Bayesian Filtering with Online Gaussian Process Latent Variable Models

Size: px
Start display at page:

Download "Bayesian Filtering with Online Gaussian Process Latent Variable Models"

Transcription

1 Bayesian Filering wih Online Gaussian Process Laen Variable Models Yali Wang Laval Universiy Marcus A. Brubaker TTI Chicago Brahim Chaib-draa Laval Universiy Raquel Urasun Universiy of Torono Absrac In his paper we presen a novel non-parameric approach o Bayesian filering, where he predicion and observaion models are learned in an online fashion. Our approach is able o handle mulimodal disribuions over boh models by employing a mixure model represenaion wih Gaussian Processes (GP) based componens. To cope wih he increasing complexiy of he esimaion process, we explore wo compuaionally efficien GP varians, sparse online GP and local GP, which help o manage compuaion requiremens for each mixure componen. Our experimens demonsrae ha our approach can rack human moion much more accuraely han exising approaches ha learn he predicion and observaion models offline and do no updae hese models wih he incoming daa sream. 1 INTRODUCTION Many real world problems involve high dimensional daa. In his paper we are ineresed in modeling and racking human moion. In his seing, dimensionaliy reducion echniques are widely employed o avoid he curse of dimensionaliy. Linear approaches such as principle componen analysis (PCA) are very popular as hey are simple o use. However, hey ofen fail o capure complex dependencies due o heir assumpion of lineariy. Non-linear dimensionaliy reducion echniques ha aemp o preserve he local srucure of he manifold (e.g., Isomap [1, ], LLE [1, 1]) can capure more complex dependencies, bu ofen suffer when he manifold assumpions are violaed, e.g., in he presence of noise. Probabilisic laen variable models have he advanage of being able o ake he uncerainies ino accoun when learning he laen represenaions. Perhaps he mos successful model in he conex of modelling human moion is he Gaussian process laen variable model (GPLVM) [1], where he non-linear mapping beween he laen space and he high dimensional space is modeled wih a Gaussian process. This provides powerful prior models, which have been employed for characer animaion [,, 1] and human body racking [, 1, ]. In he conex of racking, one is ineresed in esimaing he sae of a dynamic sysem. The mos commonly used echnique for sae esimaion is Bayesian filering, which recursively esimaes he poserior probabiliy of he sae of he sysem. The wo key componens in he filer are he predicion model, which describes he emporal evoluion of he process, as well as he observaion model which links he sae and he observaions. A parameric form is ypically employed for boh models. Ko and Fox [1] inroduced he GP-BayesFiler, which defines he predicion and observaion models in a nonparameric way via Gaussian processes. This approach is well suied when accurae parameric models are difficul o obain. Is main limiaion, however, resides in he fac ha i requires ground ruh saes (as GPs are supervised), which are ypically no available. GPLVMs were employed in [11] o learn he laen space in an unsupervised manner, bypassing he need for labeled daa. This, however, can no exploi he incoming sream of daa available in he online seing as he laen space is learned offline. Furhermore, only unimodal predicion and observaion models can be capured due o he fac ha he models learned by GP are nonlinear bu Gaussian. In his paper we exend he previous non-parameric filers o learn he laen space in an online fashion as well as o handle mulimodal disribuions for boh he predicion and observaion models. Towards his goal, we employ a mixure model represenaion in he paricle filering framework. For he mixure componens, we invesigae wo compuaionally efficien GP varians which can updae he predicion and observaion models in an online fashion, and cope wih he growh in complexiy as he number of daa poins increases over ime. More specifically, he sparse

2 online GP [] selecs he acive se in a online fashion o efficienly mainain sparse approximaions o he models. Alernaively, he local GP [] reduces he compuaion by imposing local sparsiy. We demonsrae he effeciveness of our approach on a wide variey of moions, and show ha boh approaches perform beer han exising algorihms. In he remainder of he paper we firs presen a review on Bayesian filering and he GPLVM. We hen inroduce our algorihm and show our experimenal evaluaion followed by he conclusions. BACKGROUND In his secion we review Bayesian filering and Gaussian process laen variable models..1 BAYESIAN FILTERING Bayesian filering is a sequenial inference echnique ypically employed o perform sae esimaion in dynamic sysems. Specifically, he goal is o recursively compue he poserior disribuion of he curren hidden sae x given he hisory of observaions y 1: = (y 1,..., y ) up o he curren ime sep p(x y 1: ) p(y x ) p(x x 1 )p(x 1 y 1: 1 )dx 1 where p(x x 1 ) is he predicion model ha represens he sysem dynamics, and p(y x ) is he observaion model ha represens he likelihood of an observaion y given he sae x. One of he mos fundamenal Bayesian filers is he Kalman filer, which is a maximum-a-poseriori esimaor for linear and Gaussian models. Unforunaely, i is ofen no applicable in pracice since mos real dynamical sysems are nonlinear and/or non-gaussian. Two popular exensions for non-linear sysems are he exended Kalman filer (EKF) and he unscened Kalman filer (UKF) []. However, similar o he Kalman filer, he performance of EKF and UKF is poor when he models are mulimodal []. In conras, paricle filers ha are no resriced o linear and Gaussian models have been developed by using sequenial Mone Carlo sampling o represen he underlying poserior p(x y 1: ) []. More specifically, a each ime sep, N p paricles of x are drawn from he predicion model p(x x 1 ), and hen all he paricles are weighed according o he observaion model p(y x ). The poserior p(x y 1: ) is approximaed using hese N p weighed paricles. Finally, he N p paricles are resampled for he nex sep. Unforunaely, he parameric descripion of he dynamic models limis he esimaion accuracy of Bayesian filers. Recenly, a number of GP-based Bayesian filers were proposed by learning he predicion and observaion models using GP regression [1, ]. This is a promising alernaive as GPs are non-parameric and can capure complex mappings. However, raining hese mehods requires access o ground ruh daa before filering. Unforunaely, he inpus of he raining se are he hidden saes which are no always known in real-world applicaions. Two exensions were inroduced o learn he hidden saes of he raining se via a non-linear laen variable model [11] or a sparse pseudo-inpu GP regression []. However, hese mehods require offline learning procedures, which are no able o exploi he incoming daa sreams. In conras, we propose wo non-parameric paricle filers ha are able o exploi he incoming daa o learn beer models in an online fashion.. GAUSSIAN PROCESS DYNAMICAL MODEL The Gaussian Process Laen Variable Model (GPLVM) is a probabilisic dimensionaliy reducion echnique, which places a GP prior on he observaion model [1]. Wang e al. [] proposed he Gaussian Process Dynamical Model (GPDM), which enriches he GPLVM o capure emporal srucure by incorporaing a GP prior over he dynamics in he laen space. Formally, he model is: x = f x (x 1 ) + η x y = f y (x ) + η y where y R Dy represens he observaion and x R Dx he laen sae, wih D y D x. The noise processes are assumed o be Gaussian η x N (, σ xi) and η y N (, σ yi). The nonlinear funcions f i x and f i y have GP priors, i.e., f i x GP(, k x (x, x )) and f i y GP(, k y (x, x )) where k x (, ) and k y (, ) are he kernel funcions. For simpliciy, we denoe he hyperparameers of he kernel funcions by θ. Le x 1:T = (x 1,, x T ) be he laen space coordinaes from ime = 1 o ime = T. GPDM is ypically learned by minimizing he negaive log poserior log(p(x 1:T, θ y 1:T )) wih respec o x 1:T, and θ []. Afer x 1:T and θ are obained, a sandard GP predicion is used o consruc he model p(x x 1, θ, X T ) and p(y x, θ, Y T ) wih daa X T = {(x k 1, x k )} T k= and Y T = {(x k, y k )} T k=1. Tracking ( > T ) is hen performed assuming he model is fixed and can be done using, e.g., a paricle filer as described above. The major drawback of his approach is ha i is no able o adap o new observaions during racking. As shown in our experimenal evaluaion, his resuls in poor performance when he raining se is small.

