Communication-Efficient Distributed Primal-Dual Algorithm for Saddle Point Problems

Size: px
Start display at page:

Download "Communication-Efficient Distributed Primal-Dual Algorithm for Saddle Point Problems"

Transcription

1 Commucato-Effcet Dstrbuted Prmal-Dual Algorthm for Saddle Pot Problems Yaodog Yu Nayag Techologcal Uversty Sul Lu Nayag Techologcal Uversty So Jal Pa Nayag Techologcal Uversty Abstract Prmal-dual algorthms, whch are proposed to solve reformulated covex-cocave saddle pot problems, have bee prove to be effectve for solvg a geerc class of covex optmzato problems, especally whe the problems are ll-codtoed. However, the saddle pot problem stll lacs a dstrbuted optmzato framewor where prmal-dual algorthms ca be employed. I ths paper, we propose a ovel commucato-effcet dstrbuted optmzato framewor to solve the covex-cocave saddle pot problem based o prmal-dual methods. We carefully desg local subproblems ad a cetral problem such that our proposed dstrbuted optmzato framewor s commucato-effcet. We provde a covergece aalyss of our proposed algorthm, ad exted t to address o-smooth ad o-strogly covex loss fuctos. We coduct extesve expermets o several real-world datasets to demostrate compettve performace of the proposed method, especally o ll-codtoed problems. INTRODUCTION I the era of bg data, developg dstrbuted mache learg algorthms has become creasgly mportat yet challegg. I ths wor, we focus o developg a ew dstrbuted optmzato algorthm for regularzed emprcal rs mmzato ERM, whch s a geerc class of covex optmzato problems that arses ofte from mache learg. Specfcally, our goal s to m- Idcates equal cotrbutos. mze the emprcal loss defed over data samples: m x R d Px = φ a x gx, = where a,..., a R d are feature vectors of data pots, φ : R R s a covex loss fucto wth the lear predctor a x, for =,...,, ad g : Rd R s a covex regularzato fucto for the predctor x R d. Suppose that a dstrbuted system cossts of K maches, ad each mache has access oly a subset P of the data [] := {,..., }, where {P } K = s a gve partto of the dataset [], ad we deote by = P.. COMMUNICATION-EFFICIENT DISTRIBUTED OPTIMIZATION Oe of the most mportat ssues dstrbuted optmzato s the commucato effcecy because commucato betwee maches s much more expesve tha readg data from the memory o local maches. Therefore, more ad more efforts have bee made proposg commucato-effcet methods dstrbuted optmzato Jagg et al., 04, Ma et al., 05, Redd et al., 06, Shamr et al., 04, Smth et al., 05, Yag, 03. The basc dea behd these methods s to carefully desg local computato ad commucato a dstrbuted system. To acheve ths goal, exstg methods usually decompose the optmzato problem to K local subproblems, deoted by L, wth respect to local data of each mache. After each local mache performs a arbtrary optmzato method o L ad solves t approxmately, updated formato from local maches s set to a cetral ode to carry out a cetral update. Ths allows oe to cotrol the trade-off betwee commucato ad local computato, whch s more flexble dstrbuted settg. I geeral, exstg commucato-effcet methods ca be classfed to two categores. A frst category

2 s referred to as gradet-type, whch ams to decompose the problem to K local subproblems ad solve each subproblem usg gradet descet methods, such as stochastc varace reduced gradet SVRG Johso ad Zhag, 03, o each local mache depedetly, ad the update the optmzato varables by dog a average of local varables of each mache Redd et al., 06, Shamr et al., 04. The secod oe s referred to as coordate-type, whch usually focuses o solvg the dual problem of, max α R Dα = φ α g = α a, = where φ s the cojugate fucto of φ, α s the dual varable vector wth the -th elemet beg α. Ths category of methods decomposes the dual problem to K local subproblems wth regard to each mache, ad employs a coordate-type method, such as stochastc dual coordate ascet method SDCA Shalev-Shwartz ad Zhag, 03, to solve each local subproblem. After each local subproblem s approxmately solved, the cetral ode taes a average/add step accordg to the local update from each mache Jagg et al., 04, Ma et al., 05, Yag, 03. Our proposed method falls to ths category. There are two lmtatos of the coordate-type approach. Oe lmtato s that most exstg coordatetype methods fal to match the commucato complexty lower bouds proved Arjeva ad Shamr, 05. For example, whe φ s /γ-smooth, g s λ-strogly covex, ad R = max a problem, the the codto umber s defed as κ = R /λγ, 3 ad the lower boud of commucato complexty obtaed Arjeva ad Shamr, 05 s Õ κ log/ɛ order to acheve a ɛ-suboptmal soluto Nesterov, 03. Sce these coordatetype methods Jagg et al., 04, Ma et al., 05, Yag, 03 could be see as a dstrbuted verso of SDCA, ad the terato complexty of SDCA s Õ κ log/ɛ, the commucato complexty of these methods Jagg et al., 04, Ma et al., 05, Yag, 03 s Õ κ log/ɛ, whch does ot match the optmal commucato complexty. The problem could get worse whe problem s ll-codtoed,.e., κ. The other lmtato s that t s ot straght-forward to exted exstg coordate-type approaches Jagg et al., 04, Ma et al., 05, Yag, 03 to deal wth dfferet regularzato forms of the regularzato term g other tha l regularzato.. OUR APPROACH To overcome the lmtatos metoed above, we propose a ovel commucato-effcet dstrbuted framewor wth a prmal-dual algorthm for saddle pot problems Chambolle ad Poc, 0, Zhag ad Xao, 05b. The reasos that we focus o developg dstrbuted framewor for saddle pot problems are three folds:. It has bee show that saddle pot algorthms are able to obta comparable ad eve more stable performace tha other state-of-the-art techques for covex optmzato Chambolle ad Poc, 0, Yu et al., 05, Zhag ad Xao, 05b.. Prmal-dual coordate algorthms for saddle pot problems are able to reach the optmal terato complexty as proved by Zhag ad Xao 05b. 3. Sce prmal-dual algorthms eep both prmal ad dual varables optmzato, they are able to deal wth dfferet ds of regularzato terms g, such as l -regularzato, aturally. Therefore, we am to develop a dstrbuted prmal-dual algorthm whch ca hert the above three propertes a dstrbuted evromet. To be specfc, we frst reformulate problem as a covex-cocave saddle pot problem through covex cojugato, whch eeps both prmal ad dual varables optmzato. We the decompose the saddle pot problem to carefully desged local subproblems, whch ca be solved depedetly by each mache wth respect to ts local data. After the local subproblems are solved approxmately, the updated local dual varables of each mache are set to a cetral ode. Based o all the updates of local dual varables, we costruct a cetral problem ad get the cetral update of the prmal varables o the cetral ode. Fally, the cetral ode seds the updated prmal varables ad aggregated dual varables to each local mache for ext local optmzato terato. Detals wll be descrbed Secto 3. As wll be dscussed Secto 4 ad Secto 5, by carefully desgg the local subproblems ad the cetral problem, our algorthm s able to reach the commucato complexty lower bouds, ad deal wth dfferet ds of regularzato terms. As the parameter λ of the regularzato fucto g s usually o the order of / or / for may mache learg problems, we are especally terested solvg problem uder the dstrbuted settg whe t s ll-codtoed. Wth a extremely large-scale dataset, the codto umber 3 could be relatvely large κ. We wll

3 show through expermets Secto 6 that our proposed algorthm ca obta better ad more stable performace o ll-codtoed problems compared to other commucato-effcet dstrbuted optmzato methods. Moreover, we provde a soluto to exted our algorthm to deal wth o-smooth ad o-strogly covex optmzato problems Secto 5..3 OTHER RELATED WORK Besdes the commucato-effcet dstrbuted optmzato methods revewed above, there exst other parallel ad dstrbuted optmzato techques. For example, a well-ow approach for solvg problem s to perform a gradet descet method mplemeted a dstrbuted system. Each local mache computes ts local gradet ad seds t to the cetral ode. The cetral ode aggregates the local gradet to tae a gradet step ad updates x, ad the broadcasts t bac to each local mache for the ext terato updates. If the accelerated gradet descet method Nesterov, 03 s used, oe ca obta the terato complexty Õ κ log/ɛ. Aother popular techque s dstrbuted alteratg drecto method of multplers ADMM Boyd et al., 0, Sh et al., 04, whose complexty s Õ κ log/ɛ uder certa codtos. Recetly, Zhag ad Xao 05a proposed a dstrbuted algorthm based o the exact damped Newto method, whch matches the commucato complexty lower boud Õ κ log/ɛ. For more related dstrbuted ad parallel optmzato methods, we refer the readers to Arjeva ad Shamr, 05, Chag et al., 04, Duch et al., 03, Zhag et al., 0 for a more comprehesve lterature of the related wor. PRELIMINARIES I ths secto, we troduce frst-order prmal-dual algorthms for covex-cocave saddle pot problems a sgle mache. To beg wth, we reformulate the prmal problem as a covex-cocave saddle pot problem through covex cojugato, ad the saddle pot reformulato s wdely appled mache learg Da et al., 06, Zhag ad Xao, 05b. Based o the defto of covex cojugate, we replace each compoet fucto φ a x by φ a x = sup {y a, x φ y }, y R where φ y = sup α R {αy φ α} s the covex cojugate of φ, the we arrve at a covex-cocave saddle pot problem m max x R d y R fx, y, 4 where fx, y def = y a, x φ y gx, = where y R s referred to as a vector of dual varables wth ts -th elemet deoted by y, ad each y s assocated wth a data pot a. Uder the assumpto that φ s /γ-smooth, g s λ-strogly covex, ad each φ s γ-strogly covex, the saddle pot problem 4 has a uque soluto, whch s deoted by x, y. Based o the reformulated problem 4, we could rewrte the optmzato problem as m max fx, y = x R d y R Ay, x Φ y gx, where A = [a,..., a ] ad Φ y = = φ y. Ths s the geerc saddle pot problem Chambolle ad Poc, 0. To solve ths optmzato problem, the basc dea s to alteratgly maxmze f wth repect to y, ad mmze f wth respect to x, whch s y t = arg max y R Ay, xt Φ y y yt, σ x t = arg m x R Ayt, x gx x xt, d x t = x t θ x t x t, where the parameters τ ad σ cotrol the quadratc regularzato terms wth respect to x ad y, respectvely, whch s smlar to the use of step sze prmal methods. Ad x t s the extrapolato from x t ad x t wth parameter θ [0, ]. Here, θ x t x t s smlar to the mometum term Nesterov accelerato Nesterov, 03. There have bee varous prmal-dual coordate algorthms desged based o the above scheme, such as Chambolle ad Poc, 0, Yu et al., 05, Zhag ad Xao, 05b. Amog them, Yu et al. 05, Zhag ad Xao 05b developed stochastc versos of the prmal-dual algorthm for the saddle pot problem. Suppose φ s /γ-smooth ad g s λ-strogly covex, most exstg algorthms Chambolle ad Poc, 0, Yu et al., 05, Zhag ad Xao, 05b ca acheve lear covergece rate to obta a ɛ-accurate soluto. The expected terato complexty of these methods s Õ κ log/ɛ, whch s desred o llcodtoed problems. 3 DISTRIBUTED SADDLE-POINT FRAMEWORK I ths secto, we preset our proposed commucatoeffcet Dstrbuted Saddle Pot Algorthm DSPA

