independence. Example: Independence. Example: Independence? Conditional Independence

Probabilistic Models. CS 188: Artificial Intelligence Fall Independence. Example: Independence. Example: Independence? Conditional Independence C 188: Aificial Inelligence Fall 2007 obabilisic Models A pobabilisic model is a join disibuion ove a se of vaiables Lecue 15: Bayes Nes 10/18/2007 Given a join disibuion, we can eason abou unobseved vaiables

More information

New sufficient conditions of robust recovery for low-rank matrices

New sufficient conditions of robust recovery for low-rank matrices New sufficien condiions of ous ecovey fo low-an maices Jianwen Huang a, Jianjun Wang a,, Feng Zhang a, Wendong Wang a a School of Mahemaics and Saisics, Souhwes Univesiy, Chongqing, 400715, China Reseach

More information

Pressure Vessels Thin and Thick-Walled Stress Analysis

Pressure Vessels Thin and Thick-Walled Stress Analysis Pessue Vessels Thin and Thick-Walled Sess Analysis y James Doane, PhD, PE Conens 1.0 Couse Oveview... 3.0 Thin-Walled Pessue Vessels... 3.1 Inoducion... 3. Sesses in Cylindical Conaines... 4..1 Hoop Sess...

More information

THE MODULAR INEQUALITIES FOR A CLASS OF CONVOLUTION OPERATORS ON MONOTONE FUNCTIONS

THE MODULAR INEQUALITIES FOR A CLASS OF CONVOLUTION OPERATORS ON MONOTONE FUNCTIONS PROCEEDINGS OF THE AMERICAN MATHEMATICAL SOCIETY Volume 5, Numbe 8, Augus 997, Pages 93 35 S -9939(973867-7 THE MODULAR INEQUALITIES FOR A CLASS OF CONVOLUTION OPERATORS ON MONOTONE FUNCTIONS JIM QILE

More information

1 Review of Zero-Sum Games

1 Review of Zero-Sum Games COS 5: heoreical Machine Learning Lecurer: Rob Schapire Lecure #23 Scribe: Eugene Brevdo April 30, 2008 Review of Zero-Sum Games Las ime we inroduced a mahemaical model for wo player zero-sum games. Any

More information

WORK POWER AND ENERGY Consevaive foce a) A foce is said o be consevaive if he wok done by i is independen of pah followed by he body b) Wok done by a consevaive foce fo a closed pah is zeo c) Wok done

More information

Unsupervised Segmentation of Moving MPEG Blocks Based on Classification of Temporal Information

Unsupervised Segmentation of Moving MPEG Blocks Based on Classification of Temporal Information Unsupevised Segmenaion of Moving MPEG Blocs Based on Classificaion of Tempoal Infomaion Ofe Mille 1, Ami Avebuch 1, and Yosi Kelle 2 1 School of Compue Science,Tel-Aviv Univesiy, Tel-Aviv 69978, Isael

More information

Chapter 2. First Order Scalar Equations

Chapter 2. First Order Scalar Equations Chaper. Firs Order Scalar Equaions We sar our sudy of differenial equaions in he same way he pioneers in his field did. We show paricular echniques o solve paricular ypes of firs order differenial equaions.

More information

MATHEMATICAL FOUNDATIONS FOR APPROXIMATING PARTICLE BEHAVIOUR AT RADIUS OF THE PLANCK LENGTH

MATHEMATICAL FOUNDATIONS FOR APPROXIMATING PARTICLE BEHAVIOUR AT RADIUS OF THE PLANCK LENGTH Fundamenal Jounal of Mahemaical Phsics Vol 3 Issue 013 Pages 55-6 Published online a hp://wwwfdincom/ MATHEMATICAL FOUNDATIONS FOR APPROXIMATING PARTICLE BEHAVIOUR AT RADIUS OF THE PLANCK LENGTH Univesias

More information

Circular Motion. Radians. One revolution is equivalent to which is also equivalent to 2π radians. Therefore we can.