3 ONLINE GP PARTICLE FILTER In order o solve he above-menioned difficulies in learning and filering wih dynamic sysems, we propose an Online GP Paricle Filer framework o learn and refine he model during racking, i.e., he predicion p(x x 1 ) and observaion p(y x ) models are updaed online in he paricle filering framework. To accoun for mulimodaliy and he significan amoun of uncerainy ha can be presen, we propose o represen he predicion and observaion models by a mixure model. For each mixure componen, we will invesigae wo differen GP varians. Le he predicion and observaion models a 1 be p(x x 1, Θ 1,M ) = 1 RM R M i=1 p(x x 1, Θ i 1,M )(1) p(y x, Θ 1,O ) = 1 RO R O i=1 p(y x, Θ i 1,O ) () where Θ i 1,M and Θi 1,O represens he parameers of he i-h componen, Θ 1,M = {Θ i 1,M }R M i=1 and Θ 1,O = {Θ i 1,O }R O i=1 are he parameers of all componens. A he -h ime sep, we run a sandard paricle filer o obain a number of weighed paricles. The laen space represenaions a ime can be obained by resampling he weighed paricles. Then, we assign each paricle o he mos likely mixure componen of p(x x 1, Θ 1,M ) and p(y x, Θ 1,O ) o capure he muli-modaliy of he predicion and observaion models. Finally, we compue he mean laen saes of he assigned paricles and use his mean sae o updae he corresponding componens parameers, Θ i,m for he predicion (or moion) model and Θi,O for he observaion model. The whole framework is summarized in Algorihm 1. Wha remains is o specify how he parameers of individual componens are represened and updaed (lines 1 and in Algorihm 1). As noed above, we aim o use a GP model for each mixure componen. However, a sandard implemenaion would require O( ) operaions and O( ) memory. As grows linearly over ime, he paricle filer will quickly become oo compuaionally and memory inensive. Thus a primary challenge is how o efficienly updae he GP mixure componens in he predicion and observaion models. In order o efficienly updae Θ i,m and Θi,O in an online manner, we consider wo fas GP-based sraegies: Sparse Online GP (SOGP) and Local GPs (LGP) in which he reducion in memory and/or compuaion is achieved by an online sparsificaion and a local expers mechanism respecively. A deailed review of fas GP approaches can be found in [1, 1]. The specific conens of Θ i,m or Θi,O will vary depending on he mehod used. In he case of SOGP i will conain some compued quaniies and he acive se while for LGP i will simply be he se of all raining poins. While we will Algorihm 1 Online GP-Paricle Filer 1: Iniialize model parameers Θ based on y 1:T : Iniialize paricle se x (1:N P ) T based on y 1:T : for = T + 1 o T do : for i = 1 o N p do : x (i) : ŵ (i) p(x x (i) 1, Θ 1,M ) = p(y x (i), Θ 1,O ) : end for : Normalize weighs w (i) = ŵ (i) /( N p i=1 ŵ(i) ) : Resample paricle se wih probabiliies w (1:Np) 1: for i = 1 o N p do 11: ηm i = arg max j p(x (i) x (i) 1, Θj 1,M ) 1: ηo i = arg max j p(y x (i), Θ j 1,O ) 1: end for 1: for j = 1 o R M do 1: n j 1 = N p i=1 δ(ηi M = j) 1: x j 1 = 1 Np n j i=1 δ(ηi M = j)x(i) 1 1 1: x j = 1 Np n j i=1 δ(ηi M = j)x(i) 1 1: Updae Θ j,m wih ( xj 1, xj ) 1: end for : for j = 1 o R O do 1: n j 1 = N p i=1 δ(ηi O = j) : x j = 1 Np n j i=1 δ(ηi O = j)x(i) 1 : Updae Θ j,o wih ( xj, y ) : end for : end for focus on hese wo sraegies, we noe ha in principle any similar updae sraegy could be used insead, such as informaive vecor machines [1] or local regression approaches [,, ]. In wha follows, o avoid confusion wih he noaions of he laen sae and observaion, we will use a and b o indicae he inpu and oupu when we describe SOGP and LGP regression in which we consider modeling a generic funcion b = f(a) + ξ, wih ξ N (, σ I)..1 SPARSE ONLINE GAUSSIAN PROCESS The Sparse Online Gaussian Process (SOGP) of [, ] is a well-known algorihm for online learning of GP models. To cope wih he fac ha daa arrives in an online manner, SOGP rains a GP model sequenially by updaing he poserior mean and covariance of he laen funcion values of he raining se. This online procedure is coupled wih a sparsificaion sraegy which ieraively selecs a fixed-size subse of poins o form he acive se, prevening he oherwise unbounded growh of he compuaion and memory load. The key of SOGP is o mainain he join poserior over he laen funcion values of he fixed-size acive se D 1, i.e.,

4 N (µ 1, Σ 1 ), by recursively updaing µ 1 and Σ 1. When a new observaion (a, b ) 1 is available, we perform he following updae o ake he new daa poin ino accoun [] q = Q 1 k 1 (a ) () ρ = k(a, a ) k 1 (a ) T Q 1 k 1 (a ) () ˆσ = σ + ρ + q T Σ 1 q () [ ] Σ δ = 1 q ρ + q T () Σ 1 q [ ] µ 1 µ = q T + ˆσ µ (b q T µ 1 )δ () 1 [ ] Σ 1 Σ Σ = 1 q q T Σ 1 ρ + q T ˆσ Σ 1 q δ δ T () where k 1 (a ) is he kernel vecor which is consruced from a and he acive se D 1, and Q 1 is he inverse kernel marix of he acive se D 1. One of he key seps in his algorihm is how o decide when o add he new poin o he acive se. We employed he sraegy suggesed by [, ], and ignore he new poin wih ρ < ɛ for some small value of ɛ (we use ɛ = 1 ). In his case, he µ, Σ are updaed as µ [µ ] i, Σ [Σ ] i, i where i = is he index of he new poin, [ ] i removes he i-h enry of a vecor, and [ ] i, i removes he i-h row and column of a marix. Addiionally, he inverse kernel marix is simply Q = Q 1 because he new poin is no included in he acive se. When ρ ɛ, we add he new poin o he acive se D = D 1 {(a, b )}. The µ, Σ are hen he same as Eq. () and (), and he inverse kernel marix is updaed o be [] Q = [ Q 1 ] + ρ [ q q T ] q q T 1 When he size of he acive se is larger han he fixed size N A because a new poin was added, we mus remove a poin. This is done by selecing he one which minimally affecs he predicions according o he squared prediced error. Following [, ], we remove he j-h daa poin wih j = arg min j () ( ) [Q µ ] j (1) [Q ] j,j where [ ] j selecs he j-h enry of a vecor and [ ] j,j selec he jh diagonal enry of a marix. Once a poin has been seleced for removal, µ, Σ and Q are updaed as µ [µ ] j (11) Σ [Σ ] j, j (1) Q [Q ] j, j [Q ] j,j [Q ] T j,j [Q ] j,j (1) 1 For simpliciy of presenaion, we assume ha b is a scalar. The exension o vecor valued b is sraighforward. Algorihm SOGP Updae inpu Previous poserior quaniies µ 1, Σ 1, Q 1 inpu Previous acive se D 1 inpu New inpu-oupu observaion (a, b ) pair 1: Compue ρ, µ and Σ as in Equaions (), () and (). : if ρ < ɛ hen : Perform updae µ [µ ] i, Σ [Σ ] i, i where i is he index of he newly added row o µ. : Se Q = Q 1, D = D 1. : else : Compue Q as in Equaion (). : Add o acive se D = D 1 {(a, b )}. : end if : if D > N A hen 1: Selec poin j o remove using Equaion (1). 11: Perform updae µ [µ ] j, Σ [Σ ] j, j and Q [Q ] j, j [Q] j,j[q]t j,j [Q ] j,j. 1: Remove j from acive se D D \ {(a j, b j )}. 1: end if oupu µ, Σ, Q and D where [ ] j,j selecs he j-h column of he marix wih he j-h row removed and he poin is removed from he acive se D D \ {(a j, b j )}. The join poserior a ime can be used o consruc he predicive disribuion for a new inpu a p SOGP (b a, D, Θ) = N (b b, σ ) (1) where b = k (a ) T Q µ and σ = σ + k(a, a ) + k (a ) T (Q Σ Q Q )k (a ). We summarize he SOGP updaes in Algorihm.. LOCAL GAUSSIAN PROCESSES l=1 An alernaive o he SOGP approach is o use Local Gaussian Processes (LGP), which was developed specifically o deal wih large, muli-modal regression problems [1, ]. In LGP, given a es inpu a and a se of inpu-oupu pairs D = {(a i, b i )} N i=1 }, he M a-neares neighbors D a = {(a l, b l )} Ma are seleced based on he disance in he inpu space d l = a l a. Then, for each of he M a neighbors, M b -neares neighbors D bl = {(a j, b j )} M b j=1 are seleced based on he disance in he oupu space o b l. These neighbors are hen combined o form a local GP exper which makes a Gaussian predicion wih mean and covariance µ l = B Dbl K 1 D bl,d bl k Dbl (a ) σ l = k(a, a ) k Dbl (a ) T K 1 D bl,d bl k Dbl (a ) + σ where B Dbl is he marix whose columns are he M b neares neighbors of b l, k Dbl (a ) is he vecor of kernel funcion values for he inpu a and he poins in D bl, and