4 detal. Assume that the dataset {a } = s dstrbuted over K local maches, each mache has access to a subset P of the data wth sze of. Here we defe K local vectors of y by usg the otato y [] R for =,..., K: { y, f P, y [] = 0, otherwse, ad K local data matrx of A by usg the otato A [] R d for =,..., K: { a, f P, A [] = 0, otherwse, where A [] deotes the -th colum of A [], ad 0 s a d-dmesoal vector of all zeros. Accordg to the data partto, we decompose the problem 4 to K local subproblems, ad defe the assocated cetral problem. The core dea s to solve the defed local subproblems o each local mache depedetly, ad the cetralze the updated formato of each mache to solve a easy cetral problem. 3. LOCAL SUBPROBLEMS We defe a local subproblem of the orgal saddle pot problem 4 for mache, whch oly requres accessg data that s avalable locally,.e., P. Specfcally, each local subproblem of mache at t-terato s defed as m max x, y [], 5 where L t x, y [] := ad x R d y [] R Lt yj a j, x φ j y j gx j P y t / P r t x, y [] = x xt a, x r t x, y [], [] K y [] y t σ Each local subproblem s related to the soluto obtaed the prevous terato x t, y t, ad we wll defe the parameters τ, σ later. The frst two terms ca be regarded as a local verso of the orgal saddle pot problem 4. The thrd term / P y t a, x = l A [l]y t [l], x 5 ca be terpreted as the teracto wth other local subproblems. Ad the quadratc regularzato term r t x, y [] s used to eforce x ad y [] ot to move too far away from soluto obtaed at the prevous terato.. To solve the local subproblem L t x, y [], we could apply prmal-dual coordate type of methods Yu et al., 05, Zhag ad Xao, 05b. The put eeded for solvg the local subproblem would be x t ad Ay t, as the term l A [l]y t [l] ca be expressed as Ay t A [] y t [] ad mache has access to A [], y t []. By deotg our local optmzato algorthm used as Local-DSPA, we could deote the procedure of solvg the local subproblem as x t, yt [] Local-DSPAx t, Ay t. Note that our framewor, each local subproblem s oly requred to be solved approxmately. x t, yt [] 3. CENTRAL PROBLEM Let deote the approxmate soluto obtaed by solvg L t x, y [] o mache. Let y t =, we defe a cetral problem at t-terato as K = yt [] where C t x := x t = arg m C t x, 6 x R d = y t a, x gx x xt. The cetral update volves two steps. Frstly, updated local A [] y t [] R d are set to the cetral ode. After that, we solve 6 o the cetral ode, ad the sed the updated x t R d ad Ay t R d of the t-terato bac to each local mache. The cetral update could be vewed as mmzg the orgal saddle pot problem 4 over x wth updated y t. Also, we should observe that the parameter τ defed 5 ad 6 s the same parameter, whch s oe of the ey pots of our framewor. 3.3 OVERALL ALGORITHM DESCRIPTION The overall algorthm of DSPA s preseted Algorthm. Frstly, we dstrbute the tal dual varable y 0 to K local maches, ad aggregate A [] y 0 [] from all maches to compute Ay 0 K = A []y 0 []. After that, the tal prmal varables x 0 ad the trasformed dual varables Ay 0 are set to each mache Steps - 4. Secodly, each mache performs optmzato o ts ow local subproblem L t x, y [], usg a prmaldual algorthm, such as SPDC Zhag ad Xao, 05b, ad seds ts local update A [] y t [] R d to the cetral

5 Algorthm DSPAf, x 0, y 0, τ, σ, T : Iput: Data pots {a } = dstrbuted across K maches {P } K =, parameters τ, σ R, tal prmal varables x 0 R d ad tal dual varables y 0 R. : Dstrbute y 0 to K maches as y 0 [] for each. 3: Each mache compute A [] y 0 [], ad sed t bac to the cetral ode. 4: Perform Ay 0 K = A []y 0 [] o the cetral ode, ad sed x 0 ad Ay 0 to each mache 5: for t =,,..., T do 6: for [K] parallel over all maches do 7: Local-DSPAx t, Ay t x t, yt [] 8: Sed A [] y t [] to the cetral ode 9: ed for 0: Ay t K = A []y t [] o the cetral ode C t x o the cetral ode : x t arg m x R d : Sed x t ad Ay t bac to each mache 3: ed for 4: Output: x T, y T ode. The goal of ths teral procedure s to approxmately solve each local subproblem. We refer ths procedure as the er loop, whch s stated Steps 6-9 of Algorthm. Thrdly, wth the local updates A [] y t [] s, the cetral ode aggregates them to compute Ay t, ad solves the defed cetral problem C t x to get the update of x t Steps 0- of Algorthm. Fally, the cetral ode seds x t ad Ay t bac to each mache order to start the ext roud of local optmzato Step, we refer Steps 5-3 as the outer loop of our algorthm. Note that, o oe had, the essetal formato that the cetral { problem} requres ca be represeted by the vectors A [] y t [] s of each local mache, whch are of dmeso of d. O the other had, each local mache oly requres x t, Ay t, whch s of d dmesos, from the cetral ode. 4 CONVERGENCE ANALYSIS Before we preset our ma covergece results, we mae some assumptos o the objectve fuctos ad the local suboptmalty o each local subproblem. Assumpto. Each φ s covex ad dfferetable, ad φ s /γ-smooth,.e., φ a φ b /γ a b, a, b R. Assumpto. g s λ-strogly covex.e. for x, y R d ad g y gy, gx gy g yx y λ x y. Assumpto 3. There exsts costats Ω x, Ω y > 0 such that max x t x Ω x, t max y t y Ω y, t where x t, y t s t-terato update of Algorthm, ad x, y s the optmal soluto of 4. Assumpto 4. Each local mache produces a approxmate soluto that satsfes ad [ y t E E[fˆx t, ŷt [] ŷt [] x t [] fˆxt Θ where ˆΘ, Θ <, ad, yt [] ] ˆΘ, yt [] ] fˆx t y t [], ŷt ˆx t, ŷt [] x, y[]. ŷ t, [], yt [], [] fˆxt s the optmal soluto of local subproblem L t Assumpto 5. There exst costats c, c > 0, ad the approxmate soluto y t = K = yt [] satsfes E[ ] Θfx t, y fx, y t, ad s defed as = K fˆx t =, ŷt [] fˆxt, yt [] M yt ŷ t, where M = c Ω x c Ω y, ad Θ 0, s a pre-defed costat. Note that Assumptos 4 ad 5 mply that the local subproblems eed to be solved approxmately to some extet. As the local subproblems are saddle pot problems, we ca apply dfferet ds of algorthms for saddle pot problems to solve them, such as Chambolle ad Poc, 0, Yu et al., 05, Zhag ad Xao, 05b, whch clude both stochastc ad o-stochastc optmzato methods. To aalyse the covergece behavor of our dstrbuted algorthm, we eed to characterze the coecto betwee local subproblems ad cetral update o the cetral ode. Based o the relatoshp betwee the cetral update x t ad the local optmal soluto ˆx t, ŷt [] we ca get the covergece guaratee of Algorthm. All proofs ca be foud Appedx.,

6 Lemma. Suppose that Assumptos -5 hold. Let x, y be the uque saddle pot of fx, y defed 4, ad defe t =fx t, y fx, y t τλ x t x K Kσγ y t y. 4σK If the parameters τ ad σ are chose such that τσ = 4R ad σ > Kγ, the for t, the proposed DSPA Algorthm acheves E[ t ] ΘE[ t ], 7 where Θ = max{ σγ K K, τλ }. Theorem. Suppose that Assumptos -5 hold, ad the parameters τ = γ/4r, σ = /γ ad K κ. I order for Algorthm to obta E[ x T x ] ɛ, t suffces to have the umber of commucato teratos T satsfy T 4R C log, λγ ɛ where C = 0 λ 4R λγ. Based o Lemma ad Theorem, we ca get the commucato complexty of Algorthm. By tag τ = γ 4R ad σ = γ, the commucato complexty of Algorthm s Õ κ log/ɛ. Ths does ot match the optmal commucato complexty. I the ext secto, we derve a accelerated verso of DSPA that attas the optmal commucato complexty. 5 EXTENSION OF DISPA I ths secto, we derve two extesos of DSPA. The frst exteso apples a geerc accelerato scheme proposed by L et al. 05 to obta a accelerated verso of DSPA. The secod oe exteds our algorthm to hadle o-smooth ad o-strogly covex loss fuctos wth covergece guaratee. 5. ACCELERATION The commucato complexty of our algorthm derved Theorem s Õ κ log/ɛ order to acheve a ɛ-suboptmal soluto, whch does ot match the lower boud proved Arjeva ad Shamr, 05. Based o the Catalyst accelerato scheme proposed by L et al. 05, we develop the accelerated DSPA to acheve the optmal commucato complexty. Catalyst s a geerc scheme for acceleratg frst-order optmzato methods, smlar to classcal gradet descet schemes of Nesterov accelerato. It s a atural choce for dstrbuted optmzato algorthms that solve local subproblem approxmately each commucato terato. Based o Catalyst accelerato, we modfy the orgal objectve fucto by addg a quadratc term f t x, y = fx, y ϑ x zt, where ϑ > 0 s a parameter defed Catalyst, z t s obtaed by a extrapolato step smlar to the accelerato scheme Nesterov, 03. Wth a carefully selected parameter ϑ, we ca develop a accelerated verso of DSPA descrbed Algorthm, whch acheves the optmal commucato complexty. We refer to the accelerated dstrbuted saddle pot algorthm as A-DSPA. Theorem. Suppose Assumpto -5 hold ad assume ϑ = 4R γ λ > 0, f the parameters Catalyst are chose as T s = Õ, q = λ/λ ϑ, α 0 = q. The total commucato teratos of Algorthm for achevg Px T Px < ɛ s Õ κlog ɛ. The Catalyst accelerato scheme s appled wdely to accelerate dfferet ds of optmzato methods Redd et al., 06. As the accelerato quadratc term s related to the prmal varable x ad the saddle pot problem eeps the prmal varable, t s coveet to apply ths accelerato scheme to DSPA ad acheve the optmal commucato complexty. 5. NON-SMOOTH OR/AND NON-STRONGLY CONVEX FUNCTIONS The commucato complexty bouds Theorem ad Theorem are developed uder the Assumptos ad, whch meas that the dervatve of φ eeds to be /γ-lpschtz cotuous ad g eeds to be λ-strogly covex. For geeral loss fuctos mache learg, both assumptos may fal, for example, whe φ s hge loss ad g s l -regularzato term. Therefore, we am to develop a exteso of Algorthm to deal wth o-smooth ad o-strogly covex loss fuctos. To be cocse, we oly cosder the case whe φ s o-smooth ad g s o-strogly covex. The ey dea here s to apply a perturbato term to aalyze the o-smooth ad o-strogly covex settg, whch s commoly used may optmzato methods Redd et al., 06, Zhag ad Xao, 05b. Here we assume that ether φ s smooth or g s strogly covex.

