Unsupervised Cross-Domain Transfer in Policy Gradient Reinforcement Learning via Manifold Alignment

Size: px
Start display at page:

Download "Unsupervised Cross-Domain Transfer in Policy Gradient Reinforcement Learning via Manifold Alignment"

Transcription

1 Unsupevsed Coss-Doman ansfe n Polcy Gaden Renfocemen Leanng va Manfold Algnmen Haham Bou Amma Unv. of Pennsylvana hahamb@seas.upenn.edu Ec Eaon Unv. of Pennsylvana eeaon@cs.upenn.edu Paul Ruvolo Oln College of Engneeng paul.uvolo@oln.edu Mahew E. aylo Washngon Sae Unv. aylom@eecs.wsu.edu Absac he success of applyng polcy gaden enfocemen leanng RL o dffcul conol asks hnges cucally on he ably o deemne a sensble nalzaon fo he polcy. ansfe leanng mehods ackle hs poblem by eusng knowledge gleaned fom solvng ohe elaed asks. In he case of mulple ask domans, hese algohms eque an ne-ask mappng o faclae knowledge ansfe acoss domans. Howeve, hee ae cuenly no geneal mehods o lean an ne-ask mappng whou equng ehe backgound knowledge ha s no ypcally pesen n RL sengs, o an expensve analyss of an exponenal numbe of ne-ask mappngs n he sze of he sae and acon spaces. hs pape noduces an auonomous famewok ha uses unsupevsed manfold algnmen o lean neask mappngs and effecvely ansfe samples beween dffeen ask domans. Empcal esuls on dvese dynamcal sysems, ncludng an applcaon o quadoo conol, demonsae s effecveness fo coss-doman ansfe n he conex of polcy gaden RL. Inoducon Polcy gaden enfocemen leanng RL algohms have been appled wh consdeable success o solve hghdmensonal conol poblems, such as hose asng n oboc conol and coodnaon Pees & Schaal hese algohms use gaden ascen o une he paamees of a polcy o maxmze s expeced pefomance. Unfounaely, hs gaden ascen pocedue s pone o becomng apped n local maxma, and hus has been wdely ecognzed ha nalzng he polcy n a sensble manne s cucal fo achevng opmal pefomance. o nsance, one ypcal saegy s o nalze he polcy usng human demonsaons Pees & Schaal 2006, whch may be nfeasble when he ask canno be easly solved by a human. hs pape exploes a dffeen appoach: nsead of nalzng he polcy a andom.e., abula asa o va human demonsaons, we nsead use ansfe leanng L o nalze he polcy fo a new age doman based on knowledge fom one o moe souce asks. In RL ansfe, he souce and age asks may dffe n he fomulaons aylo & Sone In pacula, Copygh c 2015, Assocaon fo he Advancemen of Afcal Inellgence All ghs eseved. when he souce and age asks have dffeen sae and/o acon spaces, an ne-ask mappng aylo e al. 2007a ha descbes he elaonshp beween he wo asks s ypcally needed. hs pape noduces a famewok fo auonomously leanng an ne-ask mappng fo coss-doman ansfe n polcy gaden RL. s, we lean an ne-sae mappng.e., a mappng beween saes n wo asks usng unsupevsed manfold algnmen. Manfold algnmen povdes a poweful and geneal famewok ha can dscove a shaed laen epesenaon o capue nnsc elaons beween dffeen asks, especve of he dmensonaly. he algnmen also yelds an mplc ne-acon mappng ha s geneaed by mappng ackng saes fom he souce o he age. Gven he mappng beween ask domans, souce ask aecoes ae hen used o nalze a polcy n he age ask, sgnfcanly mpovng he speed of subsequen leanng ove an unnfomed nalzaon. hs pape povdes he followng conbuons. s, we noduce a novel unsupevsed mehod fo leanng nesae mappngs usng manfold algnmen. Second, we show ha he dscoveed subspace can be used o nalze he age polcy. hd, ou empcal valdaon conduced on fou dssmla and dynamcally chaoc ask domans e.g., conollng a hee-lnk ca-pole and a quadoo aeal vehcle shows ha ou appoach can a auomacally lean an ne-sae mappng acoss MDPs fom he same doman, b auomacally lean an ne-sae mappng acoss MDPs fom vey dffeen domans, and c ansfe nfomave nal polces o acheve hghe nal pefomance and educe he me needed fo convegence o nea-opmal behavo. Relaed Wok Leanng an ne-ask mappng has been of mao nees n he ansfe leanng communy because of s pomse of auonomous ansfe beween vey dffeen asks aylo & Sone Howeve, he maoy of exsng wok assumes ha a he souce ask and age ask ae smla enough ha no mappng s needed Baneee & Sone 2007; Kondas & Bao 2007, o b an ne-ask mappng s povded o he agen aylo e al. 2007a; oey e al he man dffeence beween hese mehods and hs pape s ha we ae neesed n leanng a mappng beween asks. hee has been some ecen wok on leanng such mappngs. o example, mappngs may be based on seman-