5 K Dbl,D bl is he kernel marix for he poins in D bl. The final predicive disribuion is hen formed by combining all local expers in a mixure model M a p LGP (b a, D, Θ) = w l N (b µ l, σl I) (1) wih weighs w l 1/d l. l=1 EXPERIMENTAL EVALUATION To illusrae our approach we choose very differen moions, i.e., walking, golf swing, swimming as well as exercises (composed of side wis and squa). The daa consiss of moion capure from he CMU daase [], where each observaion is a dimensional vecor conaining he D roaions of all joins. We normalize he daa o be mean zero and subsample he observaions o reduce he correlaion beween consecuive frames. We use a frequency of 1 frames/s for walking and swimming, frames/s for golf swing and frames/s for he exercise moion. We compue all resuls averaged over rials and repor he average roo mean squared error as our measure of performance. In all he experimens, he laen space dimensionaliy is se o be as is common for human moion models []. We use PCA o iniialize he laen space and K-means o obain he daa poins used for he mixure componens. We choose he compound kernel funcion k(x, x ) = σf exp(. x x /γ ) + l x T x for boh predicion and observaion mappings. Unless oherwise saed, we use paricles, a raining se of size of /// and /// mixure componens for walking/golf/swimming/exercise moions respecively. For LGP, he number of local GP expers is ///, and he size of each local exper is ///. For SOGP, he size of he acive se is ///. The parameer values were chosen o balance compuaional cos wih he predicion accuracy and in our experimens we demonsrae he robusness of our approach o hese parameers..1 COMPARISON TO STATE-OF-THE-ART We compare our approaches o wo baselines: The firs one is he approach of Ko and Fox [11] where a GPDM is learned offline wih gradien descen [] before performing paricle filering for sae esimaion. The second baseline is similar, bu learns he GPDM offline using sochasic gradien descen []. We esed he baselines in wo differen seings. Firs, only he iniial raining se is available o learn he predicion and observaion models. Second, all he daa (including fuure sreamed examples) are used o learn he predicion and observaion models. Noe ha he laer represens he oracle for Ko and Fox [11]. Number of Paricles: We evaluae how he accuracy changes as a funcion of he number of paricles, N p. As expeced, he predicion error is reduced in all he mehods when he number of paricles increases. As shown in he firs row of Fig. 1, our approaches are superior o he baselines. Imporanly, we ouperformed he oracle baseline as we are able o represen muli-modal disribuions effecively. This is paricularly imporan in he exercise sequence as he dynamics are clearly mulimodal due o he differen moions ha are performed in he sequence. Furhermore, our LGP varian ouperforms SOGP. We believe his is due o he fac ha SOGP has a fixed capaciy while LGP is able o leverage more raining daa when making predicions. Influence of noise: In his experimen we evaluae he robusness of all approaches o addiive noise in he observaions. The second row of Fig. 1 shows ha our LGP paricle filer significanly ouperforms he baselines, paricularly in he exercise sequence. Our SOGP ouperforms all baselines ha have access o he same raining se, and is only beaen by he oracle for walking. Size of Training Se: We nex evaluae how he accuracy depends on he size of he inial raining se, T. The firs row of Fig. clearly indicaes ha our mehods perform well even when he raining se is very small. In conras, he wo baselines require bigger raining ses o achieve comparable performance. This is expeced as he baselines do no updae he laen space o ake he incoming observaions ino accoun.. QUALITATIVE EXPERIMENTS Fig. shows he laen space of boh SOGP and LGP filers when employing paricles for each ime sep (depiced in blue). From he D laen space and prediced skeleons, we find ha he manifolds of boh LGP and SOGP paricle filers have a good represenaion of he high-dimensional human moion daa.. PROPERTIES OF OUR METHODS We nex discuss various aspecs of our mehod and evaluae he influence of he parameers of SOGP and LGP filers. For LGP, due o he fac ha he daa sizes of walking, golf and swimming moions are small, we reduced he number (size) of local expers o be able o increase he size (number) of he local expers. Compuaional Complexiy: Overall he compuaional complexiy of our mehod (Alg 1) is mainly deermined by he complexiy of consrucing a predicion disribuion for each componens (lines - and 11-1) and model updaes (line 1 and line ). Specifically, for an individual componen which is eiher SOGP or LGP, compuing he predicion disribuion is O(N A ) or O(M am b + T M am b ) respecively where N A is he size of acive se, M a is he number of local expers, M b are he number of neigh-

6 GPDM+PF GPDM(Oracle)+PF Sochasic GPDM+PF Sochasic GPDM(Oracle)+PF Our SOGP+PF Our LGP+PF Number of Paricles 1. 1 Number of Paricles. Number of Paricles 1 Number of Paricles Sandard Deviaion of Noise 1 1. Sandard Deviaion of Noise.. Sandard Deviaion of Noise 1 Sandard Deviaion of Noise Figure 1: Roo mean squared error as a funcion of (1s row) he number of paricles in he paricle filer, (nd row) he sandard deviaion of noise added o he observaions. The columns (from lef o righ) represen walking, golf, swimming and exercise moions. Noe ha our approach ouperforms he baselines in all seings. Handling mulimodal disribuions is paricularly imporan in he exercise example as i is composed of a variey of differen moions. GPDM+PF Sochasic GPDM+PF Our SOGP+PF Our LGP+PF Size of Iniial daa 1. Size of Iniial Daa Size of Iniial Daa Size of Iniial Daa Missing Dimension Missing Dimension 1 1 Missing Dimension 1 1 Missing Dimension Figure : Roo Mean Squared Error as a funcion of (1s row) he number of iniial raining poins, (nd row) he number of he missing dimensions. The columns (from lef o righ) are respecively for walking, golf, swimming and exercise moions. bors in he oupu space and T M a M b comes from he KNN search. The model updaes for he mixure componens (lines 1 and ) have a compuaional complexiy of O(NA ) and O(1) for SOGP and LGP respecively.

7 (a) D Walk (b) D Golf (c) D Swim (d) D Exercise (e) D Walk (f) D Golf (g) D Swim (h) D Exercise (i) Walk (=,, ) (j) Golf (=, 1, ) (k) Swim (=1,) (l) Exercise (=,,1) Figure : D laen spaces learned while racking: The firs row depics he resuls of our SOGP varian, while he second row shows our LGP varian. In he walk/golf/swimming plos he red curve represens he prediced mean of he laen sae sequence and he blue crosses are he paricles a each sep. In he plos for exercise (las column), he red, blue and green curves are he prediced mean of he laen sae for hree moions in he exercise sequence, and he black crosses are he paricles a each sep. The hird row depics he prediced skeleons, where ground ruh is shown in green, our SOGP varian in blue and our LGP varian in red. Our SOGP+PF Our SOGP(no updae)+pf Our LGP+PF Our LGP(no updae)+pf Number of Mixure Componens 1. 1 Number of Mixure Componens. 1 Number of Mixure Componens 1 Number of Mixure Componens Figure : Roo Mean Squared Error as a funcion of he number of mixure componens. The columns (from lef o righ) are respecively for walking, golf, swimming and exercise moions. Number of Mixure Componens: Fig. shows performance as a funcion of he number of mixure componens, R M and R O, for boh SOGP and LGP. For LGP+PF in walk/golf/swim/exercise, he number of local GPs are 1/1// and he size of each local GP are ///. In all cases, we se R M = R O. Noe ha performance ypically increases wih he number of mixure componens, for SOGP, bu less so for LGP. Furhermore, our approaches ouperform he baselines in which he model is no updaed during filering indicaing ha he online model updaing is very imporan in pracice. Also noe ha while LGP generally ouperforms SOGP, he difference quickly declines

8 Walk Golf Swim Exercise. 1 1 (a) Size of Acive Daa (b) Number of Local GPs (c) Size of Local GPs Figure : Roo mean squared error as a funcion of he size of he acive se in SOGP, he number of he local GP expers and he size of each local GP exper in LGP. In he subplo (c), he op x-axis is for exercise and he boom one for he oher moions. Figure : Prediced skeleon for missing pars (Walk: wo legs; Golf, Swim and Exercise: lef arm). The ground ruh is shown in green, our SOGP paricle filer in blue and LGP paricle filer in red. We show he prediced performance a =,, for walk, =,, for golf, =1, for swim, =,, 11 for exercise. as he number of mixure componens increases. This suggess ha, when he fixed memory requiremens of SOGP is desirable, a larger number of mixure componens will achieve performance comparable o LGP. Acive Se size in SOGP: To explore he effec of he size of he acive se, N A, on performance we se he number of mixure componens, R M and R O, o be /1// for walk/golf/swim/exercise, and use he same seings as before for he oher parameers. Resuls are shown in Fig. (a). As expeced performance improves when he size of he acive se increases. Number and Size of Local Expers in LGP: Figs. (b) and (c) show he performance of our approach as a funcion of he number of local GP expers, M a, as well as heir size, M b. For his experimen we se he number of mixure componens, R M and R O, o be 1//1/ and used he same seings as before for he oher parameers excep when evaluaing he size of each local GP exper where we se he number of local GP expers o ///. As shown in he figure, even wih he small number (size) of local GP expers, we sill achieve good performance.. HANDLING MISSING DATA In his seing, we evaluae he capabiliies of our approaches o handle missing daa. We assume ha he iniial se has no missing values, bu a fixed se of join angles are missing for all incoming frames. Our approach is able o cope wih missing daa wih only wo small modificaions. Firs, paricles are weighed only based on he observed dimensions. Furhermore, when updaing he predicion and observaion models, we employ mean impuaion for he missing observaion dimensions. Fig. shows reconsrucions of he missing dimensions for all our moions, which consiss of he wo legs for walking, he lef arm for golf swing, swimming and exercise moions. We can see ha our approach is able o reconsruc he missing pars well. Finally, o evaluae he racking performance as a funcion of he number of missing dimensions, we randomly generae he indices for he missing dimensions and use he same missing dimensions for all incoming frames. The second row of Fig. shows ha, compared o he baselines, our mehods perform well even when he number of missing dimensions is (1/ of he skeleon) for all he moions. In addiion, our LGP paricle filer ouperforms our SOGP varian. CONCLUSION In his paper we have presened a novel non-parameric approach o Bayesian filering, where he observaion and predicion models are consruced using a mixure model wih GP componens learned in an online fashion. We have demonsraed ha our approach can capure he mul-