Circular Motion. Radians. One revolution is equivalent to which is also equivalent to 2π radians. Therefore we can. 1 Cicula Moion Radians One evoluion is equivalen o 360 0 which is also equivalen o 2π adians. Theefoe we can say ha 360 = 2π adians, 180 = π adians, 90 = π 2 adians. Hence 1 adian = 360 2π Convesions Rule

More information

EFFECT OF PERMISSIBLE DELAY ON TWO-WAREHOUSE INVENTORY MODEL FOR DETERIORATING ITEMS WITH SHORTAGES

EFFECT OF PERMISSIBLE DELAY ON TWO-WAREHOUSE INVENTORY MODEL FOR DETERIORATING ITEMS WITH SHORTAGES Volume, ssue 3, Mach 03 SSN 39-4847 EFFEC OF PERMSSBLE DELAY ON WO-WAREHOUSE NVENORY MODEL FOR DEERORANG EMS WH SHORAGES D. Ajay Singh Yadav, Ms. Anupam Swami Assisan Pofesso, Depamen of Mahemaics, SRM

More information

Two-dimensional Effects on the CSR Interaction Forces for an Energy-Chirped Bunch. Rui Li, J. Bisognano, R. Legg, and R. Bosch

Two-dimensional Effects on the CSR Interaction Forces for an Energy-Chirped Bunch. Rui Li, J. Bisognano, R. Legg, and R. Bosch Two-dimensional Effecs on he CS Ineacion Foces fo an Enegy-Chiped Bunch ui Li, J. Bisognano,. Legg, and. Bosch Ouline 1. Inoducion 2. Pevious 1D and 2D esuls fo Effecive CS Foce 3. Bunch Disibuion Vaiaion

More information

A Weighted Moving Average Process for Forecasting. Shou Hsing Shih Chris P. Tsokos

A Weighted Moving Average Process for Forecasting. Shou Hsing Shih Chris P. Tsokos A Weighed Moving Aveage Pocess fo Foecasing Shou Hsing Shih Chis P. Tsokos Depamen of Mahemaics and Saisics Univesiy of Souh Floida, USA Absac The objec of he pesen sudy is o popose a foecasing model fo

More information

P h y s i c s F a c t s h e e t

P h y s i c s F a c t s h e e t P h y s i c s F a c s h e e Sepembe 2001 Numbe 20 Simple Hamonic Moion Basic Conceps This Facshee will:! eplain wha is mean by simple hamonic moion! eplain how o use he equaions fo simple hamonic moion!

More information

On the Semi-Discrete Davey-Stewartson System with Self-Consistent Sources

On the Semi-Discrete Davey-Stewartson System with Self-Consistent Sources Jounal of Applied Mahemaics and Physics 25 3 478-487 Published Online May 25 in SciRes. hp://www.scip.og/jounal/jamp hp://dx.doi.og/.4236/jamp.25.356 On he Semi-Discee Davey-Sewason Sysem wih Self-Consisen

More information

New and Faster Filters for. Multiple Approximate String Matching. University of Chile. Blanco Encalada Santiago - Chile

New and Faster Filters for. Multiple Approximate String Matching. University of Chile. Blanco Encalada Santiago - Chile New and Fase Files fo Muliple Appoximae Sing Maching Ricado Baeza-Yaes Gonzalo Navao Depamen of Compue Science Univesiy of Chile Blanco Encalada 22 - Saniago - Chile fbaeza,gnavaog@dcc.uchile.cl Absac

More information

Chapter Finite Difference Method for Ordinary Differential Equations

Chapter Finite Difference Method for Ordinary Differential Equations Chape 8.7 Finie Diffeence Mehod fo Odinay Diffeenial Eqaions Afe eading his chape, yo shold be able o. Undesand wha he finie diffeence mehod is and how o se i o solve poblems. Wha is he finie diffeence

More information

GRADIENT ESTIMATES, POINCARÉ INEQUALITIES, DE GIORGI PROPERTY AND THEIR CONSEQUENCES

GRADIENT ESTIMATES, POINCARÉ INEQUALITIES, DE GIORGI PROPERTY AND THEIR CONSEQUENCES GRADIENT ESTIMATES, POINCARÉ INEQUALITIES, DE GIORGI PROPERTY AND THEIR CONSEQUENCES FRÉDÉRIC ERNICOT, THIERRY COULHON, DOROTHEE FREY Absac. On a doubling meic measue space endowed wih a caé du champ,