7 Algorthm A-DSPA : Iput: Data pots {a } = dstrbuted across K maches {P } K =, parameters τ, σ, ϑ, α 0 R, tal prmal varable x 0 geerated o the cetral ode, tal local dual varables y 0 geerated o local maches, optmzato algorthm DSPA, umber of er teratos T s of DSPA. : Italzato: q = λ/λ ϑ, z 0 = x 0, t = 3: whle stoppg crtero s ot satsfed do 4: x t, y t = DSPAf t, x t, y t, τ, σ, T s 5: Compute α t 0, from equato α t = α t α t qα t 6: Compute z t = x t β t x t x t, where β t = αt αt α αt 7: t = t 8: ed whle 9: Output x t I partcular, we cosder the followg modfed saddle pot fucto f ɛ x, y def = = y a, x φ y ɛ y gx ɛ x, 8 where ɛ > 0 s a pre-defed scalar. f ɛ x, y ca be regarded as a approxmato of the orgal saddle pot problem fx, y. We the obta that both φ y ɛ y ad gx ɛ x are ɛ-strogly covex. I the followg corollary, we show that the perturbed fucto f ɛ x, y s a good approxmato of the org fucto fx, y. Corollary. Assume that φ s covex ad L φ -Lpschtz cotuous, ad g s covex. Let x, y be the uque saddle pot of fx, y, ad x ɛ, y ɛ be the uque saddle pot of f ɛ x, y. The we have Px ɛ Px ɛ x L φ. As a result, we could apply Algorthm to the perturbed fucto f ɛ x, y. By applyg Theorem ad the relatoshp betwee f ɛ x, y ad fx, y, we obta that the commucato complexty of Algorthm for o-smooth ad o-strogly covex fuctos s Õ /ɛ log ɛ, ad the proof s smlar to the oe Zhag ad Xao, 05b. 6 NUMERICAL EXPERIMENTS I ths secto, we coduct umercal expermets of DSPA ad A-DSPA o several real-world dstrbuted datasets. We compare the performace of our algorthms wth the state-of-the-art dstrbuted optmzato algorthm CoCoA Ma et al., 05. For our expermets, we solve the stadard bary classfcato tas wth datasets obtaed from LIBSVM datasets Chag ad L, 0. The statstcs of the datasets are summarzed Table. Our goal s to mmze the regularzed emprcal rs wth smoothed hge loss: m Px = φ a x R d x λ x. = For each tas, the data pot taes the form of a, b, where a s a feature vector, b {, } s the correspodg class label ad assocated wth loss fucto φ. The smoothed hge loss fucto Shalev-Shwartz ad Zhag, 03 φ s defed as 0 f b z, φ z = b z γ f b z γ, γ b z otherwse, where φ s /γ-smooth. We set γ = our expermets. The cojugate fucto of φ s φ β = b β γ β for b β [, 0] ad otherwse. Table : Three Datasets for Numercal Expermets DATASET # SAMPLES # FEATURES d RCV 677,399 47,36 Realsm 7,309 0,958 Covtype 58, IMPLEMENTATION DETAILS We mplemet DSPA, A-DSPA ad CoCoA Petuum Xg et al., 05. For DSPA ad A-DSPA, we use SPDC Zhag ad Xao, 05b as the local solver, ad use SDCA Shalev-Shwartz ad Zhag, 03 as the local solver for CoCoA. Whe dog comparso amog dfferet algorthms, we ru the same umber of local teratos o local maches.e., teratos of SDCA or SPDC before commucato wth the cetral ode. We compare dfferet algorthms based o commucato teratos. From our results, we fd that DSPA coverges faster whe parameters τ ad σ are large, whch s smlar to the fdgs reported Johso ad Zhag, 03, Zhag ad Xao, 05a. We adopt

8 Px - Px * 0 0 CoCoA DSPA A-DSPA Px - Px * 0 - CoCoA DSPA A-DSPA Px - Px * 0 - CoCoA DSPA A-DSPA Number of Commucatos a RCV, λ = Number of Commucatos b RCV, λ = Number of Commucatos c RCV, λ = 0 8 Px - Px * 0 0 CoCoA DSPA A-DSPA Px - Px * 0 - CoCoA DSPA A-DSPA Px - Px * 0 - CoCoA DSPA A-DSPA Number of Commucatos d Realsm, λ = Number of Commucatos e Realsm, λ = Number of Commucatos f Realsm, λ = 0 8 Px - Px * 0 0 CoCoA DSPA A-DSPA Px - Px * CoCoA 0 - DSPA A-DSPA Px - Px * CoCoA 0 - DSPA A-DSPA Number of Commucatos g Covtype, λ = Number of Commucatos h Covtype, λ = Number of Commucatos Covtype, λ = 0 8 Fgure : Comparso of DSPA, A-DSPA ad CoCoA o three datasets: RCV, Realsm, Covtype. The horzotal axs s the umber of commucatos, ad the vertcal axs s the optmalty gap P x T P x. Each plot cotas a comparso of CoCoA blue, crcle, DSPA red, asters ad A-DSPA yellow, damod. All the plots are show o log-y scale. The umber of local teratos performed each commucato terato s 5,000 for Realsm, ad 0,000 teratos for RCV ad Covtype. tued parameters τ ad σ from a predefed rage. For CoCoA, we preset the results of the selected optmal parameter σ. 6. COMPARISON BETWEEN DISPA AND COCOA We compare DSPA ad CoCoA o three datasets: RCV, Realsm, Covtype, across dfferet values of the regularzato parameter λ. To be specfc, the regularzato parameter λ s set to be λ { 0 6, 0 7, 0 8} o the three datasets. Both DSPA ad CoCoA are mplemeted o 6 maches K = 6. For both DSPA ad CoCoA, the umbers of local teratos performed each commucato terato are the same. I Fgure, the results show that the performace of DSPA s comparable wth CoCoA. We ca see that whe λ = 0 6, DSPA has comparable covergece performace wth regard to CoCoA. O all the three datasets, whe λ { 0 7, 0 8}, DSPA coverges to the optmal soluto faster ad more stable tha Co- CoA. We also otce that the performace of Co-

9 Px - Px * 0 0 K=4 K=8 K=6 K=3 K= Px - Px * 0 - K=4 K=8 K=6 K=3 K= Number of Commucatos a Realsm, λ = Number of Commucatos b RCV, λ = 0 7 Fgure : The performace of DSPA wth creasg umber of maches, K {4, 8, 6, 3, 64}. The local teratos performed each commucato terato s,000 for Realsm, ad 5,000 teratos for RCV. CoA o Covtype s ot stable compared to that o the other two datasets, especally whe λ s very small,.e., λ = 0 7 or λ = 0 8, whch s cosstet wth the expermetal results of CoCoA o Covtype Redd et al. 06. I Secto 6.3, we wll show the superor performace of A-DSPA o ll-codtoed problems. 6.3 COMPARISON OF DISPA AND A-DISPA I Fgure, we also compare A-DSPA ad DSPA to show the accelerato obtaed by applyg Catalyst accelerato. We compare DSPA ad A-DSPA o three datasets RCV, Realsm, Covtype across dfferet regularzato parameter uder the same settgs as metoed Secto 6.. The results show that whe problem s ll-codtoed,.e. κ = R / λγ, A- DSPA coverges substatally faster tha DSPA. Ths cofrms our theoretcal aalyss, as the commucato complexty of A-DSPA s Õ κlog ɛ ad DSPA s Õ κlog ɛ. 6.4 SCALABILITY Sce scalablty s a mportat metrc of dstrbuted optmzato algorthms, Fgure, we study the scalablty of DSPA by observg the umercal performace wth creasg umber of maches. We coduct expermets o two datasets RCV, Realsm, ad the mache umber s set to be K {4, 8, 6, 3, 64}. For comparso, we eep the teratos ad parameters τ, σ the same for each dataset. The expermets show that DSPA ca scale effectvely wth umber of maches, whch cofrms our theory Secto 4. We observe that performace slghtly drops o RCV wth 64 maches, ths s possbly because that local subproblems are solved wth hgher accuracy whch may affect the effectveess of aggregato the cetral update. 7 CONCLUSION I ths wor, we preset a ovel dstrbuted optmzato framewor for solvg the saddle pot problem, whch ca be appled to solvg a geerc class of covex optmzato problems. We provde the theoretcal guaratee of our algorthms, ad show that the accelerated algorthm ca acheve the optmal commucato complexty. We also exted our algorthm for solvg o-smooth ad o-strogly covex loss fuctos. The expermetal results demostrate that our algorthms obta better performace compared to the state-of-the-art dstrbuted optmzato method. Acowledgemets Ths wor s supported by NTU Sgapore Nayag Assstat Professorshp NAP grat M , Sgapore MOE AcRF Ter- grat MOE06-T--060, ad MOE AcRF Ter- grat 06-T Refereces Y. Arjeva ad O. Shamr. Commucato complexty of dstrbuted covex learg ad optmzato. I Advaces Neural Iformato Processg Systems, pages , 05. S. Boyd, N. Parh, E. Chu, B. Peleato, ad J. Ecste. Dstrbuted optmzato ad statstcal learg va the alteratg drecto method of multplers.