9 imodaliy accuraely and efficienly by online updaes. We have explored wo fas GP varians for updaing which keep memory and compuaion bounded for individual mixure componens. We have demonsraed he effeciveness of our approach when racking differen human moions and explored he impac of various parameers on performance. The Local GP paricle filer proved superior o our SOGP varian, however hese differences can be miigaed by using more mixure componens when using SOGP. In he fuure, we plan o invesigae he usefulness of our approach in oher seings such as shape deformaion esimaion and financial ime series. References [1] K. Chalupka, C. K. I. Williams, and I. Murray. A Framework for Evaluaing Approximaion Mehods for Gaussian Process Regression. JMLR, 1. [] CMU Mocap Daabase. CMU Mocap Daabase. hp://mocap.cs.cmu.edu/. [] L. Csao and M. Opper. Sparse online gaussian processes. In Neural Compuaion,. [] M. P. Deisenroh, M. F. Huber, and U. D. Hanebeck. Analyic momen-based gaussian process filering. In ICML,. [] A. Douce, S. Godsill, and C. Andrieu. On sequenial mone carlo sampling mehods for bayesian filering. Saisics and Compuing,. [] T. Hasie and C. Loader. Local Regression: Auomaic Kernel Carpenry. Saisical Science, 1. [] R. A. Jacobs, M. I. Jordan, S. J. Nowlan, and G. E. Hinon. Adapive Mixures of Local Expers. Neural Compuaion, 11. [] O. C. Jenkins and M. Maarić. A spaio-emporal exension o Isomap nonlinear dimension reducion. In ICML,. [] S. J. Julier, J. K. Uhlmann, and H. F. Durran-Whye. A new approach for filering nonlinear sysems. In American Conrol Conference, 1. [1] J. Ko and D. Fox. Gp-bayesfilers: Bayesian filering using gaussian process predicion and observaion models. In IROS,. [11] J. Ko and D. Fox. Learning gp-bayesfilers via gaussian process laen variable models. In RSS,. [1] N. Lawrence. Probabilisic non-linear principal componen analysis wih gaussian process laen variable models. JMLR, :1 11,. [1] N. Lawrence, M. Seeger, and R. Herbrich. Fas Sparse Gaussian Process Mehods: The Informaive Vecor Machine. In NIPS,. [1] C-S. Lee and A. Elgammal. Coupled Visual and Kinemaics Manifold Models for Human Moion Analysis. IJCV, 1. [1] S. Levine, J. Wang, A. Haraux, Z. Popovic, and V. Kolun. Coninuous characer conrol wih lowdimensional embeddings. In SIGGRAPH, 1. [1] K. Moon and V. Pavlovic. Impac of dynamics on subspace embedding and racking of sequences. In CVPR, pages 1,. [1] D. Nguyen-Tuong, J. Peers, and M. Seeger. Local gaussian process regression for real ime online model learning and conrol. In NIPS,. [1] J. Quiñonero-Candela and C. E. Rasmussen. A Unifying View of Sparse Approximae Gaussian Process Regression. JMLR,. [1] S. Roweis and L. Saul. Nonlinear Dimensionaliy Reducion by Locally Linear Embedding. Science, ():,. [] S. Schaal and C. G. Akeson. Consrucive Incremenal Learning From Only Local Informaion. Neural Compuaion, 1. [1] J. Tenenbaum, V. de Silva, and J. Langford. A Global Geomeric Framework for Nonlinear Dimensionaliy Reducion. Science,. [] R. Turner, M. P. Deisenroh, and C. E. Rasmussen. Sae-space inference and learning wih gaussian processes. In AISTATS, 1. [] R. Urasun and T. Darrell. Sparse probabilisic regression for aciviy-independen human pose inference. In CVPR,. [] R. Urasun, D. Flee, A. Herzman, and P. Fua. Priors for people racking from small raining ses. In ICCV,. [] R. Urasun, D. Flee, and P. Fua. d people racking wih gaussian process dynamical models. In CVPR,. [] R. Urasun, D Flee, A. Geiger, J. Popovic, T. Darrell, and N. Lawrence. Topologically-consrained laen variable models. In ICML,. [] S. Van Vaerenbergh, M. Lázaro-Gredilla, and I. Sanamaría. Kernel recursive leas-squares racker for ime-varying regression. IEEE Transacions on Neural Neworks and Learning Sysems, 1. [] J. Wang, D. Flee, and A. Herzmann. Gaussian process dynamical models for human moion. PAMI, ():,. ISSN 1-. [] A. Yao, J. Gall, L. V. Gool, and R. Urasun. Learning probabilisic non-linear laen variable models for racking complex aciviies. In NIPS, 11.

Real-time Particle Filters

Real-time Particle Filters Real-ime Paricle Filers Cody Kwok Dieer Fox Marina Meilă Dep. of Compuer Science & Engineering, Dep. of Saisics Universiy of Washingon Seale, WA 9895 ckwok,fox @cs.washingon.edu, mmp@sa.washingon.edu Absrac

More information

Maintaining Multi-Modality through Mixture Tracking

Maintaining Multi-Modality through Mixture Tracking Mainaining Muli-Modaliy hrough Mixure Tracking Jaco Vermaak, Arnaud Douce Cambridge Universiy Engineering Deparmen Cambridge, CB2 1PZ, UK Parick Pérez Microsof Research Cambridge, CB3 0FB, UK Absrac In

More information

An Online Learning-based Framework for Tracking

An Online Learning-based Framework for Tracking An Online Learning-based Framework for Tracking Kamalika Chaudhuri Compuer Science and Engineering Universiy of California, San Diego La Jolla, CA 9293 Yoav Freund Compuer Science and Engineering Universiy

More information

Journal Of Business & Economics Research September 2005 Volume 3, Number 9

Journal Of Business & Economics Research September 2005 Volume 3, Number 9 Opion Pricing And Mone Carlo Simulaions George M. Jabbour, (Email: jabbour@gwu.edu), George Washingon Universiy Yi-Kang Liu, (yikang@gwu.edu), George Washingon Universiy ABSTRACT The advanage of Mone Carlo

More information

SELF-EVALUATION FOR VIDEO TRACKING SYSTEMS

SELF-EVALUATION FOR VIDEO TRACKING SYSTEMS SELF-EVALUATION FOR VIDEO TRACKING SYSTEMS Hao Wu and Qinfen Zheng Cenre for Auomaion Research Dep. of Elecrical and Compuer Engineering Universiy of Maryland, College Park, MD-20742 {wh2003, qinfen}@cfar.umd.edu

More information

Principal components of stock market dynamics. Methodology and applications in brief (to be updated ) Andrei Bouzaev, bouzaev@ya.

Principal components of stock market dynamics. Methodology and applications in brief (to be updated ) Andrei Bouzaev, bouzaev@ya. Principal componens of sock marke dynamics Mehodology and applicaions in brief o be updaed Andrei Bouzaev, bouzaev@ya.ru Why principal componens are needed Objecives undersand he evidence of more han one

More information

SPEC model selection algorithm for ARCH models: an options pricing evaluation framework

SPEC model selection algorithm for ARCH models: an options pricing evaluation framework Applied Financial Economics Leers, 2008, 4, 419 423 SEC model selecion algorihm for ARCH models: an opions pricing evaluaion framework Savros Degiannakis a, * and Evdokia Xekalaki a,b a Deparmen of Saisics,

More information

A Bayesian framework with auxiliary particle filter for GMTI based ground vehicle tracking aided by domain knowledge

A Bayesian framework with auxiliary particle filter for GMTI based ground vehicle tracking aided by domain knowledge A Bayesian framework wih auxiliary paricle filer for GMTI based ground vehicle racking aided by domain knowledge Miao Yu a, Cunjia Liu a, Wen-hua Chen a and Jonahon Chambers b a Deparmen of Aeronauical

More information

How To Predict A Person'S Behavior

How To Predict A Person'S Behavior Informaion Theoreic Approaches for Predicive Models: Resuls and Analysis Monica Dinculescu Supervised by Doina Precup Absrac Learning he inernal represenaion of parially observable environmens has proven

More information

A Natural Feature-Based 3D Object Tracking Method for Wearable Augmented Reality

A Natural Feature-Based 3D Object Tracking Method for Wearable Augmented Reality A Naural Feaure-Based 3D Objec Tracking Mehod for Wearable Augmened Realiy Takashi Okuma Columbia Universiy / AIST Email: okuma@cs.columbia.edu Takeshi Kuraa Universiy of Washingon / AIST Email: kuraa@ieee.org

More information

Improving Unreliable Mobile GIS with Swarm-based Particle Filters

Improving Unreliable Mobile GIS with Swarm-based Particle Filters Improving Unreliable Mobile GIS wih Swarm-based Paricle Filers Fama Hrizi, Jérôme Härri, Chrisian Bonne EURECOM, Mobile Communicaions Deparmen Campus SophiaTech, 450 Roue des Chappes Bio, France {hrizi,

More information

Improved Techniques for Grid Mapping with Rao-Blackwellized Particle Filters

Improved Techniques for Grid Mapping with Rao-Blackwellized Particle Filters 1 Improved Techniques for Grid Mapping wih Rao-Blackwellized Paricle Filers Giorgio Grisei Cyrill Sachniss Wolfram Burgard Universiy of Freiburg, Dep. of Compuer Science, Georges-Köhler-Allee 79, D-79110

