Scaling Up POMDPs for Dialog Management: The Summary POMDP Method. Jason D. Williams and Steve Young
|
|
- William Rice
- 8 years ago
- Views:
Transcription
1 Scalng Up POMDPs for Dalog Managemen: The Summary POMDP Mehod Jason D. Wllams and Seve Young Cambrdge Unversy Engneerng Deparmen Trumpngon Sree, Cambrdge CB2 1PZ, UK BSTRCT Parally Observable Markov Decson Processes (POMDPs have been shown o be a promsng framework for dalog managemen n spoken dalog sysems. However, o dae, POMDPs have been lmed o arfcally small asks. In hs work, we presen a novel mehod called a Summary POMDP for scalng -fllng POMDP-based dalog managers o cope wh asks of a realsc sze. n example dalog problem ncorporang a user model bul from real dalog daa s presened. dalog manager s creaed usng hs mehod and evaluaed usng a second user model creaed from held-ou dalog daa. Resuls confrm ha summary POMDP polces scale well, and also show ha summary POMDP polces are reasonably robus o varaons n user behavor. 1. INTRODUCTION Dalog managemen for spoken dalog sysems can be approached as plannng under uncerany. (Fully-observable Markov decson processes ((FOMDPs and parally observable Markov decson processes (POMDPs offer a prncpled framework n hs pursu [5]. MDPs offer an accessble body of opmzaon echnques and have been successfully used o creae dalog managers for problems of realsc szes [4, 6, 9, 10]. However, MDPs do no ake proper accoun of he corrupon nroduced by he speech recognon channel. The POMDP framework, whch s desgned o cope wh nosy npu, has been found o conssenly ouperform MDP-based sysems [8, 14, 15, 16]. Despe her promse, POMDPs o dae have been lmed o arfcally small oy problems. Sysems n he leraure handle fewer han 10 dsnc user goals [8, 14, 15, 16]. In hs paper, we presen a mehod for scalng up POMDPs for dalog managemen he Summary POMDP mehod and demonsrae s ably o creae a robus -based dalog manager capable of handlng a realscally szed dalog problem. Ths paper s organzed as follows: Secon 2 brefly revews background on POMDPs. Secon 3 presens he summary POMDP mehod. Secon 4 brefly demonsraes he mehod, Secon 5 provdes an evaluaon, and Secon 6 concludes. 2. OVERVIEW OF POMDPS Formally, a POMDP s defned as a uple {S,, T, R, O, Z}, where S s a se of saes; s a se of acons ha an agen may ake; T defnes a ranson probably P ( s s, a ; R defnes he expeced (mmedae, real-valued reward r ( s, a ; O s a se of observaons; and Z defnes an observaon probably, P ( o s, a. The POMDP operaes as follows. each me-sep, he machne s n some unobserved sae s. Snce s s no known exacly, we manan a dsrbuon over saes called a belef sae, b. We wre b(s o ndcae he probably of beng n a parcular sae s. Based on b(s, he machne selecs an acon a, receves a reward r, and ransons o (unobserved sae s, where s depends only on s and a. The machne hen receves an observaon o whch s dependen on s and a. each mesep, we updae b as follows: b ( s = k P( o s, a P( s a, s b( s, (1 s S where k s a normalzaon consan [2]. We refer o mananng he value of b a each me-sep as belef monorng. The varables a, o, and s may be facored no componens. For example, s could be decomposed no s = ( s1, s2, K, sn such ha s1 S1, s2 S2, K, s N S N. In hs case, b(s s wren b( s = b( s1, s2, K, sn, P ( o s, a s wren P( o s1', s2', K, sn ', a, ec. Ths facorng may allow condonal ndependence o be exploed, reducng he parameers requred o express T, R, and Z. The cumulave, nfne-horzon, dscouned reward s called he reurn: = 0 s S λ b ( s r( s, a (2 where b ndcaes he dsrbuon over all saes a me ; b (s ndcaes he probably of beng n sae s a mesep ; a s he acon a aken a me ; and λ s a geomerc dscoun facor, 0 λ 1. polcy π ( b specfes an acon o ake gven a belef sae. 1 The goal of he planner s o fnd a polcy ha maxmzes he reurn. The doman b of a polcy π (b s a pon n S-dmensonal space, called belef space. In general a polcy s a paronng of belef space, where each paron corresponds o an acon. In fac, he sze of he polcy space grows exponenally wh he sze of he observaon se and doubly exponenally wh he 1 We wll assume he plannng horzon for a polcy s nfne unless oherwse saed.
2 dsance (n me-seps from he horzon [2]. Neverheless, realworld problems ofen possess small polces of hgh qualy. In hs work, POMDP opmzaon s performed wh Perseus [11]. Perseus heurscally selecs a small se of represenave belef pons, and hen eravely apples value updaes o jus hose pons nsead of all of belef space, achevng a sgnfcan speed-up over exac mehod. Increasng he number of belef pons enables he algorhm more complex polces a he expense of more compuaon. 3. THE SUMMRY POMDP METHOD Ths paper s concerned wh so-called -fllng dalogs, whch are common n he spoken dalog sysem leraure. In -fllng dalogs, here exs Ν s, where akes on one of N values. The user eners he dalog wh a goal.e., a desred value for each and he am of he machne s o correcly subm he user s goal. Tradonal POMDP-based dalog managers scale poorly because her sae space, acon se, and observaon se grow as Π N. For example, consder a problem wh wo s n whch each akes on one of N 1 = N 2 = M = 1000 values. In hs case, he POMDP s performng plannng on a dsrbuon over M = 10 saes wh more han 10 acons and observaons. Even wh recen echnques, creang polces on POMDPs of hs sze s nracable. By conras, he summary POMDP mehod consders only he mos lkely hypohess for each, effecvely reducng he sze of each o 2 values. The acon and observaon ses are smlarly reduced. Overall hs approach reduces he growh Ν Ν facor from M o 2. The summary POMDP mehod consss of hree phases: consrucon, samplng & opmzaon, and execuon. s he mehod s explaned, an example dalog problem s presened for llusraon Consrucon of he maser POMDP Frs, a POMDP called he maser POMDP s consruced. The sae varable of he maser POMDP s facored no hree groups of varables whch express he user s goal, he user s acon, and he sae of he dalog from he perspecve of he user. Noe ha, from he machne s perspecve, all of hese varables are hdden. Ths facorng faclaes esmang he ranson and observaon funcons and specfyng he reward funcon [14]. The mehod requres ha he sae space componen for he user s goal s furher subdvded no Ν s, wh one componen for each, S1 KS N. The componens for he user s acon and dalog sae may, f desred, be furher decomposed. n llusrave maser POMDP n he ravel doman s shown as an nfluence dagram n Fgure 1, wh componens descrbed n Table 1. In hs dalog problem, he user s ryng o ravel from a from locaon o a o locaon n a world wh 1000 places. Thus he user s goal ncludes 2 componens represenng 2 s, S from and, each of whch akes on 1 of 1000 values. The model of user s acon has been decomposed no he eq 2 componens S, S, S, S, S, and S. Two condonal probably ables form he core of he user model. The dsrbuon P ( s s gves he probably ha he user provdes varous wh-responses (.e., ways of sang a value such as london or o(london for a gven sysem acon. The eq dsrbuon P ( s s, s gves he probably ha he user ncludes a yn-response (yes or no n her uerance for a gven eq sysem acon. The oher user acon componens ( S, S, 2 S, and S are deermnsc funcons and smply enable S and S o be expressed succncly. Fnally, he node S expresses dalog sae,.e., wheher s, from he sandpon of he user, no-saed, saed, or confrmed. In he summary POMDP mehod, he acon se s also facored, no a = a, a, K, a, such ha a S ( 1 from o and a. represens he se of locuonary s avalable o he sysem. In Fgure 1, he sysem acon s formed a = ( a, a, a. conans 5 values: askfrom, confrm-from, ask-o, confrm-o, subm. Compose sysem acons express nclude, for example, confrm-o(london, subm(boson,london, or ask-from. The observaon se O represens he oupu of he speech recognon and parsng process, and ncludes all possble user uerances a he concep level. In he example dalog problem, he observaon se ncludes all combnaons of user acons (as observed by he parser. For example, one observaon mgh be he parse {yes, place(london, o(place(boson}. In general and as n he example, he observaon se wll be oo large o esmae a condonal probably able. Insead, we creae a funcon f ( o, perr P( o s whch esmaes P ( o s for a gven o and per-concep error rae p err. In he example problem, he error rae p err specfes he lkelhood ha a sngle concep s msrecognzed (.e., subsued or deleed. The reward funcon s specfed by he sysem desgner. In he example dalog problem, he reward funcon encourages he sysem o correcly denfy he user s goal as quckly as possble whle observng conversaonal norms. For acons whch do no end he dalog, per-urn penales are assgned whch seek o reward approprae behavor, as lsed n Table 2. For example, a hgher per-urn penaly s gven for confrmng a ha hasn been saed han for confrmng one ha has. The subm acon, whch s he only acon ha ends he dalog, assgns +50 f he user s goal s correcly denfed and -50 f no. The process of belef monorng nvolves nferrng b(s based on he observaon o as n Eq. 1. Inally, b(s s a fla dsrbuon bu as he dalog proceeds, sharpens around he mos lkely values of s. The provson of an accurae user acon model and observaon funcon grealy faclaes hs process Consrucon of he summary POMDP Whereas radonal mehods would aemp o opmze he maser POMDP drecly (e.g., [14], here we form a second POMDP called he summary POMDP n whch opmzaon wll be conduced. The sae space of he summary POMDP, S, conans a componen S N correspondng o each componen S n