10 Foudatos ad Treds R Mache Learg, 3:, 0. A. Chambolle ad T. Poc. A frst-order prmal-dual algorthm for covex problems wth applcatos to magg. Joural of Mathematcal Imagg ad Vso, 40:0 45, 0. C.-C. Chag ad C.-J. L. LIBSVM: A lbrary for support vector maches. ACM Trasactos o Itellget Systems ad Techology, :7: 7:7, 0. Software avalable at edu.tw/ cjl/lbsvm. T.-H. Chag, A. Nedc, ad A. Scagloe. Dstrbuted costraed optmzato by cosesus-based prmaldual perturbato method. IEEE Trasactos o Automatc Cotrol, 596:54 538, 04. B. Da, N. He, Y. Pa, B. Boots, ad L. Sog. Learg from codtoal dstrbutos va dual erel embeddgs. arxv preprt arxv: , 06. J. Duch, M. I. Jorda, ad B. McMaha. Estmato, optmzato, ad parallelsm whe data s sparse. I Advaces Neural Iformato Processg Systems, pages , 03. M. Jagg, V. Smth, M. Taác, J. Terhorst, S. Krsha, T. Hofma, ad M. I. Jorda. Commucatoeffcet dstrbuted dual coordate ascet. I Advaces Neural Iformato Processg Systems, pages , 04. R. Johso ad T. Zhag. Acceleratg stochastc gradet descet usg predctve varace reducto. I Advaces Neural Iformato Processg Systems, pages 35 33, 03. H. L, J. Maral, ad Z. Harchaou. A uversal catalyst for frst-order optmzato. I Advaces Neural Iformato Processg Systems, pages , 05. C. Ma, V. Smth, M. Jagg, M. Jorda, P. Rchtar, ad M. Taac. Addg vs. averagg dstrbuted prmaldual optmzato. I Proceedgs of The 3d Iteratoal Coferece o Mache Learg, pages , 05. Y. Nesterov. Itroductory lectures o covex optmzato: A basc course, volume 87. Sprger Scece & Busess Meda, 03. S. J. Redd, J. Koečỳ, P. Rchtár, B. Póczós, ad A. Smola. Ade: Fast ad commucato effcet dstrbuted optmzato. arxv: , 06. arxv preprt S. Shalev-Shwartz ad T. Zhag. Stochastc dual coordate ascet methods for regularzed loss mmzato. Joural of Mache Learg Research, 4Feb: , 03. O. Shamr, N. Srebro, ad T. Zhag. Commucatoeffcet dstrbuted optmzato usg a approxmate ewto-type method. I Proceedgs of The 3st Iteratoal Coferece o Mache Learg, pages , 04. W. Sh, Q. Lg, K. Yua, G. Wu, ad W. Y. O the lear covergece of the admm decetralzed cosesus optmzato. IEEE Trasactos o Sgal Processg, 67:750 76, 04. V. Smth, S. Forte, M. I. Jorda, ad M. Jagg. L-regularzed dstrbuted optmzato: A commucato-effcet prmal-dual framewor. arxv preprt arxv:5.040, 05. E. P. Xg, Q. Ho, W. Da, J. K. Km, J. We, S. Lee, X. Zheg, P. Xe, A. Kumar, ad Y. Yu. Petuum: A ew platform for dstrbuted mache learg o bg data. IEEE Trasactos o Bg Data, :49 67, 05. T. Yag. Tradg computato for commucato: Dstrbuted stochastc dual coordate ascet. I Advaces Neural Iformato Processg Systems, pages , 03. A. W. Yu, Q. L, ad T. Yag. Doubly stochastc prmal-dual coordate method for regularzed emprcal rs mmzato wth factorzed data. CoRR, abs/ , 05. Y. Zhag ad L. Xao. Dsco: Dstrbuted optmzato for self-cocordat emprcal loss. I Proceedgs of the 3d Iteratoal Coferece o Mache Learg, pages , 05a. Y. Zhag ad L. Xao. Stochastc prmal-dual coordate method for regularzed emprcal rs mmzato. I Proceedgs of The 3d Iteratoal Coferece o Mache Learg, pages , 05b. Y. Zhag, M. J. Wawrght, ad J. C. Duch. Commucato-effcet algorthms for statstcal optmzato. I Advaces Neural Iformato Processg Systems, pages 50 50, 0.

11 Appedx A Lemmas for Covergece Aalyss of DSPA We frst troduce the lemma that characterzes the optmalty codto of local subproblem L t x, y []. Lemma. Assume that each φ s /γ-smooth ad g s λ-strogly covex, R = max{ a,... a }. Let ˆx t, ŷt [] be the optmal soluto of Lt L t x, ŷt x, y [], =,,..., K. Based o the strogly covexty, we have [] Lt ˆxt, ŷt [] λ x ˆx t, x R d L t ˆxt, y [] L t K ˆxt, ŷt [] σ γ y [] ŷ t [], y [] R Proof. Based o the defto of the saddle pot, we ca otce that L t ˆxt, y [] s a K σ γ -strogly covex fucto ad mmzed by ŷ t [], whch mples L t Also otce that L t ˆxt, y [] L t ˆxt, ŷ t K [] σ γ x, ŷt L t ˆxt, ŷt [] λ y [] ŷ t [], y [] R [] s a τ λ -strogly covex fucto mmzed by ˆx t, whch mples x, ŷt [] Lt x ˆx t, x R d Based o the optmalty codto of L t x, y [] ad the cetral update x t o cetral worer, we ca get the coecto betwee local subproblems ad cetral update o cetral worer. Lemma 3 Relatoshp betwee local optmal soluto ad cetral update. Assume that each φ s /γ-smooth ad g s λ-strogly covex, R = max{ a,... a }. Let ˆx t, ŷt [] be the optmal soluto of Lt x, y [], =,,..., K. Let x t = arg m C t x, t holds that x R d where Λ = K K 3 4σK yt y t Λ 4τ λ x t ˆx t 9 = K = ŷt [] yt [] A x t ˆx t ad /τ = 4σR. Proof. Based o lemma, we could get that for =,,..., K, ŷ t j j P ˆx t xt a j, x t ˆx t λ y t / P x t ˆx t a, x t ˆx t Sce x t mmzes the fucto C t x, whch meas that for =,,..., K, we have = y t a, ˆx t gxt gˆx t xt x t xt gˆx t gxt ˆxt xt xt x t λ x t ˆx t

12 Sum up the above two equaltes, we ca get ŷ t j j P a j, x t ˆx t j P ŷ t j y t j y t / P a j, x t ˆx t a, x t ˆx t / P ŷt [] yt [] A x t ˆx t l y t y t [l] = y t y t a, ˆx t xt τ λ x t ˆx t a, x t ˆx t τ λ x t ˆx t y t [l] A x t ˆx t τ λ x t ˆx t We eed to upper boud the secod term o the left-had-sde of the above equalty. Sce a R, ad we assume that /τ = 4σR, the y t [l] y t [l] A x t ˆx t xt t ˆx l Ayt [l] y t [l] 4τ /τ l Combe the above two equaltes, K 4σK l y t [l] xt ˆx t 4τ / P y t xt ˆx t K 4τ 4σK 4σR l y t [l] ŷt [] yt [] A x t ˆx t 3 4τ λ y t a y t [l] y t [l] x t ˆx t Sum up the above equalty over =,,..., K, ad deote Λ = K = ŷt [] yt [] A x t ˆx t, we have K K 3 4σK yt y t Λ 4τ λ x t ˆx t = Lemma 3 shows that the dstace betwee x t ad ˆx t ca be cotrol by the update of yt each terato. B Proofs of Covergece for DSPA ad A-DSPA B. Proof of Lemma Proof. We start from characterzg the relatoshp betwee x t ad x after the t-update Algorthm. Accordg to the defto of x t, we have C t x C t x t λ x t x,.e., = y t a, x x t gx gx t x x t xt x t λ We also could derve the equalty characterzg the relato betwee ŷ t = K we have ŷ t [] = arg max L t ˆxt, y [] y [] R = ŷt x t ˆx 0 [] ad y. For =,,..., K, Sce φ s/γ-smooth, we have φ s γ-strogly covex, therefore, for =,,..., K, j P ŷ t j yj a j, ˆx t φ j yj φ j ŷ t j Ky j yt j Kŷt j y t j K σ σ σ γ ŷ t j yj

13 Sum up the above equalty over j, we have ŷ t j yj a j, ˆx t φ j yj φ j ŷ t j K y [] yt σ j P j P [] K ŷt [] yt σ [] K σ γ y[] ŷt [] Sum up the above equalty over =,,..., K ad multplyg both sdes by /, K ŷ t j j P = K ŷt y t σ yj a j, ˆx t = K σ γ y ŷ t φ yj φ ŷ t j K y y t σ I addto, we cosder a combato of the saddle-pot fucto values at dfferet pots, we have fx t, y fx, y fx, y fx, y t = y a, x t φ y gx t y t a, x = = = = gx t gx φ y t φ y y a, x t = = = = φ y t gx y t a, x The we add 0 ad to the above equalty, whch mples K ŷ t j yj a j, ˆx y t a, x x t y a, x t = j P = = φ y t φ ŷ t x x t K y y t σ = fx t, y fx, y fx, y fx, y t xt x t K ŷt y t σ λ = y t a, x K x t x σ γ y ŷ t whch mples that K ŷ t j yj a j, ˆx x t j P = = ŷ t y t a, x t x x t K y y t σ fx t, y fx, y fx, y fx, y t xt x t λ K x t x σ γ y ŷ t = φ y t φ ŷ K ŷt y t σ

14 We eed to upper boud the frst term o the left-had-sde of the above equalty, assume that /τ = 4σR, we have ŷ t j j P y j a j, ˆx t xt = ŷt [] y [] A ˆx t xt the we ca get the upper boud ˆxt xt 4τ ˆxt xt 4τ ˆxt xt 4τ Aŷt [] y [] /τ j P ŷ t j y j a j 4σR ŷt [] y [] 4σK ext we deote that K K ŷ t j yj a j, ˆx x t j P = Λ = = ŷ t = y t a, x t Combg the above equalty ad equato, we derve that K = ˆx t xt ŷt y 4τ 4σK = ˆx t xt x x t K y y t Λ 4τ σ fx t, y fx, y λ x t x fx, y fx, y t K σ γ 4σK Based o the equalty 9 lemma 3, we could get that K = φ y t φ ŷ xt x t y ŷ t K ŷt y t σ ˆx t xt K 4τ 4σK yt y t x x t K y y t Λ Λ σ fx t, y fx, y fx, y fx, y t xt x t K 3 4τ λ x t ˆx t λ x t x = K σ γ 4σK K ŷt y t σ y ŷ t Sce we ca rewrte y ŷ t = y y t y t ŷ t y y t y t ŷ t, the we deote that K Λ = σ γ y t ŷ t y y t y t ŷ t 4σK K σ Based o the defto of Λ, we derve that y t ŷ t y t y t y t ŷ t x x t K y y t Λ Λ Λ σ fx t, y fx, y fx, y fx, y t λ K x t x σ γ 4σK y y t Λ

15 where we defe Λ = xt x t Assume that Θ 0, the we ca obta Θ fx t, y fx, y K λ x t ˆx t K K y t y t σ 4σK = Θ fx, y fx, y t x x t K y y t Λ σ fx t, y fx, y fx, y fx, y t λ K x t x σ γ 4σK where Λ = Λ Λ Λ ad Λ > 0. Θ fx t, y fx, y t y y t Λ I order to get the covergece guaratee of our algorthm, we eed to get the upper boud of Λ, based o the defto of Λ, we could get Λ K ŷ t [] yt = K σ γ σk based o the assumpto, the we could get that where M = 4 R Ω x 3K σ φ y t φ ŷ [] A x t ˆx t ŷt y t A x t = y y t ŷ t y t K σ yt y t ŷ t y t Λ M ŷt y t K = fˆx t, ŷt [] fˆxt, yt [] γ Ω y = c Ω x c Ω y as defed Assumpto 5. Based o the assumpto E[Λ] ˆΘ M yt ŷ t Θ K = fˆx t [] fˆxt, yt [], ŷt where ˆΘ <, Θ < are defed Assumpto 4, ad defe the parameter Θ as { } Θ = max σγ K, τλ K based o Assumpto 5, ad we could get that E[ ] = E[ M ŷt y t K = fˆx t [] fˆxt, yt [] ] Θ fx t, y fx, y t, ŷt The we could get our fal cocluso where we requre that E[ t ] ΘE[ t ] σ > Kγ, τσ = 4R 3