More information

PROFIT TEST MODELLING IN LIFE ASSURANCE USING SPREADSHEETS PART ONE

PROFIT TEST MODELLING IN LIFE ASSURANCE USING SPREADSHEETS PART ONE Profi Tes Modelling in Life Assurance Using Spreadshees PROFIT TEST MODELLING IN LIFE ASSURANCE USING SPREADSHEETS PART ONE Erik Alm Peer Millingon 2004 Profi Tes Modelling in Life Assurance Using Spreadshees

More information

DYNAMIC MODELS FOR VALUATION OF WRONGFUL DEATH PAYMENTS

DYNAMIC MODELS FOR VALUATION OF WRONGFUL DEATH PAYMENTS DYNAMIC MODELS FOR VALUATION OF WRONGFUL DEATH PAYMENTS Hong Mao, Shanghai Second Polyechnic Universiy Krzyszof M. Osaszewski, Illinois Sae Universiy Youyu Zhang, Fudan Universiy ABSTRACT Liigaion, exper

More information

Vector Autoregressions (VARs): Operational Perspectives

Vector Autoregressions (VARs): Operational Perspectives Vecor Auoregressions (VARs): Operaional Perspecives Primary Source: Sock, James H., and Mark W. Wason, Vecor Auoregressions, Journal of Economic Perspecives, Vol. 15 No. 4 (Fall 2001), 101-115. Macroeconomericians

More information

USE OF EDUCATION TECHNOLOGY IN ENGLISH CLASSES

USE OF EDUCATION TECHNOLOGY IN ENGLISH CLASSES USE OF EDUCATION TECHNOLOGY IN ENGLISH CLASSES Mehme Nuri GÖMLEKSİZ Absrac Using educaion echnology in classes helps eachers realize a beer and more effecive learning. In his sudy 150 English eachers were

More information

Chapter 8: Regression with Lagged Explanatory Variables

Chapter 8: Regression with Lagged Explanatory Variables Chaper 8: Regression wih Lagged Explanaory Variables Time series daa: Y for =1,..,T End goal: Regression model relaing a dependen variable o explanaory variables. Wih ime series new issues arise: 1. One

More information

Particle Filtering for Geometric Active Contours with Application to Tracking Moving and Deforming Objects

Particle Filtering for Geometric Active Contours with Application to Tracking Moving and Deforming Objects Paricle Filering for Geomeric Acive Conours wih Applicaion o Tracking Moving and Deforming Objecs Yogesh Rahi Namraa Vaswani Allen Tannenbaum Anhony Yezzi Georgia Insiue of Technology School of Elecrical

More information

Multiprocessor Systems-on-Chips

Multiprocessor Systems-on-Chips Par of: Muliprocessor Sysems-on-Chips Edied by: Ahmed Amine Jerraya and Wayne Wolf Morgan Kaufmann Publishers, 2005 2 Modeling Shared Resources Conex swiching implies overhead. On a processing elemen,

More information

Biology at Home - Pariion Funcion Guillaume

Biology at Home - Pariion Funcion Guillaume On Tracking The Pariion Funcion Guillaume Desjardins, Aaron Courville, Yoshua Bengio {desjagui,courvila,bengioy}@iro.umonreal.ca Déparemen d informaique e de recherche opéraionnelle Universié de Monréal

More information

A Note on Using the Svensson procedure to estimate the risk free rate in corporate valuation

A Note on Using the Svensson procedure to estimate the risk free rate in corporate valuation A Noe on Using he Svensson procedure o esimae he risk free rae in corporae valuaion By Sven Arnold, Alexander Lahmann and Bernhard Schwezler Ocober 2011 1. The risk free ineres rae in corporae valuaion

More information

The Transport Equation

The Transport Equation The Transpor Equaion Consider a fluid, flowing wih velociy, V, in a hin sraigh ube whose cross secion will be denoed by A. Suppose he fluid conains a conaminan whose concenraion a posiion a ime will be

More information

The naive method discussed in Lecture 1 uses the most recent observations to forecast future values. That is, Y ˆ t + 1

The naive method discussed in Lecture 1 uses the most recent observations to forecast future values. That is, Y ˆ t + 1 Business Condiions & Forecasing Exponenial Smoohing LECTURE 2 MOVING AVERAGES AND EXPONENTIAL SMOOTHING OVERVIEW This lecure inroduces ime-series smoohing forecasing mehods. Various models are discussed,

More information

Duration and Convexity ( ) 20 = Bond B has a maturity of 5 years and also has a required rate of return of 10%. Its price is $613.

Duration and Convexity ( ) 20 = Bond B has a maturity of 5 years and also has a required rate of return of 10%. Its price is $613. Graduae School of Business Adminisraion Universiy of Virginia UVA-F-38 Duraion and Convexiy he price of a bond is a funcion of he promised paymens and he marke required rae of reurn. Since he promised

More information

TEMPORAL PATTERN IDENTIFICATION OF TIME SERIES DATA USING PATTERN WAVELETS AND GENETIC ALGORITHMS

TEMPORAL PATTERN IDENTIFICATION OF TIME SERIES DATA USING PATTERN WAVELETS AND GENETIC ALGORITHMS TEMPORAL PATTERN IDENTIFICATION OF TIME SERIES DATA USING PATTERN WAVELETS AND GENETIC ALGORITHMS RICHARD J. POVINELLI AND XIN FENG Deparmen of Elecrical and Compuer Engineering Marquee Universiy, P.O.

More information

Combination of UWB and GPS for indoor-outdoor vehicle localization

Combination of UWB and GPS for indoor-outdoor vehicle localization ombinaion of UW and for indoor-oudoor vehicle localizaion González J., lanco J.L., Galindo., Oriz-de-Galiseo., Fernández-Madrigal J.., Moreno F.., and Marínez J.L. {jgonzalez, jlblanco,cipriano,jafma}@cima.uma.es,

More information

Information Theoretic Evaluation of Change Prediction Models for Large-Scale Software

Information Theoretic Evaluation of Change Prediction Models for Large-Scale Software Informaion Theoreic Evaluaion of Change Predicion Models for Large-Scale Sofware Mina Askari School of Compuer Science Universiy of Waerloo Waerloo, Canada maskari@uwaerloo.ca Ric Hol School of Compuer

More information

Term Structure of Prices of Asian Options

Term Structure of Prices of Asian Options Term Srucure of Prices of Asian Opions Jirô Akahori, Tsuomu Mikami, Kenji Yasuomi and Teruo Yokoa Dep. of Mahemaical Sciences, Risumeikan Universiy 1-1-1 Nojihigashi, Kusasu, Shiga 525-8577, Japan E-mail:

More information

Analogue and Digital Signal Processing. First Term Third Year CS Engineering By Dr Mukhtiar Ali Unar

Analogue and Digital Signal Processing. First Term Third Year CS Engineering By Dr Mukhtiar Ali Unar Analogue and Digial Signal Processing Firs Term Third Year CS Engineering By Dr Mukhiar Ali Unar Recommended Books Haykin S. and Van Veen B.; Signals and Sysems, John Wiley& Sons Inc. ISBN: 0-7-380-7 Ifeachor

More information

Random Scanning Algorithm for Tracking Curves in Binary Image Sequences

Random Scanning Algorithm for Tracking Curves in Binary Image Sequences Vol., No., Page 101 of 110 Copyrigh 008, TSI Press Prined in he USA. All righs reserved Random Scanning Algorihm for Tracking Curves in Binary Image Sequences Kazuhiko Kawamoo *1 and Kaoru Hiroa 1 Kyushu

More information

Random Walk in 1-D. 3 possible paths x vs n. -5 For our random walk, we assume the probabilities p,q do not depend on time (n) - stationary

Random Walk in 1-D. 3 possible paths x vs n. -5 For our random walk, we assume the probabilities p,q do not depend on time (n) - stationary Random Walk in -D Random walks appear in many cones: diffusion is a random walk process undersanding buffering, waiing imes, queuing more generally he heory of sochasic processes gambling choosing he bes

More information

Hotel Room Demand Forecasting via Observed Reservation Information

Hotel Room Demand Forecasting via Observed Reservation Information Proceedings of he Asia Pacific Indusrial Engineering & Managemen Sysems Conference 0 V. Kachivichyanuul, H.T. Luong, and R. Piaaso Eds. Hoel Room Demand Forecasing via Observed Reservaion Informaion aragain

More information

A PROPOSAL TO OBTAIN A LONG QUARTERLY CHILEAN GDP SERIES *

A PROPOSAL TO OBTAIN A LONG QUARTERLY CHILEAN GDP SERIES * CUADERNOS DE ECONOMÍA, VOL. 43 (NOVIEMBRE), PP. 285-299, 2006 A PROPOSAL TO OBTAIN A LONG QUARTERLY CHILEAN GDP SERIES * JUAN DE DIOS TENA Universidad de Concepción y Universidad Carlos III, España MIGUEL