3 o from O eq 2 o R O Tmesep Tmesep +1 eq S 2 o o from Fgure 1: Influence dagram showng he maser POMDP for he sample dalog problem, followng [1]. Unshaded nodes are observable, and shaded nodes are unobservable. Crcles are chance nodes; squares are decson nodes; and damonds are uly nodes. rrows show causal nfluence. For clary, only he o s shown, and he acon and uly nodes for he +1 me-sep have been omed. The doed boxes show compose sae and acon varables. Se Meanng S User s goal for eq S Indcaes wheher equals S S Sysem s acon as relaes o hs S -poron of user s acon, e.g., o(x or x S -poron of user s acon (yes, no S acon wh value only, e.g., london 2 S ac w/ name & value, e.g., o(london S Dalog sae: {no saed, saed, or confrmed} O Full concep srng (boh s Sys c: ask-o/from, confrm-o/from, or subm Conen of sysem acon (same se as R Reward (see Table 2 and ex Table 1: Defnon of node labels n Fgure 1. S Dalog sae Sysem acon Reward No saed ask 1 No saed confrm 3 Saed ask 2 Saed confrm 1 Confrmed ask 3 Confrmed confrm 2 Table 2: Per-urn rewards for he sample problem. he maser POMDP. Each s akes on jus 2 values, s { bes, res}, calculaed as follows: s bes, = res, f arg max( b ( s f arg max( b ( s = S S ( (. Essenally, for a gven, he sae of he summary POMDP expresses wheher he mos lkely hypohess n he maser POMDP s correc. Oher sae space componens from he maser POMDP may be ncluded n he summary POMDP sae space f desred. In he example dalog problem, he sae of he summary POMDP S ncludes a componen S whch s a copy of he dalog sae n he maser POMDP, S. The example summary POMDP ncludes a oal of 36 saes. Fgure 2 shows an nfluence dagram of he example summary POMDP, llusrang how a summary POMDP conans fewer and more compac componens han he maser POMDP, faclang opmzaon. The acon space of he summary POMDP ncludes only. Thus, n he example dalog problem, he summary acon ncludes he componen bu no from or o, yeldng a oal of fve acons n he example summary POMDP. The observaon se of he summary POMDP, O, ncludes S O wo componens O S S and O for each. The frs observaon componen s calculaed as follows: c, f o s conssen wh argmax ( b( s S O O = c, f o s nconssen wh argmax( b( s n, f o provdes no nformaon abou b( s Ths conssency relaonshp s specfed by he sysem desgner. In he example dalog problem, f he observaon conans london (whou an ndcaon of o or from, s conssen wh arg max( b ( s = london and nconssen wh any oher value for boh = o and = from. If he observaon conans o(london, s conssen wh arg max( b ( = london and provdes no nformaon abou S from S from S O S S S o S o S O Oo S S Oo R S from S o S from S o S O S S O Oo S S S Tmesep Tmesep +1 Oo s o s from. The R Fgure 2: Influence dagram showng he summary POMDP for he sample dalog problem
4 observaons yes and no alone provde no nformaon abou any S O va O.e., O S O = n. The second observaon componen s calculaed as follows: SS eq, f arg max( b 1( s = arg max( b ( s O = ne, f arg max( b 1( s arg max( b ( s S S In oher words, O ndcaes wheher he mos lkely hypohess for s has changed. For example, n he sample problem, suppose ha arg max( b ( = london, he acon s from confrm-from(london s aken, and no s observed. Ths would resul n a sgnfcan reducon n b + 1( london, and a new mos lkely hypohess, for example arg max( b + 1( s from = leeds. In hs example, london leeds, so O S from S = ne Samplng & Opmzaon To esmae he sysem dynamcs of he summary POMDP, we sample from he maser POMDP usng a random polcy. each mesep, an acon a n he summary POMDP s randomly seleced. The acon a n he maser POMDP s hen formed by seng a = arg max( b( s and combnng a a and, o form a. Fnally, he value of he reward r s calculaed. Nex, a value for he nex (unobserved sae s and observaon o of he maser POMDP are hen sampled. Then he belef sae b (s of he maser POMDP s calculaed as n Eq. 1. Fnally, he sae s and observaons o of he summary POMDP are calculaed as descrbed n 3.2. By samplng repeaedly n hs way, he ranson funcon P ( s s, a, observaon funcon P ( o s, a, and reward funcon r ( s, a of he summary POMDP are esmaed. fer samplng, opmzaon s performed on he summary POMDP o compue a polcy, π ( b ( s a, usng for example [11] Execuon To execue he conroller, belef monorng s performed n he maser POMDP. The belef sae of he summary POMDP s calculaed as s b ( s = bes = max( b( s (3 and ( 1 ( b s = res = b s = bes. The acon a s seleced based on he polcy calculaed above,.e., a = π ( b ( s. The acon a n he maser POMDP s hen formed by seng a = arg max( b( s for all s and combnng a and a, o form a. fer a s aken, a reward r and an observaon o are receved, and he nex belef sae b (s s calculaed (Eq DEMONSTRTION Fgure 5 shows an example conversaon beween he es user model and he POMDP. The small graphs show he dsrbuon of belef mass n b ( s and b ( s over he course of he dalog. Only 3 of he 1000 values of s are shown. he begnnng of he dalog, he belef mass n b ( s s spread evenly over all goals, and ( b s = bes s low. s he dalog progresses, probably mass sharpens around observed values. s explaned n he descrpon of he mehod above, a each me-sep, b ( s = bes = max( b( s. For example, afer urn U1, arg max( b( so = leeds and b ( s = bes = b( leeds. ( b s = res s equal o b for all oher values of s. One srengh of he probablsc approach s ha he sysem never comms o one value for a ; raher, b ( bes s acng as a global confdence score over he course of he dalog. For example, he no observed n U4 redsrbues probably mass n ( b s o from s o = leeds o all oher values, causng a decrease n b ( so = bes. lernavely, a second observaon provdng furher suppor for s from = cambrdge n U5 causes an ncrease n b ( s from = bes. second srengh s ha he effec on b of an observaon s scaled by he lkelhood of he correspondng user acon. For example, n U4, he user acon from(sheffeld s very unlkely, so when he observaon from(sheffeld s made ncreases b( s from = sheffeld bu only slghly. 5. EVLUTION We are neresed n deermnng how performance of he summary mehod vares wh error rae, how robus he mehod s o varaons n user behavor, and fnally how he summary mehod compares o radonal mehods. To do hs, we employ real dalog daa from he SCTI-1 corpus [13]. The SCTI-1 corpus conans 144 human-human dalogs n he ravel/ours nformaon doman usng a smulaed SR channel [12]. The corpus conans a varey of word error raes, and he behavors observed of he subjecs n he corpus are broadly conssen wh behavors observed of a user and a compuer usng a real speech recognon sysem [13]. The corpus was segmened no a ranng sub-corpus and a es sub-corpus, whch are each composed of an equal number of dalogs, he same mx of word error raes, and dsjon subjec ses. Wzard/User urn pars were annoaed, and one user model was hen esmaed from each sub-corpus. The ranng user model was nsalled n nodes S and S n he maser POMDP, gvng he wh poron of he user s response such as o(cambrdge, and he yn (yes/no poron, respecvely. To creae he summary POMDP, 20,000 dalog urns were sampled. Polcy opmzaon was performed wh 50 belef pons, 50 eraons, and a dscoun of λ = To form a baselne, 20,000 dalog urns were hen run wh he 2 The number of belef pons broadly ses an upper bound on he complexy of he resulng polcy.