16 B. Proof of Theorem Proof. By Lemma, for each t > 0, we have E[ t ] Θ t E[ 0 ], accordg to the defto of C, we have I order to obta Θ T C ɛ, whch eeds Suppose the parameters τ, σ are set as The accordg to the defto of Θ { Θ = max σγ K K, E[ x t x ] Θ t C 4 log C/ɛ T log Θ σ = γ, τ = γ 4R } { = max τλ K K, As log as K K K K = K λγ 4R, sce we assume K κ. The we have Thus we ca get Θ = λγ/4r = 4R /λγ log C/ɛ T = log Θ log C/ɛ log 4R /λγ } λγ/4r where we apply the equalty log x x the last equalty ad we could get the cocluso. B.3 Covergece guaratee of prmal-dual gap Next we derve the covergece rate of prmal-dual gap based o Theorem. Lemma 4 Yu et al. 05. Suppose Assumpto, holds, Let x, y s the uque saddle-pot of fx, y, ad R = max a. The for ay pot x, y domg domφ, we have Px fx, y R γ x x, Dy fx, y R λ y y 5 Corollary. Suppose Assumpto A holds ad the parameters τ, σ, Θ are set as, 3. The the terates of Algorthm satsfy E[Px T Dy T ] ɛ t suffces to have the umber of commucato teratos T satsfy, T 4R log λγ. Proof. Based o Lemma 4, we have R λγ Px t Dy t = Px t Px Dy Dy t Smlar to Theorem, we could get the cocluso. C ɛ fx t, y R γ xt x fx, y t R λ yt y R t λγ

17 B.4 Proof of Theorem Proof. If the parameters τ, σ Algorthm are set as τ = γ 4R, σ = γ the based o Lemma, we have { } { } Θ = max σγ K K, = max τλ K K, λγ assume that K, ad K K λγ 4R, whch meas we focus o the stuato that κ = R λγ. The we could get that Θ = λγ = 4R 4R λγ Based o Corollary, we could get that Px t Px Px t Dy t 4R R t C Θ t Px 0 Px λγ where C s a costat ad C > 0 ad P s defed. If we apply DSPA o f t defed Secto 5., the based o Proposto 3. of L et al. 05, the τ M defed Proposto 3. equals τ M = 4R λϑγ where ϑ s the parameter defed Secto 5.. We could get that τ M λϑγ 4R sce κ. Now apply Proposto 3. L et al. 05, we could get the global lear rate of covergece wthe parameter τ A,F, τ A,F = τ Õ M λ/ λ ϑ = Õ λ ϑ γ λ 4R λ ϑ f we tae ϑ as ϑ = 4R γ λ, the we ca get that the commucato complexty of the A-DSPA for achevg Px T Px < ɛ s Õ κlog ɛ.

arxiv: v1 [cs.lg] 22 Feb 2015

arxiv: v1 [cs.lg] 22 Feb 2015 SDCA wthout Dualty Sha Shalev-Shwartz arxv:50.0677v cs.lg Feb 05 Abstract Stochastc Dual Coordate Ascet s a popular method for solvg regularzed loss mmzato for the case of covex losses. I ths paper we

More information

Feature Selection: Part 2. 1 Greedy Algorithms (continued from the last lecture)

Feature Selection: Part 2. 1 Greedy Algorithms (continued from the last lecture) CSE 546: Mache Learg Lecture 6 Feature Selecto: Part 2 Istructor: Sham Kakade Greedy Algorthms (cotued from the last lecture) There are varety of greedy algorthms ad umerous amg covetos for these algorthms.

More information

Solving Constrained Flow-Shop Scheduling. Problems with Three Machines

Solving Constrained Flow-Shop Scheduling. Problems with Three Machines It J Cotemp Math Sceces, Vol 5, 2010, o 19, 921-929 Solvg Costraed Flow-Shop Schedulg Problems wth Three Maches P Pada ad P Rajedra Departmet of Mathematcs, School of Advaced Sceces, VIT Uversty, Vellore-632

More information

The Mathematical Appendix

The Mathematical Appendix The Mathematcal Appedx Defto A: If ( Λ, Ω, where ( λ λ λ whch the probablty dstrbutos,,..., Defto A. uppose that ( Λ,,..., s a expermet type, the σ-algebra o λ λ λ are defed s deoted by ( (,,...,, σ Ω.

More information

Analysis of Lagrange Interpolation Formula

Analysis of Lagrange Interpolation Formula P IJISET - Iteratoal Joural of Iovatve Scece, Egeerg & Techology, Vol. Issue, December 4. www.jset.com ISS 348 7968 Aalyss of Lagrage Iterpolato Formula Vjay Dahya PDepartmet of MathematcsMaharaja Surajmal

More information

Part 4b Asymptotic Results for MRR2 using PRESS. Recall that the PRESS statistic is a special type of cross validation procedure (see Allen (1971))

Part 4b Asymptotic Results for MRR2 using PRESS. Recall that the PRESS statistic is a special type of cross validation procedure (see Allen (1971)) art 4b Asymptotc Results for MRR usg RESS Recall that the RESS statstc s a specal type of cross valdato procedure (see Alle (97)) partcular to the regresso problem ad volves fdg Y $,, the estmate at the

More information

Chapter 5 Properties of a Random Sample

Chapter 5 Properties of a Random Sample Lecture 6 o BST 63: Statstcal Theory I Ku Zhag, /0/008 Revew for the prevous lecture Cocepts: t-dstrbuto, F-dstrbuto Theorems: Dstrbutos of sample mea ad sample varace, relatoshp betwee sample mea ad sample

More information

Dimensionality Reduction and Learning

Dimensionality Reduction and Learning CMSC 35900 (Sprg 009) Large Scale Learg Lecture: 3 Dmesoalty Reducto ad Learg Istructors: Sham Kakade ad Greg Shakharovch L Supervsed Methods ad Dmesoalty Reducto The theme of these two lectures s that

More information

Cubic Nonpolynomial Spline Approach to the Solution of a Second Order Two-Point Boundary Value Problem

Cubic Nonpolynomial Spline Approach to the Solution of a Second Order Two-Point Boundary Value Problem Joural of Amerca Scece ;6( Cubc Nopolyomal Sple Approach to the Soluto of a Secod Order Two-Pot Boudary Value Problem W.K. Zahra, F.A. Abd El-Salam, A.A. El-Sabbagh ad Z.A. ZAk * Departmet of Egeerg athematcs

More information

Stochastic Primal-Dual Coordinate Method for Regularized Empirical Risk Minimization

Stochastic Primal-Dual Coordinate Method for Regularized Empirical Risk Minimization Stochastc Prmal-Dual Coordate Method for Regularzed Emprcal Rsk Mmzato Yuche Zhag L Xao September 24 Abstract We cosder a geerc covex optmzato problem assocated wth regularzed emprcal rsk mmzato of lear

More information

Functions of Random Variables

Functions of Random Variables Fuctos of Radom Varables Chapter Fve Fuctos of Radom Varables 5. Itroducto A geeral egeerg aalyss model s show Fg. 5.. The model output (respose) cotas the performaces of a system or product, such as weght,

More information

Distributed Accelerated Proximal Coordinate Gradient Methods

Distributed Accelerated Proximal Coordinate Gradient Methods Dstrbuted Accelerated Proxmal Coordate Gradet Methods Yog Re, Ju Zhu Ceter for Bo-Ispred Computg Research State Key Lab for Itell. Tech. & Systems Dept. of Comp. Sc. & Tech., TNLst Lab, Tsghua Uversty

More information

Parallel Multi-splitting Proximal Method for Star Networks

Parallel Multi-splitting Proximal Method for Star Networks Parallel Mult-splttg Proxmal Method for Star Networks Erm We Departmet of Electrcal Egeerg ad Computer Scece Northwester Uversty Evasto, IL 600 erm.we@orthwester.edu Abstract We develop a parallel algorthm

More information

Rademacher Complexity. Examples

Rademacher Complexity. Examples Algorthmc Foudatos of Learg Lecture 3 Rademacher Complexty. Examples Lecturer: Patrck Rebesch Verso: October 16th 018 3.1 Itroducto I the last lecture we troduced the oto of Rademacher complexty ad showed

More information

Point Estimation: definition of estimators

Point Estimation: definition of estimators Pot Estmato: defto of estmators Pot estmator: ay fucto W (X,..., X ) of a data sample. The exercse of pot estmato s to use partcular fuctos of the data order to estmate certa ukow populato parameters.

More information

Introduction to local (nonparametric) density estimation. methods

Introduction to local (nonparametric) density estimation. methods Itroducto to local (oparametrc) desty estmato methods A slecture by Yu Lu for ECE 66 Sprg 014 1. Itroducto Ths slecture troduces two local desty estmato methods whch are Parze desty estmato ad k-earest

More information

An Accelerated Proximal Coordinate Gradient Method

An Accelerated Proximal Coordinate Gradient Method A Accelerated Proxmal Coordate Gradet Method Qhag L Uversty of Iowa Iowa Cty IA USA qhag-l@uowaedu Zhaosog Lu Smo Fraser Uversty Buraby BC Caada zhaosog@sfuca L Xao Mcrosoft Research Redmod WA USA lxao@mcrosoftcom

More information

Lecture 9: Tolerant Testing

Lecture 9: Tolerant Testing Lecture 9: Tolerat Testg Dael Kae Scrbe: Sakeerth Rao Aprl 4, 07 Abstract I ths lecture we prove a quas lear lower boud o the umber of samples eeded to do tolerat testg for L dstace. Tolerat Testg We have

More information

TESTS BASED ON MAXIMUM LIKELIHOOD

TESTS BASED ON MAXIMUM LIKELIHOOD ESE 5 Toy E. Smth. The Basc Example. TESTS BASED ON MAXIMUM LIKELIHOOD To llustrate the propertes of maxmum lkelhood estmates ad tests, we cosder the smplest possble case of estmatg the mea of the ormal

More information

X ε ) = 0, or equivalently, lim

X ε ) = 0, or equivalently, lim Revew for the prevous lecture Cocepts: order statstcs Theorems: Dstrbutos of order statstcs Examples: How to get the dstrbuto of order statstcs Chapter 5 Propertes of a Radom Sample Secto 55 Covergece