More information

MACROECONOMIC FORECASTS AT THE MOF A LOOK INTO THE REAR VIEW MIRROR

MACROECONOMIC FORECASTS AT THE MOF A LOOK INTO THE REAR VIEW MIRROR MACROECONOMIC FORECASTS AT THE MOF A LOOK INTO THE REAR VIEW MIRROR The firs experimenal publicaion, which summarised pas and expeced fuure developmen of basic economic indicaors, was published by he Minisry

More information

Predicting Stock Market Index Trading Signals Using Neural Networks

Predicting Stock Market Index Trading Signals Using Neural Networks Predicing Sock Marke Index Trading Using Neural Neworks C. D. Tilakarane, S. A. Morris, M. A. Mammadov, C. P. Hurs Cenre for Informaics and Applied Opimizaion School of Informaion Technology and Mahemaical

More information

Market Liquidity and the Impacts of the Computerized Trading System: Evidence from the Stock Exchange of Thailand

Market Liquidity and the Impacts of the Computerized Trading System: Evidence from the Stock Exchange of Thailand 36 Invesmen Managemen and Financial Innovaions, 4/4 Marke Liquidiy and he Impacs of he Compuerized Trading Sysem: Evidence from he Sock Exchange of Thailand Sorasar Sukcharoensin 1, Pariyada Srisopisawa,

More information

Measuring macroeconomic volatility Applications to export revenue data, 1970-2005

Measuring macroeconomic volatility Applications to export revenue data, 1970-2005 FONDATION POUR LES ETUDES ET RERS LE DEVELOPPEMENT INTERNATIONAL Measuring macroeconomic volailiy Applicaions o expor revenue daa, 1970-005 by Joël Cariolle Policy brief no. 47 March 01 The FERDI is a

More information

Single-machine Scheduling with Periodic Maintenance and both Preemptive and. Non-preemptive jobs in Remanufacturing System 1

Single-machine Scheduling with Periodic Maintenance and both Preemptive and. Non-preemptive jobs in Remanufacturing System 1 Absrac number: 05-0407 Single-machine Scheduling wih Periodic Mainenance and boh Preempive and Non-preempive jobs in Remanufacuring Sysem Liu Biyu hen Weida (School of Economics and Managemen Souheas Universiy

More information

Mathematics in Pharmacokinetics What and Why (A second attempt to make it clearer)

Mathematics in Pharmacokinetics What and Why (A second attempt to make it clearer) Mahemaics in Pharmacokineics Wha and Why (A second aemp o make i clearer) We have used equaions for concenraion () as a funcion of ime (). We will coninue o use hese equaions since he plasma concenraions

More information

Evolutionary building of stock trading experts in real-time systems

Evolutionary building of stock trading experts in real-time systems Evoluionary building of sock rading expers in real-ime sysems Jerzy J. Korczak Universié Louis Paseur Srasbourg, France Email: jjk@dp-info.u-srasbg.fr Absrac: This paper addresses he problem of consrucing

More information

Measuring the Effects of Monetary Policy: A Factor-Augmented Vector Autoregressive (FAVAR) Approach * Ben S. Bernanke, Federal Reserve Board

Measuring the Effects of Monetary Policy: A Factor-Augmented Vector Autoregressive (FAVAR) Approach * Ben S. Bernanke, Federal Reserve Board Measuring he Effecs of Moneary Policy: A acor-augmened Vecor Auoregressive (AVAR) Approach * Ben S. Bernanke, ederal Reserve Board Jean Boivin, Columbia Universiy and NBER Pior Eliasz, Princeon Universiy

More information

Making a Faster Cryptanalytic Time-Memory Trade-Off

Making a Faster Cryptanalytic Time-Memory Trade-Off Making a Faser Crypanalyic Time-Memory Trade-Off Philippe Oechslin Laboraoire de Securié e de Crypographie (LASEC) Ecole Polyechnique Fédérale de Lausanne Faculé I&C, 1015 Lausanne, Swizerland philippe.oechslin@epfl.ch

More information

A Distributed Multiple-Target Identity Management Algorithm in Sensor Networks

A Distributed Multiple-Target Identity Management Algorithm in Sensor Networks A Disribued Muliple-Targe Ideniy Managemen Algorihm in Sensor Neworks Inseok Hwang, Kaushik Roy, Hamsa Balakrishnan, and Claire Tomlin Dep. of Aeronauics and Asronauics, Sanford Universiy, CA 94305 Elecrical

More information

ANALYSIS AND COMPARISONS OF SOME SOLUTION CONCEPTS FOR STOCHASTIC PROGRAMMING PROBLEMS

ANALYSIS AND COMPARISONS OF SOME SOLUTION CONCEPTS FOR STOCHASTIC PROGRAMMING PROBLEMS ANALYSIS AND COMPARISONS OF SOME SOLUTION CONCEPTS FOR STOCHASTIC PROGRAMMING PROBLEMS R. Caballero, E. Cerdá, M. M. Muñoz and L. Rey () Deparmen of Applied Economics (Mahemaics), Universiy of Málaga,

More information

How Useful are the Various Volatility Estimators for Improving GARCH-based Volatility Forecasts? Evidence from the Nasdaq-100 Stock Index

How Useful are the Various Volatility Estimators for Improving GARCH-based Volatility Forecasts? Evidence from the Nasdaq-100 Stock Index Inernaional Journal of Economics and Financial Issues Vol. 4, No. 3, 04, pp.65-656 ISSN: 46-438 www.econjournals.com How Useful are he Various Volailiy Esimaors for Improving GARCH-based Volailiy Forecass?

More information

Sampling Time-Based Sliding Windows in Bounded Space

Sampling Time-Based Sliding Windows in Bounded Space Sampling Time-Based Sliding Windows in Bounded Space Rainer Gemulla Technische Universiä Dresden 01062 Dresden, Germany gemulla@inf.u-dresden.de Wolfgang Lehner Technische Universiä Dresden 01062 Dresden,

More information

Multi-camera scheduling for video production

Multi-camera scheduling for video production Muli-camera scheduling for video producion Fahad Daniyal and Andrea Cavallaro Queen Mary Universiy of London Mile End Road, E 4S London, Unied Kingdom Email: {fahad.daniyal, andrea.cavallaro}@eecs.qmul.ac.uk

More information

Distributing Human Resources among Software Development Projects 1

Distributing Human Resources among Software Development Projects 1 Disribuing Human Resources among Sofware Developmen Proecs Macario Polo, María Dolores Maeos, Mario Piaini and rancisco Ruiz Summary This paper presens a mehod for esimaing he disribuion of human resources

More information

MTH6121 Introduction to Mathematical Finance Lesson 5

MTH6121 Introduction to Mathematical Finance Lesson 5 26 MTH6121 Inroducion o Mahemaical Finance Lesson 5 Conens 2.3 Brownian moion wih drif........................... 27 2.4 Geomeric Brownian moion........................... 28 2.5 Convergence of random

More information

Bayesian model comparison with un-normalised likelihoods

Bayesian model comparison with un-normalised likelihoods Saisics and Compuing manuscrip No. (will be insered by he edior) Bayesian model comparison wih un-normalised likelihoods Richard G. Everi Adam M. Johansen Ellen Rowing Melina Evdemon-Hogan Each of hese

More information

The Greek financial crisis: growing imbalances and sovereign spreads. Heather D. Gibson, Stephan G. Hall and George S. Tavlas

The Greek financial crisis: growing imbalances and sovereign spreads. Heather D. Gibson, Stephan G. Hall and George S. Tavlas The Greek financial crisis: growing imbalances and sovereign spreads Heaher D. Gibson, Sephan G. Hall and George S. Tavlas The enry The enry of Greece ino he Eurozone in 2001 produced a dividend in he

More information

Direc Manipulaion Inerface and EGN algorithms

Direc Manipulaion Inerface and EGN algorithms A Direc Manipulaion Inerface for 3D Compuer Animaion Sco Sona Snibbe y Brown Universiy Deparmen of Compuer Science Providence, RI 02912, USA Absrac We presen a new se of inerface echniques for visualizing

More information

Forecasting Product Sales with Dynamic Linear Mixture Models. Phillip M. Yelland and Eunice Lee

Forecasting Product Sales with Dynamic Linear Mixture Models. Phillip M. Yelland and Eunice Lee Forecasing Produc Sales wih Dynamic Linear Mixure Models Phillip M. Yelland and Eunice Lee Forecasing Produc Sales wih Dynamic Linear Mixure Models Phillip M. Yelland Eunice Lee SMLI TR-2003-122 March

More information

UNDERSTANDING THE DEATH BENEFIT SWITCH OPTION IN UNIVERSAL LIFE POLICIES. Nadine Gatzert

UNDERSTANDING THE DEATH BENEFIT SWITCH OPTION IN UNIVERSAL LIFE POLICIES. Nadine Gatzert UNDERSTANDING THE DEATH BENEFIT SWITCH OPTION IN UNIVERSAL LIFE POLICIES Nadine Gazer Conac (has changed since iniial submission): Chair for Insurance Managemen Universiy of Erlangen-Nuremberg Lange Gasse

More information

Tourism demand forecasting with different neural networks models

Tourism demand forecasting with different neural networks models Insiu de Recerca en Economia Aplicada Regional i Pública Research Insiue of Applied Economics Documen de Treball 2013/21, 23 pàg. Working Paper 2013/21, 23 pag. Grup de Recerca Anàlisi Quaniaiva Regional