More information

Dual Hierarchies of a Multi-Component Camassa Holm System

Dual Hierarchies of a Multi-Component Camassa Holm System Commun. heo. Phys. 64 05 37 378 Vol. 64, No. 4, Ocobe, 05 Dual Hieachies of a Muli-Componen Camassa Holm Sysem LI Hong-Min, LI Yu-Qi, and CHEN Yong Shanghai Key Laboaoy of uswohy Compuing, Eas China Nomal

More information

INTRODUCTION TO MACHINE LEARNING 3RD EDITION

INTRODUCTION TO MACHINE LEARNING 3RD EDITION ETHEM ALPAYDIN The MIT Press, 2014 Lecure Slides for INTRODUCTION TO MACHINE LEARNING 3RD EDITION alpaydin@boun.edu.r hp://www.cmpe.boun.edu.r/~ehem/i2ml3e CHAPTER 2: SUPERVISED LEARNING Learning a Class

More information

Secure Frameproof Codes Through Biclique Covers

Secure Frameproof Codes Through Biclique Covers Discee Mahemaics and Theoeical Compue Science DMTCS vol. 4:2, 202, 26 270 Secue Famepoof Codes Though Biclique Coves Hossein Hajiabolhassan,2 and Faokhlagha Moazami 3 Depamen of Mahemaical Sciences, Shahid

More information

Section 3.5 Nonhomogeneous Equations; Method of Undetermined Coefficients

Section 3.5 Nonhomogeneous Equations; Method of Undetermined Coefficients Secion 3.5 Nonhomogeneous Equaions; Mehod of Undeermined Coefficiens Key Terms/Ideas: Linear Differenial operaor Nonlinear operaor Second order homogeneous DE Second order nonhomogeneous DE Soluion o homogeneous

More information

ÖRNEK 1: THE LINEAR IMPULSE-MOMENTUM RELATION Calculate the linear momentum of a particle of mass m=10 kg which has a. kg m s

ÖRNEK 1: THE LINEAR IMPULSE-MOMENTUM RELATION Calculate the linear momentum of a particle of mass m=10 kg which has a. kg m s MÜHENDİSLİK MEKANİĞİ. HAFTA İMPULS- MMENTUM-ÇARPIŞMA Linea oenu of a paicle: The sybol L denoes he linea oenu and is defined as he ass ies he elociy of a paicle. L ÖRNEK : THE LINEAR IMPULSE-MMENTUM RELATIN

More information

Online Convex Optimization Example And Follow-The-Leader

Online Convex Optimization Example And Follow-The-Leader CSE599s, Spring 2014, Online Learning Lecure 2-04/03/2014 Online Convex Opimizaion Example And Follow-The-Leader Lecurer: Brendan McMahan Scribe: Sephen Joe Jonany 1 Review of Online Convex Opimizaion

More information

Lecture 20: Riccati Equations and Least Squares Feedback Control

Lecture 20: Riccati Equations and Least Squares Feedback Control 34-5 LINEAR SYSTEMS Lecure : Riccai Equaions and Leas Squares Feedback Conrol 5.6.4 Sae Feedback via Riccai Equaions A recursive approach in generaing he marix-valued funcion W ( ) equaion for i for he

More information

Course Notes for EE227C (Spring 2018): Convex Optimization and Approximation

Course Notes for EE227C (Spring 2018): Convex Optimization and Approximation Course Noes for EE7C Spring 018: Convex Opimizaion and Approximaion Insrucor: Moriz Hard Email: hard+ee7c@berkeley.edu Graduae Insrucor: Max Simchowiz Email: msimchow+ee7c@berkeley.edu Ocober 15, 018 3

More information

Degree of Approximation of a Class of Function by (C, 1) (E, q) Means of Fourier Series

Degree of Approximation of a Class of Function by (C, 1) (E, q) Means of Fourier Series IAENG Inenaional Jounal of Applied Mahemaic, 4:, IJAM_4 7 Degee of Appoximaion of a Cla of Funcion by C, E, q Mean of Fouie Seie Hae Kihna Nigam and Kuum Shama Abac In hi pape, fo he fi ime, we inoduce

More information

Measures the linear dependence or the correlation between r t and r t-p. (summarizes serial dependence)

Measures the linear dependence or the correlation between r t and r t-p. (summarizes serial dependence) . Definiions Saionay Time Seies- A ime seies is saionay if he popeies of he pocess such as he mean and vaiance ae consan houghou ime. i. If he auocoelaion dies ou quickly he seies should be consideed saionay

More information

Order statistics and concentration of l r norms for log-concave vectors

Order statistics and concentration of l r norms for log-concave vectors Jounal of Funcional Analysis 61 011 681 696 www.elsevie.com/locae/jfa Ode saisics and concenaion of l noms fo log-concave vecos Rafał Laała a,b, a Insiue of Mahemaics, Univesiy of Wasaw, Banacha, 0-097

More information

Exponential and Logarithmic Equations and Properties of Logarithms. Properties. Properties. log. Exponential. Logarithmic.