More information

CHAPTER VI Statistical Analysis of Experimental Data

CHAPTER VI Statistical Analysis of Experimental Data Chapter VI Statstcal Aalyss of Expermetal Data CHAPTER VI Statstcal Aalyss of Expermetal Data Measuremets do ot lead to a uque value. Ths s a result of the multtude of errors (maly radom errors) that ca

More information

Econometric Methods. Review of Estimation

Econometric Methods. Review of Estimation Ecoometrc Methods Revew of Estmato Estmatg the populato mea Radom samplg Pot ad terval estmators Lear estmators Ubased estmators Lear Ubased Estmators (LUEs) Effcecy (mmum varace) ad Best Lear Ubased Estmators

More information

Estimation of Stress- Strength Reliability model using finite mixture of exponential distributions

Estimation of Stress- Strength Reliability model using finite mixture of exponential distributions Iteratoal Joural of Computatoal Egeerg Research Vol, 0 Issue, Estmato of Stress- Stregth Relablty model usg fte mxture of expoetal dstrbutos K.Sadhya, T.S.Umamaheswar Departmet of Mathematcs, Lal Bhadur

More information

A New Method for Decision Making Based on Soft Matrix Theory

A New Method for Decision Making Based on Soft Matrix Theory Joural of Scetfc esearch & eports 3(5): 0-7, 04; rtcle o. JS.04.5.00 SCIENCEDOMIN teratoal www.scecedoma.org New Method for Decso Mag Based o Soft Matrx Theory Zhmg Zhag * College of Mathematcs ad Computer

More information

Lecture 16: Backpropogation Algorithm Neural Networks with smooth activation functions

Lecture 16: Backpropogation Algorithm Neural Networks with smooth activation functions CO-511: Learg Theory prg 2017 Lecturer: Ro Lv Lecture 16: Bacpropogato Algorthm Dsclamer: These otes have ot bee subected to the usual scruty reserved for formal publcatos. They may be dstrbuted outsde

More information

A New Family of Transformations for Lifetime Data

A New Family of Transformations for Lifetime Data Proceedgs of the World Cogress o Egeerg 4 Vol I, WCE 4, July - 4, 4, Lodo, U.K. A New Famly of Trasformatos for Lfetme Data Lakhaa Watthaacheewakul Abstract A famly of trasformatos s the oe of several

More information

Derivation of 3-Point Block Method Formula for Solving First Order Stiff Ordinary Differential Equations

Derivation of 3-Point Block Method Formula for Solving First Order Stiff Ordinary Differential Equations Dervato of -Pot Block Method Formula for Solvg Frst Order Stff Ordary Dfferetal Equatos Kharul Hamd Kharul Auar, Kharl Iskadar Othma, Zara Bb Ibrahm Abstract Dervato of pot block method formula wth costat

More information

Ordinary Least Squares Regression. Simple Regression. Algebra and Assumptions.

Ordinary Least Squares Regression. Simple Regression. Algebra and Assumptions. Ordary Least Squares egresso. Smple egresso. Algebra ad Assumptos. I ths part of the course we are gog to study a techque for aalysg the lear relatoshp betwee two varables Y ad X. We have pars of observatos

More information

Chapter 4 (Part 1): Non-Parametric Classification (Sections ) Pattern Classification 4.3) Announcements

Chapter 4 (Part 1): Non-Parametric Classification (Sections ) Pattern Classification 4.3) Announcements Aoucemets No-Parametrc Desty Estmato Techques HW assged Most of ths lecture was o the blacboard. These sldes cover the same materal as preseted DHS Bometrcs CSE 90-a Lecture 7 CSE90a Fall 06 CSE90a Fall

More information

PROJECTION PROBLEM FOR REGULAR POLYGONS

PROJECTION PROBLEM FOR REGULAR POLYGONS Joural of Mathematcal Sceces: Advaces ad Applcatos Volume, Number, 008, Pages 95-50 PROJECTION PROBLEM FOR REGULAR POLYGONS College of Scece Bejg Forestry Uversty Bejg 0008 P. R. Cha e-mal: sl@bjfu.edu.c

More information

Analysis of Variance with Weibull Data

Analysis of Variance with Weibull Data Aalyss of Varace wth Webull Data Lahaa Watthaacheewaul Abstract I statstcal data aalyss by aalyss of varace, the usual basc assumptos are that the model s addtve ad the errors are radomly, depedetly, ad

More information

2006 Jamie Trahan, Autar Kaw, Kevin Martin University of South Florida United States of America

2006 Jamie Trahan, Autar Kaw, Kevin Martin University of South Florida United States of America SOLUTION OF SYSTEMS OF SIMULTANEOUS LINEAR EQUATIONS Gauss-Sedel Method 006 Jame Traha, Autar Kaw, Kev Mart Uversty of South Florda Uted States of Amerca kaw@eg.usf.edu Itroducto Ths worksheet demostrates

More information

A new type of optimization method based on conjugate directions

A new type of optimization method based on conjugate directions A ew type of optmzato method based o cojugate drectos Pa X Scece School aj Uversty of echology ad Educato (UE aj Cha e-mal: pax94@sacom Abstract A ew type of optmzato method based o cojugate drectos s

More information

LECTURE 24 LECTURE OUTLINE

LECTURE 24 LECTURE OUTLINE LECTURE 24 LECTURE OUTLINE Gradet proxmal mmzato method Noquadratc proxmal algorthms Etropy mmzato algorthm Expoetal augmeted Lagraga mehod Etropc descet algorthm **************************************

More information

Unimodality Tests for Global Optimization of Single Variable Functions Using Statistical Methods

Unimodality Tests for Global Optimization of Single Variable Functions Using Statistical Methods Malaysa Umodalty Joural Tests of Mathematcal for Global Optmzato Sceces (): of 05 Sgle - 5 Varable (007) Fuctos Usg Statstcal Methods Umodalty Tests for Global Optmzato of Sgle Varable Fuctos Usg Statstcal

More information

Bayes Estimator for Exponential Distribution with Extension of Jeffery Prior Information

Bayes Estimator for Exponential Distribution with Extension of Jeffery Prior Information Malaysa Joural of Mathematcal Sceces (): 97- (9) Bayes Estmator for Expoetal Dstrbuto wth Exteso of Jeffery Pror Iformato Hadeel Salm Al-Kutub ad Noor Akma Ibrahm Isttute for Mathematcal Research, Uverst

More information

CIS 800/002 The Algorithmic Foundations of Data Privacy October 13, Lecture 9. Database Update Algorithms: Multiplicative Weights

CIS 800/002 The Algorithmic Foundations of Data Privacy October 13, Lecture 9. Database Update Algorithms: Multiplicative Weights CIS 800/002 The Algorthmc Foudatos of Data Prvacy October 13, 2011 Lecturer: Aaro Roth Lecture 9 Scrbe: Aaro Roth Database Update Algorthms: Multplcatve Weghts We ll recall aga) some deftos from last tme:

More information

ESS Line Fitting

ESS Line Fitting ESS 5 014 17. Le Fttg A very commo problem data aalyss s lookg for relatoshpetwee dfferet parameters ad fttg les or surfaces to data. The smplest example s fttg a straght le ad we wll dscuss that here

More information

CHAPTER 4 RADICAL EXPRESSIONS

CHAPTER 4 RADICAL EXPRESSIONS 6 CHAPTER RADICAL EXPRESSIONS. The th Root of a Real Number A real umber a s called the th root of a real umber b f Thus, for example: s a square root of sce. s also a square root of sce ( ). s a cube

More information

BERNSTEIN COLLOCATION METHOD FOR SOLVING NONLINEAR DIFFERENTIAL EQUATIONS. Aysegul Akyuz Dascioglu and Nese Isler

BERNSTEIN COLLOCATION METHOD FOR SOLVING NONLINEAR DIFFERENTIAL EQUATIONS. Aysegul Akyuz Dascioglu and Nese Isler Mathematcal ad Computatoal Applcatos, Vol. 8, No. 3, pp. 293-300, 203 BERNSTEIN COLLOCATION METHOD FOR SOLVING NONLINEAR DIFFERENTIAL EQUATIONS Aysegul Ayuz Dascoglu ad Nese Isler Departmet of Mathematcs,

More information

Lecture 3. Sampling, sampling distributions, and parameter estimation

Lecture 3. Sampling, sampling distributions, and parameter estimation Lecture 3 Samplg, samplg dstrbutos, ad parameter estmato Samplg Defto Populato s defed as the collecto of all the possble observatos of terest. The collecto of observatos we take from the populato s called

More information

A tighter lower bound on the circuit size of the hardest Boolean functions

A tighter lower bound on the circuit size of the hardest Boolean functions Electroc Colloquum o Computatoal Complexty, Report No. 86 2011) A tghter lower boud o the crcut sze of the hardest Boolea fuctos Masak Yamamoto Abstract I [IPL2005], Fradse ad Mlterse mproved bouds o the

More information

X X X E[ ] E X E X. is the ()m n where the ( i,)th. j element is the mean of the ( i,)th., then

X X X E[ ] E X E X. is the ()m n where the ( i,)th. j element is the mean of the ( i,)th., then Secto 5 Vectors of Radom Varables Whe workg wth several radom varables,,..., to arrage them vector form x, t s ofte coveet We ca the make use of matrx algebra to help us orgaze ad mapulate large umbers

More information

Simulation Output Analysis

Simulation Output Analysis Smulato Output Aalyss Summary Examples Parameter Estmato Sample Mea ad Varace Pot ad Iterval Estmato ermatg ad o-ermatg Smulato Mea Square Errors Example: Sgle Server Queueg System x(t) S 4 S 4 S 3 S 5

More information

UNIT 2 SOLUTION OF ALGEBRAIC AND TRANSCENDENTAL EQUATIONS

UNIT 2 SOLUTION OF ALGEBRAIC AND TRANSCENDENTAL EQUATIONS Numercal Computg -I UNIT SOLUTION OF ALGEBRAIC AND TRANSCENDENTAL EQUATIONS Structure Page Nos..0 Itroducto 6. Objectves 7. Ital Approxmato to a Root 7. Bsecto Method 8.. Error Aalyss 9.4 Regula Fals Method

More information

Support vector machines

Support vector machines CS 75 Mache Learg Lecture Support vector maches Mlos Hauskrecht mlos@cs.ptt.edu 539 Seott Square CS 75 Mache Learg Outle Outle: Algorthms for lear decso boudary Support vector maches Mamum marg hyperplae.