More information

Particle Filtering for Multiple Object Tracking in Dynamic Fluorescence Microscopy Images: Application to Microtubule Growth Analysis

Particle Filtering for Multiple Object Tracking in Dynamic Fluorescence Microscopy Images: Application to Microtubule Growth Analysis Chaper Three Paricle Filering for Muliple Objec Tracking in Dynamic Fluorescence Microscopy Images: Applicaion o Microubule Growh Analysis I is remarkable ha a science which began wih he consideraion of

More information

Automatic measurement and detection of GSM interferences

Automatic measurement and detection of GSM interferences Auomaic measuremen and deecion of GSM inerferences Poor speech qualiy and dropped calls in GSM neworks may be caused by inerferences as a resul of high raffic load. The radio nework analyzers from Rohde

More information

APPLICATION OF THE KALMAN FILTER FOR ESTIMATING CONTINUOUS TIME TERM STRUCTURE MODELS: THE CASE OF UK AND GERMANY. January, 2005

APPLICATION OF THE KALMAN FILTER FOR ESTIMATING CONTINUOUS TIME TERM STRUCTURE MODELS: THE CASE OF UK AND GERMANY. January, 2005 APPLICATION OF THE KALMAN FILTER FOR ESTIMATING CONTINUOUS TIME TERM STRUCTURE MODELS: THE CASE OF UK AND GERMANY Somnah Chaeree* Deparmen of Economics Universiy of Glasgow January, 2005 Absrac The purpose

More information

Does Option Trading Have a Pervasive Impact on Underlying Stock Prices? *

Does Option Trading Have a Pervasive Impact on Underlying Stock Prices? * Does Opion Trading Have a Pervasive Impac on Underlying Sock Prices? * Neil D. Pearson Universiy of Illinois a Urbana-Champaign Allen M. Poeshman Universiy of Illinois a Urbana-Champaign Joshua Whie Universiy

More information

Supplementary Appendix for Depression Babies: Do Macroeconomic Experiences Affect Risk-Taking?

Supplementary Appendix for Depression Babies: Do Macroeconomic Experiences Affect Risk-Taking? Supplemenary Appendix for Depression Babies: Do Macroeconomic Experiences Affec Risk-Taking? Ulrike Malmendier UC Berkeley and NBER Sefan Nagel Sanford Universiy and NBER Sepember 2009 A. Deails on SCF

More information

Premium Income of Indian Life Insurance Industry

Premium Income of Indian Life Insurance Industry Premium Income of Indian Life Insurance Indusry A Toal Facor Produciviy Approach Ram Praap Sinha* Subsequen o he passage of he Insurance Regulaory and Developmen Auhoriy (IRDA) Ac, 1999, he life insurance

More information

Chapter 7. Response of First-Order RL and RC Circuits

Chapter 7. Response of First-Order RL and RC Circuits Chaper 7. esponse of Firs-Order L and C Circuis 7.1. The Naural esponse of an L Circui 7.2. The Naural esponse of an C Circui 7.3. The ep esponse of L and C Circuis 7.4. A General oluion for ep and Naural

More information

MODEL AND ALGORITHMS FOR THE REAL TIME MANAGEMENT OF RESIDENTIAL ELECTRICITY DEMAND. A. Barbato, G. Carpentieri

MODEL AND ALGORITHMS FOR THE REAL TIME MANAGEMENT OF RESIDENTIAL ELECTRICITY DEMAND. A. Barbato, G. Carpentieri MODEL AND ALGORITHMS FOR THE REAL TIME MANAGEMENT OF RESIDENTIAL ELECTRICITY DEMAND A. Barbao, G. Carpenieri Poliecnico di Milano, Diparimeno di Eleronica e Informazione, Email: barbao@ele.polimi.i, giuseppe.carpenieri@mail.polimi.i

More information

Tracking of Multiple Moving Sources Using Recursive EM Algorithm

Tracking of Multiple Moving Sources Using Recursive EM Algorithm EURASIP Journal on Applied Signal Processing 24:18, 1 11 c 24 Hindawi Publishing Corporaion Tracking of Muliple Moving Sources Using Recursive EM Algorihm Pei-Jung Chung Deparmen of Elecrical Engineering

More information

Market Efficiency or Not? The Behaviour of China s Stock Prices in Response to the Announcement of Bonus Issues

Market Efficiency or Not? The Behaviour of China s Stock Prices in Response to the Announcement of Bonus Issues Discussion Paper No. 0120 Marke Efficiency or No? The Behaviour of China s Sock Prices in Response o he Announcemen of Bonus Issues Michelle L. Barnes and Shiguang Ma May 2001 Adelaide Universiy SA 5005,

More information

17 Laplace transform. Solving linear ODE with piecewise continuous right hand sides

17 Laplace transform. Solving linear ODE with piecewise continuous right hand sides 7 Laplace ransform. Solving linear ODE wih piecewise coninuous righ hand sides In his lecure I will show how o apply he Laplace ransform o he ODE Ly = f wih piecewise coninuous f. Definiion. A funcion

More information

CHARGE AND DISCHARGE OF A CAPACITOR

CHARGE AND DISCHARGE OF A CAPACITOR REFERENCES RC Circuis: Elecrical Insrumens: Mos Inroducory Physics exs (e.g. A. Halliday and Resnick, Physics ; M. Sernheim and J. Kane, General Physics.) This Laboraory Manual: Commonly Used Insrumens:

More information

Load Prediction Using Hybrid Model for Computational Grid

Load Prediction Using Hybrid Model for Computational Grid Load Predicion Using Hybrid Model for Compuaional Grid Yongwei Wu, Yulai Yuan, Guangwen Yang 3, Weimin Zheng 4 Deparmen of Compuer Science and Technology, Tsinghua Universiy, Beijing 00084, China, 3, 4

More information

AP Calculus AB 2013 Scoring Guidelines

AP Calculus AB 2013 Scoring Guidelines AP Calculus AB 1 Scoring Guidelines The College Board The College Board is a mission-driven no-for-profi organizaion ha connecs sudens o college success and opporuniy. Founded in 19, he College Board was

More information

Mobile Broadband Rollout Business Case: Risk Analyses of the Forecast Uncertainties

Mobile Broadband Rollout Business Case: Risk Analyses of the Forecast Uncertainties ISF 2009, Hong Kong, 2-24 June 2009 Mobile Broadband Rollou Business Case: Risk Analyses of he Forecas Uncerainies Nils Krisian Elnegaard, Telenor R&I Agenda Moivaion Modelling long erm forecass for MBB

More information

Dynamic programming models and algorithms for the mutual fund cash balance problem

Dynamic programming models and algorithms for the mutual fund cash balance problem Submied o Managemen Science manuscrip Dynamic programming models and algorihms for he muual fund cash balance problem Juliana Nascimeno Deparmen of Operaions Research and Financial Engineering, Princeon

More information

11/6/2013. Chapter 14: Dynamic AD-AS. Introduction. Introduction. Keeping track of time. The model s elements

11/6/2013. Chapter 14: Dynamic AD-AS. Introduction. Introduction. Keeping track of time. The model s elements Inroducion Chaper 14: Dynamic D-S dynamic model of aggregae and aggregae supply gives us more insigh ino how he economy works in he shor run. I is a simplified version of a DSGE model, used in cuing-edge

More information

Time-Expanded Sampling (TES) For Ensemble-based Data Assimilation Applied To Conventional And Satellite Observations

Time-Expanded Sampling (TES) For Ensemble-based Data Assimilation Applied To Conventional And Satellite Observations 27 h WAF/23 rd NWP, 29 June 3 July 2015, Chicago IL. 1 Time-Expanded Sampling (TES) For Ensemble-based Daa Assimilaion Applied To Convenional And Saellie Observaions Allen Zhao 1, Qin Xu 2, Yi Jin 1, Jusin

More information

Stock Price Prediction Using the ARIMA Model

Stock Price Prediction Using the ARIMA Model 2014 UKSim-AMSS 16h Inernaional Conference on Compuer Modelling and Simulaion Sock Price Predicion Using he ARIMA Model 1 Ayodele A. Adebiyi., 2 Aderemi O. Adewumi 1,2 School of Mahemaic, Saisics & Compuer

More information

The option pricing framework

The option pricing framework Chaper 2 The opion pricing framework The opion markes based on swap raes or he LIBOR have become he larges fixed income markes, and caps (floors) and swapions are he mos imporan derivaives wihin hese markes.