5 Reward ganed / urn verage or expeced reurn User model bul from ranng se User model bul from es se Fgure 3: p err vs. reward ganed per urn for he sample dalog problem p err M (Number of dsnc values Summary POMDP Baselne Fgure 4: n (number of dsnc values vs. average or expeced reurn for a smplfed 1- dalog problem. The baselne s a drec soluon of he maser POMDP. resulng polcy and he ranng user model. Nex, he es user model was nsalled no he Maser POMDP, and 20,000 dalog urns were run wh he polcy creaed from he ranng user model. Ths process was repeaed for values of p err from 0.05 o Fgure 3 shows resuls for a range of values of p err. The Y- axs shows average reward ganed per dalog urn, and error bars ndcae he 95% confdence nerval for he rue average reward ganed per dalog urn. s speech recognon errors ncrease, he average reward per urn decreases, conssen wh radonal mehods [8, 14, 15]. For all values of p err excep 0.05, performance on he es user model s no less han 0.5 pons of he ranng user model and s generally very close. In oher words, he polcy creaed for he ranng user model s performng smlarly on he esng user model, mplyng ha he mehod s reasonably robus o varaons n paerns of user behavor. In some cases, e.g., p err = 0.05, he polcy performs beer on he es user model han on he ranng user model. Ths s possble because dfferen user models can provde dfferen amouns of nformaon. In hs example, he es user model provdes slghly more nformaon han he ranng user model, whch enables he polcy o perform beer on he esng user model a ceran error raes. Fnally, he sysem was run n a smpler one confguraon whch allowed he maser POMDP o be opmzed drecly. Drec soluons were compued wh 1,000 belef pons and 50 eraons. The summary POMDP mehod was hen appled usng 20,000 dalog urns, 50 belef pons and 50 eraons. Ths process was repeaed for values of M (.e., dsnc number of values from 3 o The concep error rae was se o p err = 0.30, and he dscoun o λ = for all expermens. Fgure 4 shows M vs. average or expeced reurn for he summary mehod and he baselne. The soluon algorhm was no able o opmze he maser POMDP drecly for M > 100. Error bars show 95% confdence nervals for rue expeced reurn. For small problems,.e., lower values of M, he summary mehod performs equvalenly o he baselne. For larger problems, he summary mehod ouperforms he baselne by an ncreasng margn. Moreover, he summary POMDP polcy was derved usng 95% fewer belef pons han he baselne;.e., he summary mehod s polces scale o large problems and are much more compac. I s neresng o noe ha as M ncreases, he performance of he summary POMDP mehod appears o ncrease oward an asympoe. Ths rend s due o he fac ha all confusons are equally lkely n hs model. For a gven error rae, he more conceps n he model, he less lkely conssen confusons are. Thus, havng more conceps helps he polcy denfy spurous evdence over he course of a dalog. 6. CONCLUSION POMDP-based dalog managers have been shown o ouperform MDP-based dalog managers, bu o dae have been lmed o arfcally small problems. Ths paper has demonsraed a mehod for scalng POMDP-based dalog managers o problems of a realsc sze. The mehod has been demonsraed wh a 2- problem ncorporang a user model esmaed from real dalog daa, and expermens have shown ha he resulng dalog manager copes well wh changes n paerns of user behavor. Moreover, summary dalog polces appear o be as good as hose derved drecly for he full model. s presened here, he sze of he sae space and observaon space n he summary POMDP sll grow exponenally n he number of s. Whle many real-world problems use a handful of s (e.g. 2, 3 or 4, ohers use many s. Thus one heorecal ssue o address n fuure work s how o scale o large numbers of s. The facored naure of he summary POMDP may be of some help here, for example [7]. Fnally, fuure praccal work wll aemp o valdae he mehod by consrucng an end-o-end sysem. 7. CKNOWLEDGEMENTS The auhors would lke o hank Pascal Poupar for several helpful, nsghful dscussons. Thanks also o Ma Sule and Jos Schazmann for helpful commens on he presenaon of hs paper. The work repored n hs paper was suppored by he EU FP6 Talk Projec. 11. REFERENCES [1] Fnn V. Jensen. Bayesan Neworks and Decson Graphs. New York: Sprnger Verlang, 2001.
6 [2] Lesle Pack Kaelblng, Mchael L. Lman and nhony R. Cassandra. Plannng and cng n Parally Observable Sochasc Domans. rfcal Inellgence, Vol. 101, [3] Saffan Larsson and Davd Traum. Informaon sae and dalogue managemen n he rnd dalogue move engne oolk. Naural Language Engneerng, 5(3 4: , [4] Esher Levn, Robero Peraccn, and Weland Ecker. Sochasc Model of Human-Machne Ineracon for Learnng Dalogue Sraeges. IEEE Transacons on Speech and udo Processng, Volume 8, No. 1, 11-23, [5] Esher Levn and Robero Peraccn. Sochasc Model of Compuer-Human Ineracon For Learnng Dalogue Sraeges. Eurospeech, Rhodes, Greece, [6] Olver Pequn. Framework for Unsupervsed Learnng of Dalogue Sraeges. Ph D hess, Faculy of Engneerng, Mons, Belgum, [7] Pascal Poupar and Crag Bouler. VDCBPI: an pproxmae Scalable lgorhm for Large Scale POMDPs. dvances n Neural Informaon Processng Sysems 17 (NIPS- 2004, Vancouver, BC, pp Walker. Opmzng Dalogue Managemen wh Renmen Leanng: Expermens wh he NJFun Sysem. Journal of rfcal Inellgence, Vol. 16, , [11] Mahjs T. J. Spaan and Nkos Vlasss. Perseus: randomzed pon-based value eraon for POMDPs. Techncal Repor IS-UV-04-02, Informacs Insue, Unversy of mserdam, [12] Mahew Sule, Jason D. Wllams, and Seve Young. Framework for Wzard-of-Oz Expermens wh a Smulaed SR-Channel. Inernaonal Conferences on Spoken Language Processng (ICSLP-2004, Jeju, Souh Korea, [13] Jason D. Wllams and Seve Young. Characerzng Task- Orened Dalog usng a Smulaed SR Channel. Inernaonal Conference on Spoken Language Processng (ICSLP, Ocober 2004, Jeju, Souh Korea. [14] Jason D. Wllams, Pascal Poupar, and Seve Young. Facored Parally Observable Markov Decson Processes for Dalogue Managemen. 4h Workshop on Knowledge and Reasonng n Praccal Dalog Sysems, Inernaonal Jon Conference on rfcal Inellgence (IJCI, ugus 2005, Ednburgh. [8] Ncholas Roy, Joelle Pneau and Sebasan Thrun. Spoken Dalogue Managemen Usng Probablsc Reasonng. nnual meeng of he ssocaon for Compuaonal Lnguscs (CL [9] Konrad Scheffler and Seve Young. uomac learnng of dalogue sraegy usng dalogue smulaon and renmen learnng. Proc. Human Language Technology (HLT-2002, San Dego, pp [10] Sander Sngh, Dane Lman, Mchael Kearns and Marlyn [15] Jason D. Wllams, Pascal Poupar, and Seve Young. Parally Observable Markov Decson Processes wh Connuous Observaons for Dalogue Managemen. In Proc. 6h SgDal Workshop on Dscourse and Dalogue, Sepember 2005, Lsbon. [16] Zhang Bo, Ca Qngsheng, Mao Janfeng, and Guo Banng. Plannng and cng under Uncerany: New Model for Spoken Dalogue Sysem. Proceedngs of he 17h nnual Conference on Uncerany n rfcal Inellgence (UI-01. San Francsco, US, s from s o Turn Noes b ( s from b ( b ( s o b ( [pror o dalog sar] S1: Where are you gong o? U1: London [msrecognzed as Leeds] S2: Where are you leavng from? U2: From Cambrdge [msrec as To Oxford] S3: Where are you leavng from? U3: From Cambrdge [reco ok] S4: To Leeds, s ha rgh? U4: No, o London [msrec as No, from Sheffeld] S5: Where are you gong o? U5: To London from Cambrdge [reco ok] [Sysem prns cke from London o Cambrdge] Pror o dalog sar, probably mass s spread evenly over all values. Top hypohess for o s now leeds; from has receved no nformaon. User model predcs hs obs. s unlkely: o(oxford ges mnmal belef mass. Top hypohess for from s now cambrdge. o(leeds reduced; from(sheffeld has small effec b/c of user model. Top hypohess for o s now london; from s now very ceran. CMB. CMB. CMB. CMB. CMB. CMB. Fgure 5: Example conversaon and belef saes. The user s ryng o ravel from Cambrdge o London.