More information

Chapter 2 - Free Vibration of Multi-Degree-of-Freedom Systems - II

Chapter 2 - Free Vibration of Multi-Degree-of-Freedom Systems - II CEE49b Chapter - Free Vbrato of Mult-Degree-of-Freedom Systems - II We ca obta a approxmate soluto to the fudametal atural frequecy through a approxmate formula developed usg eergy prcples by Lord Raylegh

More information

4 Inner Product Spaces

4 Inner Product Spaces 11.MH1 LINEAR ALGEBRA Summary Notes 4 Ier Product Spaces Ier product s the abstracto to geeral vector spaces of the famlar dea of the scalar product of two vectors or 3. I what follows, keep these key

More information

Investigation of Partially Conditional RP Model with Response Error. Ed Stanek

Investigation of Partially Conditional RP Model with Response Error. Ed Stanek Partally Codtoal Radom Permutato Model 7- vestgato of Partally Codtoal RP Model wth Respose Error TRODUCTO Ed Staek We explore the predctor that wll result a smple radom sample wth respose error whe a

More information

A conic cutting surface method for linear-quadraticsemidefinite

A conic cutting surface method for linear-quadraticsemidefinite A coc cuttg surface method for lear-quadratcsemdefte programmg Mohammad R. Osoorouch Calfora State Uversty Sa Marcos Sa Marcos, CA Jot wor wth Joh E. Mtchell RPI July 3, 2008 Outle: Secod-order coe: defto

More information

MATH 247/Winter Notes on the adjoint and on normal operators.

MATH 247/Winter Notes on the adjoint and on normal operators. MATH 47/Wter 00 Notes o the adjot ad o ormal operators I these otes, V s a fte dmesoal er product space over, wth gve er * product uv, T, S, T, are lear operators o V U, W are subspaces of V Whe we say

More information

CS 2750 Machine Learning. Lecture 8. Linear regression. CS 2750 Machine Learning. Linear regression. is a linear combination of input components x

CS 2750 Machine Learning. Lecture 8. Linear regression. CS 2750 Machine Learning. Linear regression. is a linear combination of input components x CS 75 Mache Learg Lecture 8 Lear regresso Mlos Hauskrecht mlos@cs.ptt.edu 539 Seott Square CS 75 Mache Learg Lear regresso Fucto f : X Y s a lear combato of put compoets f + + + K d d K k - parameters

More information

18.413: Error Correcting Codes Lab March 2, Lecture 8

18.413: Error Correcting Codes Lab March 2, Lecture 8 18.413: Error Correctg Codes Lab March 2, 2004 Lecturer: Dael A. Spelma Lecture 8 8.1 Vector Spaces A set C {0, 1} s a vector space f for x all C ad y C, x + y C, where we take addto to be compoet wse

More information

Summary of the lecture in Biostatistics

Summary of the lecture in Biostatistics Summary of the lecture Bostatstcs Probablty Desty Fucto For a cotuos radom varable, a probablty desty fucto s a fucto such that: 0 dx a b) b a dx A probablty desty fucto provdes a smple descrpto of the

More information

ANALYSIS ON THE NATURE OF THE BASIC EQUATIONS IN SYNERGETIC INTER-REPRESENTATION NETWORK

ANALYSIS ON THE NATURE OF THE BASIC EQUATIONS IN SYNERGETIC INTER-REPRESENTATION NETWORK Far East Joural of Appled Mathematcs Volume, Number, 2008, Pages Ths paper s avalable ole at http://www.pphm.com 2008 Pushpa Publshg House ANALYSIS ON THE NATURE OF THE ASI EQUATIONS IN SYNERGETI INTER-REPRESENTATION

More information

Generating Multivariate Nonnormal Distribution Random Numbers Based on Copula Function

Generating Multivariate Nonnormal Distribution Random Numbers Based on Copula Function 7659, Eglad, UK Joural of Iformato ad Computg Scece Vol. 2, No. 3, 2007, pp. 9-96 Geeratg Multvarate Noormal Dstrbuto Radom Numbers Based o Copula Fucto Xaopg Hu +, Jam He ad Hogsheg Ly School of Ecoomcs

More information

Comparison of Dual to Ratio-Cum-Product Estimators of Population Mean

Comparison of Dual to Ratio-Cum-Product Estimators of Population Mean Research Joural of Mathematcal ad Statstcal Sceces ISS 30 6047 Vol. 1(), 5-1, ovember (013) Res. J. Mathematcal ad Statstcal Sc. Comparso of Dual to Rato-Cum-Product Estmators of Populato Mea Abstract

More information

Lecture 3 Probability review (cont d)

Lecture 3 Probability review (cont d) STATS 00: Itroducto to Statstcal Iferece Autum 06 Lecture 3 Probablty revew (cot d) 3. Jot dstrbutos If radom varables X,..., X k are depedet, the ther dstrbuto may be specfed by specfyg the dvdual dstrbuto

More information

On Modified Interval Symmetric Single-Step Procedure ISS2-5D for the Simultaneous Inclusion of Polynomial Zeros

On Modified Interval Symmetric Single-Step Procedure ISS2-5D for the Simultaneous Inclusion of Polynomial Zeros It. Joural of Math. Aalyss, Vol. 7, 2013, o. 20, 983-988 HIKARI Ltd, www.m-hkar.com O Modfed Iterval Symmetrc Sgle-Step Procedure ISS2-5D for the Smultaeous Icluso of Polyomal Zeros 1 Nora Jamalud, 1 Masor

More information

Research Article A New Iterative Method for Common Fixed Points of a Finite Family of Nonexpansive Mappings

Research Article A New Iterative Method for Common Fixed Points of a Finite Family of Nonexpansive Mappings Hdaw Publshg Corporato Iteratoal Joural of Mathematcs ad Mathematcal Sceces Volume 009, Artcle ID 391839, 9 pages do:10.1155/009/391839 Research Artcle A New Iteratve Method for Commo Fxed Pots of a Fte

More information

Entropy ISSN by MDPI

Entropy ISSN by MDPI Etropy 2003, 5, 233-238 Etropy ISSN 1099-4300 2003 by MDPI www.mdp.org/etropy O the Measure Etropy of Addtve Cellular Automata Hasa Aı Arts ad Sceces Faculty, Departmet of Mathematcs, Harra Uversty; 63100,

More information

( ) = ( ) ( ) Chapter 13 Asymptotic Theory and Stochastic Regressors. Stochastic regressors model

( ) = ( ) ( ) Chapter 13 Asymptotic Theory and Stochastic Regressors. Stochastic regressors model Chapter 3 Asmptotc Theor ad Stochastc Regressors The ature of eplaator varable s assumed to be o-stochastc or fed repeated samples a regresso aalss Such a assumpto s approprate for those epermets whch

More information

Solution of General Dual Fuzzy Linear Systems. Using ABS Algorithm

Solution of General Dual Fuzzy Linear Systems. Using ABS Algorithm Appled Mathematcal Sceces, Vol 6, 0, o 4, 63-7 Soluto of Geeral Dual Fuzzy Lear Systems Usg ABS Algorthm M A Farborz Aragh * ad M M ossezadeh Departmet of Mathematcs, Islamc Azad Uversty Cetral ehra Brach,

More information

{ }{ ( )} (, ) = ( ) ( ) ( ) Chapter 14 Exercises in Sampling Theory. Exercise 1 (Simple random sampling): Solution:

{ }{ ( )} (, ) = ( ) ( ) ( ) Chapter 14 Exercises in Sampling Theory. Exercise 1 (Simple random sampling): Solution: Chapter 4 Exercses Samplg Theory Exercse (Smple radom samplg: Let there be two correlated radom varables X ad A sample of sze s draw from a populato by smple radom samplg wthout replacemet The observed

More information

Chapter 9 Jordan Block Matrices

Chapter 9 Jordan Block Matrices Chapter 9 Jorda Block atrces I ths chapter we wll solve the followg problem. Gve a lear operator T fd a bass R of F such that the matrx R (T) s as smple as possble. f course smple s a matter of taste.

More information

CS 1675 Introduction to Machine Learning Lecture 12 Support vector machines

CS 1675 Introduction to Machine Learning Lecture 12 Support vector machines CS 675 Itroducto to Mache Learg Lecture Support vector maches Mlos Hauskrecht mlos@cs.ptt.edu 539 Seott Square Mdterm eam October 9, 7 I-class eam Closed book Stud materal: Lecture otes Correspodg chapters

More information

Investigating Cellular Automata

Investigating Cellular Automata Researcher: Taylor Dupuy Advsor: Aaro Wootto Semester: Fall 4 Ivestgatg Cellular Automata A Overvew of Cellular Automata: Cellular Automata are smple computer programs that geerate rows of black ad whte

More information

Supervised learning: Linear regression Logistic regression

Supervised learning: Linear regression Logistic regression CS 57 Itroducto to AI Lecture 4 Supervsed learg: Lear regresso Logstc regresso Mlos Hauskrecht mlos@cs.ptt.edu 539 Seott Square CS 57 Itro to AI Data: D { D D.. D D Supervsed learg d a set of eamples s

More information

Multivariate Transformation of Variables and Maximum Likelihood Estimation

Multivariate Transformation of Variables and Maximum Likelihood Estimation Marquette Uversty Multvarate Trasformato of Varables ad Maxmum Lkelhood Estmato Dael B. Rowe, Ph.D. Assocate Professor Departmet of Mathematcs, Statstcs, ad Computer Scece Copyrght 03 by Marquette Uversty

More information

Chapter 8. Inferences about More Than Two Population Central Values

Chapter 8. Inferences about More Than Two Population Central Values Chapter 8. Ifereces about More Tha Two Populato Cetral Values Case tudy: Effect of Tmg of the Treatmet of Port-We tas wth Lasers ) To vestgate whether treatmet at a youg age would yeld better results tha

More information

Maximum Likelihood Estimation

Maximum Likelihood Estimation Marquette Uverst Maxmum Lkelhood Estmato Dael B. Rowe, Ph.D. Professor Departmet of Mathematcs, Statstcs, ad Computer Scece Coprght 08 b Marquette Uverst Maxmum Lkelhood Estmato We have bee sag that ~

More information

Lecture 07: Poles and Zeros

Lecture 07: Poles and Zeros Lecture 07: Poles ad Zeros Defto of poles ad zeros The trasfer fucto provdes a bass for determg mportat system respose characterstcs wthout solvg the complete dfferetal equato. As defed, the trasfer fucto

More information

1 0, x? x x. 1 Root finding. 1.1 Introduction. Solve[x^2-1 0,x] {{x -1},{x 1}} Plot[x^2-1,{x,-2,2}] 3

1 0, x? x x. 1 Root finding. 1.1 Introduction. Solve[x^2-1 0,x] {{x -1},{x 1}} Plot[x^2-1,{x,-2,2}] 3 Adrew Powuk - http://www.powuk.com- Math 49 (Numercal Aalyss) Root fdg. Itroducto f ( ),?,? Solve[^-,] {{-},{}} Plot[^-,{,-,}] Cubc equato https://e.wkpeda.org/wk/cubc_fucto Quartc equato https://e.wkpeda.org/wk/quartc_fucto

More information

Bayes Interval Estimation for binomial proportion and difference of two binomial proportions with Simulation Study

Bayes Interval Estimation for binomial proportion and difference of two binomial proportions with Simulation Study IJIEST Iteratoal Joural of Iovatve Scece, Egeerg & Techology, Vol. Issue 5, July 04. Bayes Iterval Estmato for bomal proporto ad dfferece of two bomal proportos wth Smulato Study Masoud Gaj, Solmaz hlmad

More information

C-1: Aerodynamics of Airfoils 1 C-2: Aerodynamics of Airfoils 2 C-3: Panel Methods C-4: Thin Airfoil Theory

C-1: Aerodynamics of Airfoils 1 C-2: Aerodynamics of Airfoils 2 C-3: Panel Methods C-4: Thin Airfoil Theory ROAD MAP... AE301 Aerodyamcs I UNIT C: 2-D Arfols C-1: Aerodyamcs of Arfols 1 C-2: Aerodyamcs of Arfols 2 C-3: Pael Methods C-4: Th Arfol Theory AE301 Aerodyamcs I Ut C-3: Lst of Subects Problem Solutos?