More information

Optimal Investment and Consumption Decision of Family with Life Insurance

Optimal Investment and Consumption Decision of Family with Life Insurance Opimal Invesmen and Consumpion Decision of Family wih Life Insurance Minsuk Kwak 1 2 Yong Hyun Shin 3 U Jin Choi 4 6h World Congress of he Bachelier Finance Sociey Torono, Canada June 25, 2010 1 Speaker

More information

Distributed Online Localization in Sensor Networks Using a Moving Target

Distributed Online Localization in Sensor Networks Using a Moving Target Disribued Online Localizaion in Sensor Neworks Using a Moving Targe Aram Galsyan 1, Bhaskar Krishnamachari 2, Krisina Lerman 1, and Sundeep Paem 2 1 Informaion Sciences Insiue 2 Deparmen of Elecrical Engineering-Sysems

More information

Working Paper A Monte Carlo Comparison between the Free Cash Flow and Discounted Cash Flow Approaches

Working Paper A Monte Carlo Comparison between the Free Cash Flow and Discounted Cash Flow Approaches econsor www.econsor.eu Der Open-Access-Publikaionsserver der ZBW Leibniz-Informaionszenrum Wirschaf The Open Access Publicaion Server of he ZBW Leibniz Informaion Cenre for Economics Akalu, Mehari Mekonnen;

More information

Hedging with Forwards and Futures

Hedging with Forwards and Futures Hedging wih orwards and uures Hedging in mos cases is sraighforward. You plan o buy 10,000 barrels of oil in six monhs and you wish o eliminae he price risk. If you ake he buy-side of a forward/fuures

More information

Statistical Analysis with Little s Law. Supplementary Material: More on the Call Center Data. by Song-Hee Kim and Ward Whitt

Statistical Analysis with Little s Law. Supplementary Material: More on the Call Center Data. by Song-Hee Kim and Ward Whitt Saisical Analysis wih Lile s Law Supplemenary Maerial: More on he Call Cener Daa by Song-Hee Kim and Ward Whi Deparmen of Indusrial Engineering and Operaions Research Columbia Universiy, New York, NY 17-99

More information

Acceleration Lab Teacher s Guide

Acceleration Lab Teacher s Guide Acceleraion Lab Teacher s Guide Objecives:. Use graphs of disance vs. ime and velociy vs. ime o find acceleraion of a oy car.. Observe he relaionship beween he angle of an inclined plane and he acceleraion

More information

Performance Center Overview. Performance Center Overview 1

Performance Center Overview. Performance Center Overview 1 Performance Cener Overview Performance Cener Overview 1 ODJFS Performance Cener ce Cener New Performance Cener Model Performance Cener Projec Meeings Performance Cener Execuive Meeings Performance Cener

More information

II.1. Debt reduction and fiscal multipliers. dbt da dpbal da dg. bal

II.1. Debt reduction and fiscal multipliers. dbt da dpbal da dg. bal Quarerly Repor on he Euro Area 3/202 II.. Deb reducion and fiscal mulipliers The deerioraion of public finances in he firs years of he crisis has led mos Member Saes o adop sizeable consolidaion packages.

More information

E0 370 Statistical Learning Theory Lecture 20 (Nov 17, 2011)

E0 370 Statistical Learning Theory Lecture 20 (Nov 17, 2011) E0 370 Saisical Learning Theory Lecure 0 (ov 7, 0 Online Learning from Expers: Weighed Majoriy and Hedge Lecurer: Shivani Agarwal Scribe: Saradha R Inroducion In his lecure, we will look a he problem of

More information

The Real Business Cycle paradigm. The RBC model emphasizes supply (technology) disturbances as the main source of

The Real Business Cycle paradigm. The RBC model emphasizes supply (technology) disturbances as the main source of Prof. Harris Dellas Advanced Macroeconomics Winer 2001/01 The Real Business Cycle paradigm The RBC model emphasizes supply (echnology) disurbances as he main source of macroeconomic flucuaions in a world

More information

The Kinetics of the Stock Markets

The Kinetics of the Stock Markets Asia Pacific Managemen Review (00) 7(1), 1-4 The Kineics of he Sock Markes Hsinan Hsu * and Bin-Juin Lin ** (received July 001; revision received Ocober 001;acceped November 001) This paper applies he

More information

Monte Carlo Observer for a Stochastic Model of Bioreactors

Monte Carlo Observer for a Stochastic Model of Bioreactors Mone Carlo Observer for a Sochasic Model of Bioreacors Marc Joannides, Irène Larramendy Valverde, and Vivien Rossi 2 Insiu de Mahémaiques e Modélisaion de Monpellier (I3M UMR 549 CNRS Place Eugène Baaillon

More information

TSG-RAN Working Group 1 (Radio Layer 1) meeting #3 Nynashamn, Sweden 22 nd 26 th March 1999

TSG-RAN Working Group 1 (Radio Layer 1) meeting #3 Nynashamn, Sweden 22 nd 26 th March 1999 TSG-RAN Working Group 1 (Radio Layer 1) meeing #3 Nynashamn, Sweden 22 nd 26 h March 1999 RAN TSGW1#3(99)196 Agenda Iem: 9.1 Source: Tile: Documen for: Moorola Macro-diversiy for he PRACH Discussion/Decision

More information

This document is downloaded from DR-NTU, Nanyang Technological University Library, Singapore.

This document is downloaded from DR-NTU, Nanyang Technological University Library, Singapore. This documen is downloaded from DR-NTU, Nanyang Technological Universiy Library, Singapore. Tile A Bayesian mulivariae risk-neural mehod for pricing reverse morgages Auhor(s) Kogure, Asuyuki; Li, Jackie;

More information

4 Convolution. Recommended Problems. x2[n] 1 2[n]

4 Convolution. Recommended Problems. x2[n] 1 2[n] 4 Convoluion Recommended Problems P4.1 This problem is a simple example of he use of superposiion. Suppose ha a discree-ime linear sysem has oupus y[n] for he given inpus x[n] as shown in Figure P4.1-1.

More information

Efficient One-time Signature Schemes for Stream Authentication *

Efficient One-time Signature Schemes for Stream Authentication * JOURNAL OF INFORMATION SCIENCE AND ENGINEERING, 611-64 (006) Efficien One-ime Signaure Schemes for Sream Auhenicaion * YONGSU PARK AND YOOKUN CHO + College of Informaion and Communicaions Hanyang Universiy

More information

PRACTICES AND ISSUES IN OPERATIONAL RISK MODELING UNDER BASEL II

PRACTICES AND ISSUES IN OPERATIONAL RISK MODELING UNDER BASEL II Lihuanian Mahemaical Journal, Vol. 51, No. 2, April, 2011, pp. 180 193 PRACTICES AND ISSUES IN OPERATIONAL RISK MODELING UNDER BASEL II Paul Embrechs and Marius Hofer 1 RiskLab, Deparmen of Mahemaics,

More information

Accelerated Gradient Methods for Stochastic Optimization and Online Learning

Accelerated Gradient Methods for Stochastic Optimization and Online Learning Acceleraed Gradien Mehods for Sochasic Opimizaion and Online Learning Chonghai Hu, James T. Kwok, Weike Pan Deparmen of Compuer Science and Engineering Hong Kong Universiy of Science and Technology Clear

More information

Robust 3D Head Tracking by Online Feature Registration

Robust 3D Head Tracking by Online Feature Registration Robus 3D Head Tracking by Online Feaure Regisraion Jun-Su Jang and Takeo Kanade Roboics Insiue, Carnegie Mellon Universiy, Pisburgh, PA 15213 {jsjang, k}@cs.cmu.edu Absrac This paper presens a robus mehod

More information

Segment and combine approach for non-parametric time-series classification

Segment and combine approach for non-parametric time-series classification Segmen and combine approach for non-parameric ime-series classificaion Pierre Geurs and Louis Wehenkel Universiy of Liège, Deparmen of Elecrical Engineering and Compuer Science, Sar-Tilman B28, B4000 Liège,

More information

Working Paper No. 482. Net Intergenerational Transfers from an Increase in Social Security Benefits

Working Paper No. 482. Net Intergenerational Transfers from an Increase in Social Security Benefits Working Paper No. 482 Ne Inergeneraional Transfers from an Increase in Social Securiy Benefis By Li Gan Texas A&M and NBER Guan Gong Shanghai Universiy of Finance and Economics Michael Hurd RAND Corporaion

More information

How To Calculate Price Elasiciy Per Capia Per Capi

How To Calculate Price Elasiciy Per Capia Per Capi Price elasiciy of demand for crude oil: esimaes for 23 counries John C.B. Cooper Absrac This paper uses a muliple regression model derived from an adapaion of Nerlove s parial adjusmen model o esimae boh

More information

COMPUTATION OF CENTILES AND Z-SCORES FOR HEIGHT-FOR-AGE, WEIGHT-FOR-AGE AND BMI-FOR-AGE

COMPUTATION OF CENTILES AND Z-SCORES FOR HEIGHT-FOR-AGE, WEIGHT-FOR-AGE AND BMI-FOR-AGE COMPUTATION OF CENTILES AND Z-SCORES FOR HEIGHT-FOR-AGE, WEIGHT-FOR-AGE AND BMI-FOR-AGE The mehod used o consruc he 2007 WHO references relied on GAMLSS wih he Box-Cox power exponenial disribuion (Rigby

More information

Large Scale Online Learning.

Large Scale Online Learning. Large Scale Online Learning. Léon Boou NEC Labs America Princeon NJ 08540 leon@boou.org Yann Le Cun NEC Labs America Princeon NJ 08540 yann@lecun.com Absrac We consider siuaions where raining daa is abundan

More information

The Relationship between Stock Return Volatility and. Trading Volume: The case of The Philippines*

The Relationship between Stock Return Volatility and. Trading Volume: The case of The Philippines* The Relaionship beween Sock Reurn Volailiy and Trading Volume: The case of The Philippines* Manabu Asai Faculy of Economics Soka Universiy Angelo Unie Economics Deparmen De La Salle Universiy Manila May

More information