Capacity Planning. Operations Planning
Operaons Plannng Capacy Plannng Sales and Operaons Plannng Forecasng Capacy plannng Invenory opmzaon How much capacy assgned o each producon un? Realsc capacy esmaes Sraegc level Moderaely long me horzon
More informationAn Anti-spam Filter Combination Framework for Text-and-Image Emails through Incremental Learning
An An-spam Fler Combnaon Framework for Tex-and-Image Emals hrough Incremenal Learnng 1 Byungk Byun, 1 Chn-Hu Lee, 2 Seve Webb, 2 Danesh Iran, and 2 Calon Pu 1 School of Elecrcal & Compuer Engr. Georga
More informationHow To Calculate Backup From A Backup From An Oal To A Daa
6 IJCSNS Inernaonal Journal of Compuer Scence and Nework Secury, VOL.4 No.7, July 04 Mahemacal Model of Daa Backup and Recovery Karel Burda The Faculy of Elecrcal Engneerng and Communcaon Brno Unversy
More informationThe Virtual Machine Resource Allocation based on Service Features in Cloud Computing Environment
Send Orders for Reprns o reprns@benhamscence.ae The Open Cybernecs & Sysemcs Journal, 2015, 9, 639-647 639 Open Access The Vrual Machne Resource Allocaon based on Servce Feaures n Cloud Compung Envronmen
More informationA Background Layer Model for Object Tracking through Occlusion
A Background Layer Model for Obec Trackng hrough Occluson Yue Zhou and Ha Tao Deparmen of Compuer Engneerng Unversy of Calforna, Sana Cruz, CA 95064 {zhou,ao}@soe.ucsc.edu Absrac Moon layer esmaon has
More informationMORE ON TVM, "SIX FUNCTIONS OF A DOLLAR", FINANCIAL MECHANICS. Copyright 2004, S. Malpezzi
MORE ON VM, "SIX FUNCIONS OF A DOLLAR", FINANCIAL MECHANICS Copyrgh 2004, S. Malpezz I wan everyone o be very clear on boh he "rees" (our basc fnancal funcons) and he "fores" (he dea of he cash flow model).
More informationSpline. Computer Graphics. B-splines. B-Splines (for basis splines) Generating a curve. Basis Functions. Lecture 14 Curves and Surfaces II
Lecure 4 Curves and Surfaces II Splne A long flexble srps of meal used by drafspersons o lay ou he surfaces of arplanes, cars and shps Ducks weghs aached o he splnes were used o pull he splne n dfferen
More informationBoosting for Learning Multiple Classes with Imbalanced Class Distribution
Boosng for Learnng Mulple Classes wh Imbalanced Class Dsrbuon Yanmn Sun Deparmen of Elecrcal and Compuer Engneerng Unversy of Waerloo Waerloo, Onaro, Canada y8sun@engmal.uwaerloo.ca Mohamed S. Kamel Deparmen
More informationPedro M. Castro Iiro Harjunkoski Ignacio E. Grossmann. Lisbon, Portugal Ladenburg, Germany Pittsburgh, USA
Pedro M. Casro Iro Harjunkosk Ignaco E. Grossmann Lsbon Porugal Ladenburg Germany Psburgh USA 1 Process operaons are ofen subjec o energy consrans Heang and coolng ules elecrcal power Avalably Prce Challengng
More information12/7/2011. Procedures to be Covered. Time Series Analysis Using Statgraphics Centurion. Time Series Analysis. Example #1 U.S.
Tme Seres Analyss Usng Sagraphcs Cenuron Nel W. Polhemus, CTO, SaPon Technologes, Inc. Procedures o be Covered Descrpve Mehods (me sequence plos, auocorrelaon funcons, perodograms) Smoohng Seasonal Decomposon
More informationINTERNATIONAL JOURNAL OF STRATEGIC MANAGEMENT
IJSM, Volume, Number, 0 ISSN: 555-4 INTERNATIONAL JOURNAL OF STRATEGIC MANAGEMENT SPONSORED BY: Angelo Sae Unversy San Angelo, Texas, USA www.angelo.edu Managng Edors: Professor Alan S. Khade, Ph.D. Calforna
More informationMethodology of the CBOE S&P 500 PutWrite Index (PUT SM ) (with supplemental information regarding the CBOE S&P 500 PutWrite T-W Index (PWT SM ))
ehodology of he CBOE S&P 500 PuWre Index (PUT S ) (wh supplemenal nformaon regardng he CBOE S&P 500 PuWre T-W Index (PWT S )) The CBOE S&P 500 PuWre Index (cker symbol PUT ) racks he value of a passve
More informationLecture 40 Induction. Review Inductors Self-induction RL circuits Energy stored in a Magnetic Field
ecure 4 nducon evew nducors Self-nducon crcus nergy sored n a Magnec Feld 1 evew nducon end nergy Transfers mf Bv Mechancal energy ransform n elecrc and hen n hermal energy P Fv B v evew eformulaon of
More informationSPC-based Inventory Control Policy to Improve Supply Chain Dynamics
Francesco Cosanno e al. / Inernaonal Journal of Engneerng and Technology (IJET) SPC-based Invenory Conrol Polcy o Improve Supply Chan ynamcs Francesco Cosanno #, Gulo Gravo #, Ahmed Shaban #3,*, Massmo
More informationKalman filtering as a performance monitoring technique for a propensity scorecard
Kalman flerng as a performance monorng echnque for a propensy scorecard Kaarzyna Bjak * Unversy of Souhampon, Souhampon, UK, and Buro Informacj Kredyowej S.A., Warsaw, Poland Absrac Propensy scorecards
More informationHEURISTIC ALGORITHM FOR SINGLE RESOURCE CONSTRAINED PROJECT SCHEDULING PROBLEM BASED ON THE DYNAMIC PROGRAMMING
Yugoslav Journal o Operaons Research Volume 19 (2009) Number 2, 281-298 DOI:10.2298/YUJOR0902281S HEURISTIC ALGORITHM FOR SINGLE RESOURCE CONSTRAINED PROJECT SCHEDULING PROBLEM BASED ON THE DYNAMIC PROGRAMMING
More informationPerfCenter: A Methodology and Tool for Performance Analysis of Application Hosting Centers
PerfCener: A Mehodology and Tool for Performance Analyss of Applcaon Hosng Ceners Rukma P. Verlekar, Varsha Ape, Prakhar Goyal, Bhavsh Aggarwal Dep. of Compuer Scence and Engneerng Indan Insue of Technology
More informationAn Architecture to Support Distributed Data Mining Services in E-Commerce Environments
An Archecure o Suppor Dsrbued Daa Mnng Servces n E-Commerce Envronmens S. Krshnaswamy 1, A. Zaslavsky 1, S.W. Loke 2 School of Compuer Scence & Sofware Engneerng, Monash Unversy 1 900 Dandenong Road, Caulfeld
More informationA Heuristic Solution Method to a Stochastic Vehicle Routing Problem
A Heursc Soluon Mehod o a Sochasc Vehcle Roung Problem Lars M. Hvaum Unversy of Bergen, Bergen, Norway. larsmh@.ub.no Arne Løkkeangen Molde Unversy College, 6411 Molde, Norway. Arne.Lokkeangen@hmolde.no
More informationProceedings of the 2008 Winter Simulation Conference S. J. Mason, R. R. Hill, L. Mönch, O. Rose, T. Jefferson, J. W. Fowler eds.