Exponential and Logarithmic Equations and Properties of Logarithms. Properties. Properties. log. Exponential. Logarithmic. Eponenial and Logaihmic Equaions and Popeies of Logaihms Popeies Eponenial a a s = a +s a /a s = a -s (a ) s = a s a b = (ab) Logaihmic log s = log + logs log/s = log - logs log s = s log log a b = loga

More information

Numerical solution of fuzzy differential equations by Milne s predictor-corrector method and the dependency problem

Numerical solution of fuzzy differential equations by Milne s predictor-corrector method and the dependency problem Applied Maemaics and Sciences: An Inenaional Jounal (MaSJ ) Vol. No. Augus 04 Numeical soluion o uzz dieenial equaions b Milne s pedico-coeco meod and e dependenc poblem Kanagaajan K Indakuma S Muukuma

More information

MECHANICS OF MATERIALS Poisson s Ratio

MECHANICS OF MATERIALS Poisson s Ratio Fouh diion MCHANICS OF MATRIALS Poisson s Raio Bee Johnson DeWolf Fo a slende ba subjeced o aial loading: 0 The elongaion in he -diecion is accompanied b a conacion in he ohe diecions. Assuming ha he maeial

More information

Notes for Lecture 17-18

Notes for Lecture 17-18 U.C. Berkeley CS278: Compuaional Complexiy Handou N7-8 Professor Luca Trevisan April 3-8, 2008 Noes for Lecure 7-8 In hese wo lecures we prove he firs half of he PCP Theorem, he Amplificaion Lemma, up

More information

Chapter 7. Interference

Chapter 7. Interference Chape 7 Inefeence Pa I Geneal Consideaions Pinciple of Supeposiion Pinciple of Supeposiion When wo o moe opical waves mee in he same locaion, hey follow supeposiion pinciple Mos opical sensos deec opical

More information

K. G. Malyutin, T. I. Malyutina, I. I. Kozlova ON SUBHARMONIC FUNCTIONS IN THE HALF-PLANE OF INFINITE ORDER WITH RADIALLY DISTRIBUTED MEASURE

K. G. Malyutin, T. I. Malyutina, I. I. Kozlova ON SUBHARMONIC FUNCTIONS IN THE HALF-PLANE OF INFINITE ORDER WITH RADIALLY DISTRIBUTED MEASURE Математичнi Студiї. Т.4, 2 Maemaychni Sudii. V.4, No.2 УДК 57.5 K. G. Malyuin, T. I. Malyuina, I. I. Kozlova ON SUBHARMONIC FUNCTIONS IN THE HALF-PLANE OF INFINITE ORDER WITH RADIALLY DISTRIBUTED MEASURE

More information

An Open cycle and Closed cycle Gas Turbine Engines. Methods to improve the performance of simple gas turbine plants

An Open cycle and Closed cycle Gas Turbine Engines. Methods to improve the performance of simple gas turbine plants An Open cycle and losed cycle Gas ubine Engines Mehods o impove he pefomance of simple gas ubine plans I egeneaive Gas ubine ycle: he empeaue of he exhaus gases in a simple gas ubine is highe han he empeaue

More information

Dynamic Estimation of OD Matrices for Freeways and Arterials

Dynamic Estimation of OD Matrices for Freeways and Arterials Novembe 2007 Final Repo: ITS Dynamic Esimaion of OD Maices fo Feeways and Aeials Auhos: Juan Calos Heea, Sauabh Amin, Alexande Bayen, Same Madana, Michael Zhang, Yu Nie, Zhen Qian, Yingyan Lou, Yafeng

More information

Vehicle Arrival Models : Headway

Vehicle Arrival Models : Headway Chaper 12 Vehicle Arrival Models : Headway 12.1 Inroducion Modelling arrival of vehicle a secion of road is an imporan sep in raffic flow modelling. I has imporan applicaion in raffic flow simulaion where