More information

Dynamic Analysis of Axially Beam on Visco - Elastic Foundation with Elastic Supports under Moving Load

Dynamic Analysis of Axially Beam on Visco - Elastic Foundation with Elastic Supports under Moving Load Dyamc Aalyss of Axally Beam o Vsco - Elastc Foudato wth Elastc Supports uder Movg oad Saeed Mohammadzadeh, Seyed Al Mosayeb * Abstract: For dyamc aalyses of ralway track structures, the algorthm of soluto

More information

An Introduction to. Support Vector Machine

An Introduction to. Support Vector Machine A Itroducto to Support Vector Mache Support Vector Mache (SVM) A classfer derved from statstcal learg theory by Vapk, et al. 99 SVM became famous whe, usg mages as put, t gave accuracy comparable to eural-etwork

More information

Bayes (Naïve or not) Classifiers: Generative Approach

Bayes (Naïve or not) Classifiers: Generative Approach Logstc regresso Bayes (Naïve or ot) Classfers: Geeratve Approach What do we mea by Geeratve approach: Lear p(y), p(x y) ad the apply bayes rule to compute p(y x) for makg predctos Ths s essetally makg

More information

Support vector machines II

Support vector machines II CS 75 Mache Learg Lecture Support vector maches II Mlos Hauskrecht mlos@cs.ptt.edu 539 Seott Square Learl separable classes Learl separable classes: here s a hperplae that separates trag staces th o error

More information

Arithmetic Mean and Geometric Mean

Arithmetic Mean and Geometric Mean Acta Mathematca Ntresa Vol, No, p 43 48 ISSN 453-6083 Arthmetc Mea ad Geometrc Mea Mare Varga a * Peter Mchalča b a Departmet of Mathematcs, Faculty of Natural Sceces, Costate the Phlosopher Uversty Ntra,

More information

Pinaki Mitra Dept. of CSE IIT Guwahati

Pinaki Mitra Dept. of CSE IIT Guwahati Pak Mtra Dept. of CSE IIT Guwahat Hero s Problem HIGHWAY FACILITY LOCATION Faclty Hgh Way Farm A Farm B Illustrato of the Proof of Hero s Theorem p q s r r l d(p,r) + d(q,r) = d(p,q) p d(p,r ) + d(q,r

More information

The number of observed cases The number of parameters. ith case of the dichotomous dependent variable. the ith case of the jth parameter

The number of observed cases The number of parameters. ith case of the dichotomous dependent variable. the ith case of the jth parameter LOGISTIC REGRESSION Notato Model Logstc regresso regresses a dchotomous depedet varable o a set of depedet varables. Several methods are mplemeted for selectg the depedet varables. The followg otato s

More information

CS286.2 Lecture 4: Dinur s Proof of the PCP Theorem

CS286.2 Lecture 4: Dinur s Proof of the PCP Theorem CS86. Lecture 4: Dur s Proof of the PCP Theorem Scrbe: Thom Bohdaowcz Prevously, we have prove a weak verso of the PCP theorem: NP PCP 1,1/ (r = poly, q = O(1)). Wth ths result we have the desred costat

More information

Discrete Mathematics and Probability Theory Fall 2016 Seshia and Walrand DIS 10b

Discrete Mathematics and Probability Theory Fall 2016 Seshia and Walrand DIS 10b CS 70 Dscrete Mathematcs ad Probablty Theory Fall 206 Sesha ad Walrad DIS 0b. Wll I Get My Package? Seaky delvery guy of some compay s out delverg packages to customers. Not oly does he had a radom package

More information

MULTIDIMENSIONAL HETEROGENEOUS VARIABLE PREDICTION BASED ON EXPERTS STATEMENTS. Gennadiy Lbov, Maxim Gerasimov

MULTIDIMENSIONAL HETEROGENEOUS VARIABLE PREDICTION BASED ON EXPERTS STATEMENTS. Gennadiy Lbov, Maxim Gerasimov Iteratoal Boo Seres "Iformato Scece ad Computg" 97 MULTIIMNSIONAL HTROGNOUS VARIABL PRICTION BAS ON PRTS STATMNTS Geady Lbov Maxm Gerasmov Abstract: I the wors [ ] we proposed a approach of formg a cosesus

More information

Homework 1: Solutions Sid Banerjee Problem 1: (Practice with Asymptotic Notation) ORIE 4520: Stochastics at Scale Fall 2015

Homework 1: Solutions Sid Banerjee Problem 1: (Practice with Asymptotic Notation) ORIE 4520: Stochastics at Scale Fall 2015 Fall 05 Homework : Solutos Problem : (Practce wth Asymptotc Notato) A essetal requremet for uderstadg scalg behavor s comfort wth asymptotc (or bg-o ) otato. I ths problem, you wll prove some basc facts

More information

Lecture Note to Rice Chapter 8

Lecture Note to Rice Chapter 8 ECON 430 HG revsed Nov 06 Lecture Note to Rce Chapter 8 Radom matrces Let Y, =,,, m, =,,, be radom varables (r.v. s). The matrx Y Y Y Y Y Y Y Y Y Y = m m m s called a radom matrx ( wth a ot m-dmesoal dstrbuto,

More information

A Remark on the Uniform Convergence of Some Sequences of Functions

A Remark on the Uniform Convergence of Some Sequences of Functions Advaces Pure Mathematcs 05 5 57-533 Publshed Ole July 05 ScRes. http://www.scrp.org/joural/apm http://dx.do.org/0.436/apm.05.59048 A Remark o the Uform Covergece of Some Sequeces of Fuctos Guy Degla Isttut

More information

NP!= P. By Liu Ran. Table of Contents. The P vs. NP problem is a major unsolved problem in computer

NP!= P. By Liu Ran. Table of Contents. The P vs. NP problem is a major unsolved problem in computer NP!= P By Lu Ra Table of Cotets. Itroduce 2. Strategy 3. Prelmary theorem 4. Proof 5. Expla 6. Cocluso. Itroduce The P vs. NP problem s a major usolved problem computer scece. Iformally, t asks whether

More information

Simple Linear Regression

Simple Linear Regression Statstcal Methods I (EST 75) Page 139 Smple Lear Regresso Smple regresso applcatos are used to ft a model descrbg a lear relatoshp betwee two varables. The aspects of least squares regresso ad correlato

More information

( ) 2 2. Multi-Layer Refraction Problem Rafael Espericueta, Bakersfield College, November, 2006

( ) 2 2. Multi-Layer Refraction Problem Rafael Espericueta, Bakersfield College, November, 2006 Mult-Layer Refracto Problem Rafael Espercueta, Bakersfeld College, November, 006 Lght travels at dfferet speeds through dfferet meda, but refracts at layer boudares order to traverse the least-tme path.

More information

NP!= P. By Liu Ran. Table of Contents. The P versus NP problem is a major unsolved problem in computer

NP!= P. By Liu Ran. Table of Contents. The P versus NP problem is a major unsolved problem in computer NP!= P By Lu Ra Table of Cotets. Itroduce 2. Prelmary theorem 3. Proof 4. Expla 5. Cocluso. Itroduce The P versus NP problem s a major usolved problem computer scece. Iformally, t asks whether a computer

More information

Chapter 4 Multiple Random Variables

Chapter 4 Multiple Random Variables Revew for the prevous lecture: Theorems ad Examples: How to obta the pmf (pdf) of U = g (, Y) ad V = g (, Y) Chapter 4 Multple Radom Varables Chapter 44 Herarchcal Models ad Mxture Dstrbutos Examples:

More information

Block-Based Compact Thermal Modeling of Semiconductor Integrated Circuits

Block-Based Compact Thermal Modeling of Semiconductor Integrated Circuits Block-Based Compact hermal Modelg of Semcoductor Itegrated Crcuts Master s hess Defese Caddate: Jg Ba Commttee Members: Dr. Mg-Cheg Cheg Dr. Daqg Hou Dr. Robert Schllg July 27, 2009 Outle Itroducto Backgroud

More information

Bootstrap Method for Testing of Equality of Several Coefficients of Variation

Bootstrap Method for Testing of Equality of Several Coefficients of Variation Cloud Publcatos Iteratoal Joural of Advaced Mathematcs ad Statstcs Volume, pp. -6, Artcle ID Sc- Research Artcle Ope Access Bootstrap Method for Testg of Equalty of Several Coeffcets of Varato Dr. Navee

More information

ENGI 3423 Simple Linear Regression Page 12-01

ENGI 3423 Simple Linear Regression Page 12-01 ENGI 343 mple Lear Regresso Page - mple Lear Regresso ometmes a expermet s set up where the expermeter has cotrol over the values of oe or more varables X ad measures the resultg values of aother varable

More information

Unsupervised Learning and Other Neural Networks

Unsupervised Learning and Other Neural Networks CSE 53 Soft Computg NOT PART OF THE FINAL Usupervsed Learg ad Other Neural Networs Itroducto Mture Destes ad Idetfablty ML Estmates Applcato to Normal Mtures Other Neural Networs Itroducto Prevously, all

More information

ρ < 1 be five real numbers. The

ρ < 1 be five real numbers. The Lecture o BST 63: Statstcal Theory I Ku Zhag, /0/006 Revew for the prevous lecture Deftos: covarace, correlato Examples: How to calculate covarace ad correlato Theorems: propertes of correlato ad covarace

More information

Comparing Different Estimators of three Parameters for Transmuted Weibull Distribution

Comparing Different Estimators of three Parameters for Transmuted Weibull Distribution Global Joural of Pure ad Appled Mathematcs. ISSN 0973-768 Volume 3, Number 9 (207), pp. 55-528 Research Ida Publcatos http://www.rpublcato.com Comparg Dfferet Estmators of three Parameters for Trasmuted

More information

G S Power Flow Solution

G S Power Flow Solution G S Power Flow Soluto P Q I y y * 0 1, Y y Y 0 y Y Y 1, P Q ( k) ( k) * ( k 1) 1, Y Y PQ buses * 1 P Q Y ( k1) *( k) ( k) Q Im[ Y ] 1 P buses & Slack bus ( k 1) *( k) ( k) Y 1 P Re[ ] Slack bus 17 Calculato

More information