Proceedngs of he 008 Wner Smulaon Conference S. J. Mason, R. R. Hll, L. Mönch, O. Rose, T. Jefferson, J. W. Fowler eds. DEMAND FORECAST OF SEMICONDUCTOR PRODUCTS BASED ON TECHNOLOGY DIFFUSION Chen-Fu Chen,
More informationLinear Extension Cube Attack on Stream Ciphers Abstract: Keywords: 1. Introduction
Lnear Exenson Cube Aack on Sream Cphers Lren Dng Yongjuan Wang Zhufeng L (Language Engneerng Deparmen, Luo yang Unversy for Foregn Language, Luo yang cy, He nan Provnce, 47003, P. R. Chna) Absrac: Basng
More informationCooperative Distributed Scheduling for Storage Devices in Microgrids using Dynamic KKT Multipliers and Consensus Networks
Cooperave Dsrbued Schedulng for Sorage Devces n Mcrogrds usng Dynamc KK Mulplers and Consensus Newors Navd Rahbar-Asr Yuan Zhang Mo-Yuen Chow Deparmen of Elecrcal and Compuer Engneerng Norh Carolna Sae
More informationCost- and Energy-Aware Load Distribution Across Data Centers
- and Energy-Aware Load Dsrbuon Across Daa Ceners Ken Le, Rcardo Banchn, Margare Maronos, and Thu D. Nguyen Rugers Unversy Prnceon Unversy Inroducon Today, many large organzaons operae mulple daa ceners.
More informationHow Much Life Insurance is Enough?
How Much Lfe Insurance s Enough? Uly-Based pproach By LJ Rossouw BSTRCT The paper ams o nvesgae how much lfe nsurance proecon cover a uly maxmsng ndvdual should buy. Ths queson s relevan n he nsurance
More informationBoth human traders and algorithmic
Shuhao Chen s a Ph.D. canddae n sascs a Rugers Unversy n Pscaaway, NJ. bhmchen@sa.rugers.edu Rong Chen s a professor of Rugers Unversy n Pscaaway, NJ and Peng Unversy, n Bejng, Chna. rongchen@sa.rugers.edu
More informationInsurance. By Mark Dorfman, Alexander Kling, and Jochen Russ. Abstract
he Impac Of Deflaon On Insurance Companes Offerng Parcpang fe Insurance y Mar Dorfman, lexander Klng, and Jochen Russ bsrac We presen a smple model n whch he mpac of a deflaonary economy on lfe nsurers
More informationAnomaly Detection in Network Traffic Using Selected Methods of Time Series Analysis
I. J. Compuer Nework and Informaon Secury, 2015, 9, 10-18 Publshed Onlne Augus 2015 n MECS (hp://www.mecs-press.org/) DOI: 10.5815/jcns.2015.09.02 Anomaly Deecon n Nework Traffc Usng Seleced Mehods of
More informationThe Rules of the Settlement Guarantee Fund. 1. These Rules, hereinafter referred to as "the Rules", define the procedures for the formation
Vald as of May 31, 2010 The Rules of he Selemen Guaranee Fund 1 1. These Rules, herenafer referred o as "he Rules", defne he procedures for he formaon and use of he Selemen Guaranee Fund, as defned n Arcle
More informationCONTROLLER PERFORMANCE MONITORING AND DIAGNOSIS. INDUSTRIAL PERSPECTIVE
Copyrgh IFAC 5h Trennal World Congress, Barcelona, Span CONTROLLER PERFORMANCE MONITORING AND DIAGNOSIS. INDUSTRIAL PERSPECTIVE Derrck J. Kozub Shell Global Soluons USA Inc. Weshollow Technology Cener,
More informationAn Ensemble Data Mining and FLANN Combining Short-term Load Forecasting System for Abnormal Days
JOURNAL OF SOFTWARE, VOL. 6, NO. 6, JUNE 0 96 An Ensemble Daa Mnng and FLANN Combnng Shor-erm Load Forecasng Sysem for Abnormal Days Mng L College of Auomaon, Guangdong Unversy of Technology, Guangzhou,
More informationFundamental Analysis of Receivables and Bad Debt Reserves
Fundamenal Analyss of Recevables and Bad Deb Reserves Mchael Calegar Assocae Professor Deparmen of Accounng Sana Clara Unversy e-mal: mcalegar@scu.edu February 21 2005 Fundamenal Analyss of Recevables
More informationDEPARTMENT OF ECONOMETRICS AND BUSINESS STATISTICS. Exponential Smoothing for Inventory Control: Means and Variances of Lead-Time Demand
ISSN 440-77X ISBN 0 736 094 X AUSTRALIA DEPARTMENT OF ECONOMETRICS AND BUSINESS STATISTICS Exponenal Smoohng for Invenory Conrol: Means and Varances of Lead-Tme Demand Ralph D. Snyder, Anne B. Koehler,
More informationCase Study on Web Service Composition Based on Multi-Agent System
900 JOURNAL OF SOFTWARE, VOL. 8, NO. 4, APRIL 2013 Case Sudy on Web Servce Composon Based on Mul-Agen Sysem Shanlang Pan Deparmen of Compuer Scence and Technology, Nngbo Unversy, Chna PanShanLang@gmal.com
More informationGenetic Algorithm with Range Selection Mechanism for Dynamic Multiservice Load Balancing in Cloud-Based Multimedia System
ISSN : 2347-8446 (Onlne) Inernaonal Journal of Advanced Research n Genec Algorhm wh Range Selecon Mechansm for Dynamc Mulservce Load Balancng n Cloud-Based Mulmeda Sysem I Mchael Sadgun Rao Kona, II K.Purushoama
More informationMODEL-BASED APPROACH TO CHARACTERIZATION OF DIFFUSION PROCESSES VIA DISTRIBUTED CONTROL OF ACTUATED SENSOR NETWORKS
MODEL-BASED APPROACH TO CHARACTERIZATION OF DIFFUSION PROCESSES IA DISTRIBUTED CONTROL OF ACTUATED SENSOR NETWORKS Kevn L. Moore and YangQuan Chen Cener for Self-Organzng and Inellgen Sysems Uah Sae Unversy
More informationNetwork Effects on Standard Software Markets: A Simulation Model to examine Pricing Strategies
Nework Effecs on Sandard Sofware Markes Page Nework Effecs on Sandard Sofware Markes: A Smulaon Model o examne Prcng Sraeges Peer Buxmann Absrac Ths paper examnes sraeges of sandard sofware vendors, n
More informationGUIDANCE STATEMENT ON CALCULATION METHODOLOGY
GUIDANCE STATEMENT ON CALCULATION METHODOLOGY Adopon Dae: 9/28/0 Effecve Dae: //20 Reroacve Applcaon: No Requred www.gpssandards.org 204 CFA Insue Gudance Saemen on Calculaon Mehodology GIPS GUIDANCE STATEMENT
More information(Im)possibility of Safe Exchange Mechanism Design
(Im)possbly of Safe Exchange Mechansm Desgn Tuomas Sandholm Compuer Scence Deparmen Carnege Mellon Unversy 5 Forbes Avenue Psburgh, PA 15213 sandholm@cs.cmu.edu XaoFeng Wang Deparmen of Elecrcal and Compuer
More informationA Real-time Adaptive Traffic Monitoring Approach for Multimedia Content Delivery in Wireless Environment *
A Real-e Adapve Traffc Monorng Approach for Muleda Conen Delvery n Wreless Envronen * Boonl Adpa and DongSong Zhang Inforaon Syses Deparen Unversy of Maryland, Balore Couny Balore, MD, U.S.A. bdpa1@ubc.edu,
More informationPARTICLE FILTER BASED VEHICLE TRACKING APPROACH WITH IMPROVED RESAMPLING STAGE
ISS: 0976-910(OLIE) ICTACT JOURAL O IMAGE AD VIDEO PROCESSIG, FEBRUARY 014, VOLUME: 04, ISSUE: 03 PARTICLE FILTER BASED VEHICLE TRACKIG APPROACH WITH IMPROVED RESAMPLIG STAGE We Leong Khong 1, We Yeang
More informationIndex Mathematics Methodology
Index Mahemacs Mehodology S&P Dow Jones Indces: Index Mehodology Ocober 2015 Table of Conens Inroducon 4 Dfferen Varees of Indces 4 The Index Dvsor 5 Capalzaon Weghed Indces 6 Defnon 6 Adjusmens o Share
More informationGround rules. Guide to the calculation methods of the FTSE Actuaries UK Gilts Index Series v1.9
Ground rules Gude o he calculaon mehods of he FTSE Acuares UK Gls Index Seres v1.9 fserussell.com Ocober 2015 Conens 1.0 Inroducon... 4 1.1 Scope... 4 1.2 FTSE Russell... 5 1.3 Overvew of he calculaons...