More information

Appendix to Online l 1 -Dictionary Learning with Application to Novel Document Detection

Appendix to Online l 1 -Dictionary Learning with Application to Novel Document Detection Appendix o Online l -Dicionary Learning wih Applicaion o Novel Documen Deecion Shiva Prasad Kasiviswanahan Huahua Wang Arindam Banerjee Prem Melville A Background abou ADMM In his secion, we give a brief

More information

T L. t=1. Proof of Lemma 1. Using the marginal cost accounting in Equation(4) and standard arguments. t )+Π RB. t )+K 1(Q RB

T L. t=1. Proof of Lemma 1. Using the marginal cost accounting in Equation(4) and standard arguments. t )+Π RB. t )+K 1(Q RB Elecronic Companion EC.1. Proofs of Technical Lemmas and Theorems LEMMA 1. Le C(RB) be he oal cos incurred by he RB policy. Then we have, T L E[C(RB)] 3 E[Z RB ]. (EC.1) Proof of Lemma 1. Using he marginal

More information

10. State Space Methods

10. State Space Methods . Sae Space Mehods. Inroducion Sae space modelling was briefly inroduced in chaper. Here more coverage is provided of sae space mehods before some of heir uses in conrol sysem design are covered in he

More information

Simulation-Solving Dynamic Models ABE 5646 Week 2, Spring 2010

Simulation-Solving Dynamic Models ABE 5646 Week 2, Spring 2010 Simulaion-Solving Dynamic Models ABE 5646 Week 2, Spring 2010 Week Descripion Reading Maerial 2 Compuer Simulaion of Dynamic Models Finie Difference, coninuous saes, discree ime Simple Mehods Euler Trapezoid

More information

On Energy-Efficient Node Deployment in Wireless Sesnor Networks

On Energy-Efficient Node Deployment in Wireless Sesnor Networks I J Communicaions, Newok and Sysem Sciences, 008, 3, 07-83 Published Online Augus 008 in Scies (hp://wwwscipog/jounal/ijcns/) On Enegy-Efficien Node Deploymen in Wieless Sesno Newoks Hui WANG 1, KeZhong

More information

arxiv: v2 [stat.me] 13 Jul 2015

arxiv: v2 [stat.me] 13 Jul 2015 One- and wo-sample nonpaameic ess fo he al-o-noise aio based on ecod saisics axiv:1502.05367v2 [sa.me] 13 Jul 2015 Damien Challe 1,2 1 Laboaoie de mahémaiques appliquées aux sysèmes, CenaleSupélec, 92295

More information

Learning a Class from Examples. Training set X. Class C 1. Class C of a family car. Output: Input representation: x 1 : price, x 2 : engine power

Learning a Class from Examples. Training set X. Class C 1. Class C of a family car. Output: Input representation: x 1 : price, x 2 : engine power Alpaydin Chaper, Michell Chaper 7 Alpaydin slides are in urquoise. Ehem Alpaydin, copyrigh: The MIT Press, 010. alpaydin@boun.edu.r hp://www.cmpe.boun.edu.r/ ehem/imle All oher slides are based on Michell.

More information

Chapter 7: Solving Trig Equations

Chapter 7: Solving Trig Equations Haberman MTH Secion I: The Trigonomeric Funcions Chaper 7: Solving Trig Equaions Le s sar by solving a couple of equaions ha involve he sine funcion EXAMPLE a: Solve he equaion sin( ) The inverse funcions

More information

PHYS PRACTICE EXAM 2

PHYS PRACTICE EXAM 2 PHYS 1800 PRACTICE EXAM Pa I Muliple Choice Quesions [ ps each] Diecions: Cicle he one alenaive ha bes complees he saemen o answes he quesion. Unless ohewise saed, assume ideal condiions (no ai esisance,

More information

d 1 = c 1 b 2 - b 1 c 2 d 2 = c 1 b 3 - b 1 c 3

d 1 = c 1 b 2 - b 1 c 2 d 2 = c 1 b 3 - b 1 c 3 and d = c b - b c c d = c b - b c c This process is coninued unil he nh row has been compleed. The complee array of coefficiens is riangular. Noe ha in developing he array an enire row may be divided or

More information