More informationProt sharing: a stochastic control approach.
Pro sharng: a sochasc conrol approach. Donaen Hanau Aprl 2, 2009 ESC Rennes. 35065 Rennes, France. Absrac A majory of lfe nsurance conracs encompass a guaraneed neres rae and a parcpaon o earnngs of he
More informationOptimization of Nurse Scheduling Problem with a Two-Stage Mathematical Programming Model
Asa Pacfc Managemen Revew 15(4) (2010) 503-516 Opmzaon of Nurse Schedulng Problem wh a Two-Sage Mahemacal Programmng Model Chang-Chun Tsa a,*, Cheng-Jung Lee b a Deparmen of Busness Admnsraon, Trans World
More informationTHE USE IN BANKS OF VALUE AT RISK METHOD IN MARKET RISK MANAGEMENT. Ioan TRENCA *
ANALELE ŞTIINłIFICE ALE UNIVERSITĂłII ALEXANDRU IOAN CUZA DIN IAŞI Tomul LVI ŞnŃe Economce 009 THE USE IN BANKS OF VALUE AT RISK METHOD IN MARKET RISK MANAGEMENT Ioan TRENCA * Absrac In sophscaed marke
More informationA robust optimisation approach to project scheduling and resource allocation. Elodie Adida* and Pradnya Joshi
In. J. Servces Operaons and Informacs, Vol. 4, No. 2, 2009 169 A robus opmsaon approach o projec schedulng and resource allocaon Elode Adda* and Pradnya Josh Deparmen of Mechancal and Indusral Engneerng,
More informationEstimating intrinsic currency values
Cung edge Foregn exchange Esmang nrnsc currency values Forex marke praconers consanly alk abou he srenghenng or weakenng of ndvdual currences. In hs arcle, Jan Chen and Paul Dous presen a new mehodology
More informationAttribution Strategies and Return on Keyword Investment in Paid Search Advertising
Arbuon Sraeges and Reurn on Keyword Invesmen n Pad Search Adversng by Hongshuang (Alce) L, P. K. Kannan, Sva Vswanahan and Abhshek Pan * December 15, 2015 * Honshuang (Alce) L s Asssan Professor of Markeng,
More informationA Hybrid AANN-KPCA Approach to Sensor Data Validation
Proceedngs of he 7h WSEAS Inernaonal Conference on Appled Informacs and Communcaons, Ahens, Greece, Augus 4-6, 7 85 A Hybrd AANN-KPCA Approach o Sensor Daa Valdaon REZA SHARIFI, REZA LANGARI Deparmen of
More informationLevy-Grant-Schemes in Vocational Education
Levy-Gran-Schemes n Vocaonal Educaon Sefan Bornemann Munch Graduae School of Economcs Inernaonal Educaonal Economcs Conference Taru, Augus 26h, 2005 Sefan Bornemann / MGSE Srucure Movaon and Objecve Leraure
More informationCurrency Exchange Rate Forecasting from News Headlines
Currency Exchange Rae Forecasng from News Headlnes Desh Peramunelleke Raymond K. Wong School of Compuer Scence & Engneerng Unversy of New Souh Wales Sydney, NSW 2052, Ausrala deshp@cse.unsw.edu.au wong@cse.unsw.edu.au
More informationAnalysis of intelligent road network, paradigm shift and new applications
CONFERENCE ABOUT THE STATUS AND FUTURE OF THE EDUCATIONAL AND R&D SERVICES FOR THE VEHICLE INDUSTRY Analyss of nellgen road nework, paradgm shf and new applcaons Péer Tamás "Smarer Transpor" - IT for co-operave
More informationRevision: June 12, 2010 215 E Main Suite D Pullman, WA 99163 (509) 334 6306 Voice and Fax
.3: Inucors Reson: June, 5 E Man Sue D Pullman, WA 9963 59 334 636 Voce an Fax Oerew We connue our suy of energy sorage elemens wh a scusson of nucors. Inucors, lke ressors an capacors, are passe wo-ermnal
More informationThe Feedback from Stock Prices to Credit Spreads
Appled Fnance Projec Ka Fa Law (Keh) The Feedback from Sock Prces o Cred Spreads Maser n Fnancal Engneerng Program BA 3N Appled Fnance Projec Ka Fa Law (Keh) Appled Fnance Projec Ka Fa Law (Keh). Inroducon
More informationA Decision-Theoretic Generalization of On-Line Learning and an Application to Boosting*
journal of compuer and sysem scences 55, 119139 (1997) arcle no. SS971504 A Decson-heorec Generalzaon of On-Lne Learnng and an Applcaon o Boosng* Yoav Freund and Rober E. Schapre - A6 Labs, 180 Park Avenue,
More informationThe Performance of Seasoned Equity Issues in a Risk- Adjusted Environment?
The Performance of Seasoned Equy Issues n a Rsk- Adjused Envronmen? Allen, D.E., and V. Souck 2 Deparmen of Accounng, Fnance and Economcs, Edh Cowan Unversy, W.A. 2 Erdeon Group, Sngapore Emal: d.allen@ecu.edu.au
More informationA Hybrid Method for Forecasting Stock Market Trend Using Soft-Thresholding De-noise Model and SVM
A Hybrd Mehod for Forecasng Sock Marke Trend Usng Sof-Thresholdng De-nose Model and SVM Xueshen Su, Qnghua Hu, Daren Yu, Zongxa Xe, and Zhongyng Q Harbn Insue of Technology, Harbn 150001, Chna Suxueshen@Gmal.com
More informationUsing Cellular Automata for Improving KNN Based Spam Filtering
The Inernaonal Arab Journal of Informaon Technology, Vol. 11, No. 4, July 2014 345 Usng Cellular Auomaa for Improvng NN Based Spam Flerng Faha Bargou, Bouzane Beldjlal, and Baghdad Aman Compuer Scence
More informationModelling Operational Risk in Financial Institutions using Hybrid Dynamic Bayesian Networks. Authors:
Modellng Operaonal Rsk n Fnancal Insuons usng Hybrd Dynamc Bayesan Neworks Auhors: Professor Marn Nel Deparmen of Compuer Scence, Queen Mary Unversy of London, Mle nd Road, London, 1 4NS, Uned Kngdom Phone:
More informationLong Run Underperformance of Seasoned Equity Offerings: Fact or an Illusion?
Long Run Underperformance of Seasoned Equy Offerngs: Fac or an Illuson? 1 2 Allen D.E. and V. Souck 1 Edh Cowan Unversy, 2 Unversy of Wesern Ausrala, E-Mal: d.allen@ecu.edu.au Keywords: Seasoned Equy Issues,
More informationSensor Nework proposeations
008 Inernaoal Symposum on Telecommuncaons A cooperave sngle arge rackng algorhm usng bnary sensor neworks Danal Aghajaran, Reza Berang Compuer Engneerng Deparmen, Iran Unversy of Scence and Technology,
More informationThe Sarbanes-Oxley Act and Small Public Companies
The Sarbanes-Oxley Ac and Small Publc Companes Smry Prakash Randhawa * June 5 h 2009 ABSTRACT Ths sudy consrucs measures of coss as well as benefs of mplemenng Secon 404 for small publc companes. In hs
More informationWhat Explains Superior Retail Performance?
Wha Explans Superor Real Performance? Vshal Gaur, Marshall Fsher, Ananh Raman The Wharon School, Unversy of Pennsylvana vshal@grace.wharon.upenn.edu fsher@wharon.upenn.edu Harvard Busness School araman@hbs.edu
More informationEvent Based Project Scheduling Using Optimized Ant Colony Algorithm Vidya Sagar Ponnam #1, Dr.N.Geethanjali #2
Inernaonal Journal of Compuer Trends and Technology (IJCTT) Volume 18 Number 6 Dec 2014 Even Based Projec Schedulng Usng Opmzed An Colony Algorhm Vdya Sagar Ponnam #1, Dr.N.Geehanjal #2 1 Research Scholar,
More informationApplying the Theta Model to Short-Term Forecasts in Monthly Time Series
Applyng he Thea Model o Shor-Term Forecass n Monhly Tme Seres Glson Adamczuk Olvera *, Marcelo Gonçalves Trenn +, Anselmo Chaves Neo ** * Deparmen of Mechancal Engneerng, Federal Technologcal Unversy of
More informationOmar Shatnawi. Eks p l o a t a c j a i Ni e z a w o d n o s c Ma in t e n a n c e a n d Reliability Vo l.16, No. 4, 2014 585. 1.
Arcle caon nfo: Shanaw O. Measurng commercal sofware operaonal relably: an nerdscplnary modellng approach. Esploaacja Nezawodnosc Manenance and Relably 014; 16 (4): 585 594. Omar Shanaw Measurng commercal
More informationA 3D Model Retrieval System Using The Derivative Elevation And 3D-ART
3 Model Rereal Sysem Usng he erae leaon nd 3-R Jau-Lng Shh* ng-yen Huang Yu-hen Wang eparmen of ompuer Scence and Informaon ngneerng hung Hua Unersy Hsnchu awan RO -mal: sjl@chueduw bsrac In recen years
More informationContract design and insurance fraud: an experimental investigation *
Conrac desgn and nsurance fraud: an expermenal nvesgaon * Frauke Lammers and Jörg Schller Absrac Ths paper nvesgaes he mpac of nsurance conrac desgn on he behavor of flng fraudulen clams n an expermenal
More informationA Common Neural Network Model for Unsupervised Exploratory Data Analysis and Independent Component Analysis
A Common Neural Nework Model for Unsupervsed Exploraory Daa Analyss and Independen Componen Analyss Keywords: Unsupervsed Learnng, Independen Componen Analyss, Daa Cluserng, Daa Vsualsaon, Blnd Source
More informationY2K* Stephanie Schmitt-Grohé. Rutgers Uni ersity, 75 Hamilton Street, New Brunswick, New Jersey 08901 E-mail: grohe@econ.rutgers.edu.
Revew of Economc Dynamcs 2, 850856 Ž 1999. Arcle ID redy.1999.0065, avalable onlne a hp:www.dealbrary.com on Y2K* Sephane Schm-Grohé Rugers Unersy, 75 Hamlon Sree, New Brunswc, New Jersey 08901 E-mal:
More informationFinance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C.
Fnance and Economcs Dscusson Seres Dvsons of Research & Sascs and Moneary Affars Federal Reserve Board, Washngon, D.C. Prcng Counerpary Rs a he Trade Level and CVA Allocaons Mchael Pyhn and Dan Rosen 200-0
More informationThe Incentive Effects of Organizational Forms: Evidence from Florida s Non-Emergency Medicaid Transportation Programs
The Incenve Effecs of Organzaonal Forms: Evdence from Florda s Non-Emergency Medcad Transporaon Programs Chfeng Da* Deparmen of Economcs Souhern Illnos Unversy Carbondale, IL 62901 Davd Denslow Deparmen
More informationTime Series. A thesis. Submitted to the. Edith Cowan University. Perth, Western Australia. David Sheung Chi Fung. In Fulfillment of the Requirements
Mehods for he Esmaon of Mssng Values n Tme Seres A hess Submed o he Faculy of Communcaons, ealh and Scence Edh Cowan Unversy Perh, Wesern Ausrala By Davd Sheung Ch Fung In Fulfllmen of he Requremens For
More informationMULTI-WORKDAY ERGONOMIC WORKFORCE SCHEDULING WITH DAYS OFF
Proceedngs of he 4h Inernaonal Conference on Engneerng, Projec, and Producon Managemen (EPPM 203) MULTI-WORKDAY ERGONOMIC WORKFORCE SCHEDULING WITH DAYS OFF Tar Raanamanee and Suebsak Nanhavanj School
More informationTECNICHE DI DIAGNOSI AUTOMATICA DEI GUASTI. Silvio Simani silvio.simani@unife.it. References
TECNICHE DI DIAGNOSI AUTOMATICA DEI GUASTI Re Neural per l Idenfcazone d Ssem non Lnear e Paern Recognon slvo.sman@unfe. References Texbook suggesed: Neural Neworks for Idenfcaon, Predcon, and Conrol,
More informationThe Prediction Algorithm Based on Fuzzy Logic Using Time Series Data Mining Method
The Predcon Algorhm Based on Fuzzy Logc Usng Tme Seres Daa Mnng Mehod I Aydn, M Karakose, and E Akn Asrac Predcon of an even a a me seres s que mporan for engneerng and economy prolems Tme seres daa mnng
More informationDistribution Channel Strategy and Efficiency Performance of the Life insurance. Industry in Taiwan. Abstract
Dsrbuon Channel Sraegy and Effcency Performance of he Lfe nsurance Indusry n Tawan Absrac Changes n regulaons and laws he pas few decades have afeced Tawan s lfe nsurance ndusry and caused many nsurers
More informationSHIPPING ECONOMIC ANALYSIS FOR ULTRA LARGE CONTAINERSHIP
Journal of he Easern Asa Socey for Transporaon Sudes, Vol. 6, pp. 936-951, 2005 SHIPPING ECONOMIC ANALYSIS FOR ULTRA LARGE CONTAINERSHIP Chaug-Ing HSU Professor Deparen of Transporaon Technology and Manageen
More informationLoad Balancing in Internet Using Adaptive Packet Scheduling and Bursty Traffic Splitting
152 IJCSNS Inernaonal Journal of Compuer Scence and Nework Secury, VOL.8 No.1, Ocober 28 Load Balancng n Inerne Usng Adapve Packe Schedulng and Bursy Traffc Splng M. Azah Research Scholar, Anna Unversy,
More informationMarket-Clearing Electricity Prices and Energy Uplift
Marke-Clearng Elecrcy Prces and Energy Uplf Paul R. Grbk, Wllam W. Hogan, and Susan L. Pope December 31, 2007 Elecrcy marke models requre energy prces for balancng, spo and shor-erm forward ransacons.
More informationA GENERALIZED FRAMEWORK FOR CREDIT RISK PORTFOLIO MODELS
A GENERALIZED FRAMEWORK FOR CREDIT RISK PORTFOLIO MODELS H. UGUR KOYLUOGLU ANDREW HICKMAN Olver, Wyman & Company CSFP Capal, Inc. * 666 Ffh Avenue Eleven Madson Avenue New Yor, New Yor 10103 New Yor, New
More informationCLoud computing has recently emerged as a new
1 A Framework of Prce Bddng Confguraons for Resource Usage n Cloud Compung Kenl L, Member, IEEE, Chubo Lu, Keqn L, Fellow, IEEE, and Alber Y. Zomaya, Fellow, IEEE Absrac In hs paper, we focus on prce bddng
More informationAPPLICATION OF CHAOS THEORY TO ANALYSIS OF COMPUTER NETWORK TRAFFIC Liudvikas Kaklauskas, Leonidas Sakalauskas
The XIII Inernaonal Conference Appled Sochasc Models and Daa Analyss (ASMDA-2009) June 30-July 3 2009 Vlnus LITHUANIA ISBN 978-9955-28-463-5 L. Sakalauskas C. Skadas and E. K. Zavadskas (Eds.): ASMDA-2009
More informationNonlinearity or Structural Break? - Data Mining in Evolving Financial Data Sets from a Bayesian Model Combination Perspective
Proceedngs of he 38h Hawa Inernaonal Conference on Sysem Scences - 005 Nonlneary or Srucural Break? - Daa Mnng n Evolvng Fnancal Daa Ses from a Bayesan Model Combnaon Perspecve Hao Davd Zhou Managemen
More informationTHE IMPACT OF UNSECURED DEBT ON FINANCIAL DISTRESS AMONG BRITISH HOUSEHOLDS. Ana del Río and Garry Young. Documentos de Trabajo N.
THE IMPACT OF UNSECURED DEBT ON FINANCIAL DISTRESS AMONG BRITISH HOUSEHOLDS 2005 Ana del Río and Garry Young Documenos de Trabajo N.º 0512 THE IMPACT OF UNSECURED DEBT ON FINANCIAL DISTRESS AMONG BRITISH
More informationHAND: Highly Available Dynamic Deployment Infrastructure for Globus Toolkit 4
HAND: Hghly Avalable Dynamc Deploymen Infrasrucure for Globus Toolk 4 L Q 1, Ha Jn 1, Ian Foser,3, Jarek Gawor 1 Huazhong Unversy of Scence and Technology, Wuhan, 430074, Chna quck@chnagrd.edu.cn; hjn@hus.edu.cn
More informationNo. 32-2009. David Büttner and Bernd Hayo. Determinants of European Stock Market Integration
MAGKS Aachen Segen Marburg Geßen Göngen Kassel Jon Dscusson Paper Seres n Economcs by he Unverses of Aachen Geßen Göngen Kassel Marburg Segen ISSN 1867-3678 No. 32-2009 Davd Büner and Bernd Hayo Deermnans
More informationWho are the sentiment traders? Evidence from the cross-section of stock returns and demand. April 26, 2014. Luke DeVault. Richard Sias.
Who are he senmen raders? Evdence from he cross-secon of sock reurns and demand Aprl 26 2014 Luke DeVaul Rchard Sas and Laura Sarks ABSTRACT Recen work suggess ha senmen raders shf from less volale o speculave
More informationThe Multi-shift Vehicle Routing Problem with Overtime
The Mul-shf Vehcle Roung Problem wh Overme Yngao Ren, Maged Dessouy, and Fernando Ordóñez Danel J. Epsen Deparmen of Indusral and Sysems Engneerng Unversy of Souhern Calforna 3715 McClnoc Ave, Los Angeles,
More informationThe Definition and Measurement of Productivity* Mark Rogers
The Defnon and Measuremen of Producvy* Mark Rogers Melbourne Insue of Appled Economc and Socal Research The Unversy of Melbourne Melbourne Insue Workng Paper No. 9/98 ISSN 1328-4991 ISBN 0 7325 0912 6
More informationt φρ ls l ), l = o, w, g,
Reservor Smulaon Lecure noe 6 Page 1 of 12 OIL-WATER SIMULATION - IMPES SOLUTION We have prevously lsed he mulphase flow equaons for one-dmensonal, horzonal flow n a layer of consan cross seconal area
More informationWhat influences the growth of household debt?
Wha nfluences he growh of household deb? Dag Hennng Jacobsen, economs n he Secures Markes Deparmen, and Bjørn E. Naug, senor economs n he Research Deparmen 1 Household deb has ncreased by 10 11 per cen
More informationLinear methods for regression and classification with functional data
Lnear mehods for regresson and classfcaon wh funconal daa Glber Sapora Chare de Sasue Appluée & CEDRIC Conservaore Naonal des Ars e Méers 9 rue San Marn, case 44 754 Pars cedex 3, France sapora@cnam.fr
More informationModeling state-related fmri activity using change-point theory
Modelng sae-relaed fmri acvy usng change-pon heory Marn A. Lndqus 1*, Chrsan Waugh and Tor D. Wager 3 1. Deparmen of Sascs, Columba Unversy, New York, NY, 1007. Deparmen of Psychology, Unversy of Mchgan,
More informationWorking Paper Ageing, demographic risks, and pension reform. ETLA Discussion Papers, The Research Institute of the Finnish Economy (ETLA), No.
econsor www.econsor.eu Der Open-Access-Publkaonsserver der ZBW Lebnz-Informaonszenrum Wrschaf The Open Access Publcaon Server of he ZBW Lebnz Informaon Cenre for Economcs Lassla, Jukka; Valkonen, Tarmo
More informationAn Introductory Study on Time Series Modeling and Forecasting
An Inroducory Sudy on Tme Seres Modelng and Forecasng Ranadp Adhkar R. K. Agrawal ACKNOWLEDGEMENT The mely and successful compleon of he book could hardly be possble whou he helps and suppors from a lo
More informationRESOLUTION OF THE LINEAR FRACTIONAL GOAL PROGRAMMING PROBLEM
Revsa Elecrónca de Comuncacones y Trabajos de ASEPUMA. Rec@ Volumen Págnas 7 a 40. RESOLUTION OF THE LINEAR FRACTIONAL GOAL PROGRAMMING PROBLEM RAFAEL CABALLERO rafael.caballero@uma.es Unversdad de Málaga
More informationFOREIGN AID AND ECONOMIC GROWTH: NEW EVIDENCE FROM PANEL COINTEGRATION
JOURAL OF ECOOMIC DEVELOPME 7 Volume 30, umber, June 005 FOREIG AID AD ECOOMIC GROWH: EW EVIDECE FROM PAEL COIEGRAIO ABDULASSER HAEMI-J AD MAUCHEHR IRADOUS * Unversy of Skövde and Unversy of Örebro he
More informationBest estimate calculations of saving contracts by closed formulas Application to the ORSA
Bes esmae calculaons of savng conracs by closed formulas Applcaon o he ORSA - Franços BONNIN (Ala) - Frédérc LANCHE (Unversé Lyon 1, Laboraore SAF) - Marc JUILLARD (Wner & Assocés) 01.5 (verson modfée
More informationInformation-based trading, price impact of trades, and trade autocorrelation
Informaon-based radng, prce mpac of rades, and rade auocorrelaon Kee H. Chung a,, Mngsheng L b, Thomas H. McInsh c a Sae Unversy of New York (SUNY) a Buffalo, Buffalo, NY 426, USA b Unversy of Lousana
More information