2 c knowledge abou sae feaues beween wo asks Lu & Sone 2006, backgound knowledge abou he ange o ype of sae vaables aylo e al. 2007b, o anson models fo each possble mappng could be geneaed and esed aylo e al Howeve, hee ae cuenly no geneal mehods o lean an ne-ask mappng whou equng ehe backgound knowledge ha s no ypcally pesen n RL sengs, o an expensve analyss of an exponenal numbe n he sze of he acon and sae vaable ses of ne-ask mappngs. We ovecome hese ssues by auomacally dscoveng hgh-level feaues and usng hem o ansfe knowledge beween agens whou suffeng fom an exponenal exploson. In pevous wok, we used spase codng, spase poecon, and spase Gaussan pocesses o lean an ne-ask mappng beween MDPs wh abay vaaons Bou Amma e al Howeve, hs pevous wok eled on a Eucldean dsance coelaon beween souce and age ask ples, whch may fal fo hghly dssmla asks. Addonally, placed escons on he ne-ask mappng ha educed he flexbly of he leaned mappng. In ohe elaed wok, Bósc e al use manfold algnmen o asss n ansfe. he pmay dffeences wh ou wok ae ha he auhos a focus on ansfeng models beween dffeen obos, ahe han polces/samples, and b ely on souce and age obos ha ae qualavely smla. Backgound Renfocemen Leanng poblems nvolve an agen choosng sequenal acons o maxmze s expeced eun. Such poblems ae ypcally fomalzed as a Makov decson pocess MDP = S, A, P 0, P,, whee S s he poenally nfne se of saes, A s he se of acons ha he agen may execue, P 0 : S [0, 1] s a pobably dsbuon ove he nal sae, P : S A S [0, 1] s a sae anson pobably funcon descbng he ask dynamcs, and : S A S R s he ewad funcon measung he pefomance of he agen. A polcy π : S A [0, 1] s defned as a condonal pobably dsbuon ove acons gven he cuen sae. he agen s goal s o fnd a polcy π whch maxmzes he aveage expeced ewad: π = ag max π = ag max π E [ 1 H H ] s, a, s +1 π =1 p π τ Rτ dτ, whee s he se of all possble aecoes wh hozon H, Rτ = 1 H s, a, s +1, and 2 H =1 p π τ = P 0 s 1 1 H Ps +1 s, a πa s. 3 =1 Polcy Gaden mehods Suon e al. 1999; Pees e al epesen he agen s polcy π as a funcon defned ove a veco θ R d of conol paamees and a veco of sae feaues gven by he ansfomaon Φ : S R m. By subsung hs paameezaon of he conol polcy no Eqn. 2, we can compue he paamees of he opmal polcy as θ = ag max θ J θ, whee J θ = p πθτ Rτ dτ. o maxmze J, many polcy gaden mehods employ sandad supevsed funcon appoxmaon o lean θ by followng an esmaed gaden of a lowe bound on he expeced eun of J θ. Polcy gaden algohms have ganed aenon n he RL communy n pa due o he successful applcaons on eal-wold obocs Pees e al Whle such algohms have a low compuaonal cos pe updae, hghdmensonal poblems eque many updaes by acqung new ollous o acheve good pefomance. ansfe leanng can educe hs daa equemen and acceleae leanng. Snce polcy gaden mehods ae pone o becomng suck n local maxma, s cucal ha he polcy be nalzed n a sensble fashon. A common echnque Pees & Schaal 2006; Agall e al fo polcy nalzaon s o fs collec demonsaons fom a human conollng he sysem, hen use supevsed leanng o f polcy paamees ha maxmze he lkelhood of he human-demonsaed acons, and fnally use he fed paamees as he nal polcy paamees fo a polcy gaden algohm. Whle hs appoach woks well n some sengs, s napplcable n seveal common cases: a when s dffcul o nsumen he sysem n queson so ha a human can successfully pefom a demonsaon, b when an agen s consanly faced wh new asks, makng gaheng human demonsaons fo each new ask mpaccal, o c when he asks n queson canno be nuvely solved by a human demonsao. he nex secon noduces a mehod fo usng ansfe leanng o nalze he paamees of a polcy n a way ha s no suscepble o hese lmaons. Ou expemenal esuls show ha hs mehod of polcy nalzaon, when compaed o andom polcy nalzaon, s able o no only acheve bee nal pefomance, bu also oban a hghe pefomng polcy when un unl convegence. Polcy Gaden ansfe Leanng ansfe leanng ams o mpove leanng mes and/o behavo of an agen on a new age ask by eusng knowledge fom a solved souce ask. In RL sengs, each ask s descbed by an MDP: ask = S, A, P 0, P, and = S, A, P 0, P,. One way n whch knowledge fom a solved souce ask can be leveaged o solve he age ask s by mappng opmal sae, acon, nex sae ples fom he souce ask no he sae and acon spaces of he age ask. ansfeng opmal ples n hs way allows us o boh povde a bee umpsa and leanng ably o he age agen, based on he souce agen s ably. Whle he pecedng dea s aacve, complexes ase when he souce and age asks have dffeen sae and/o acon spaces. In hs case, one mus defne an ne-ask mappng χ n ode o anslae opmal ples fom he souce o he age ask. ypcally aylo & Sone 2009, χ s defned by wo sub-mappngs: 1 an ne-sae mappng χ S and 2 an ne-acon mappng χ A.

3 Souce Doman Phase I: Lean coss-doman mappng GS P0 S Phase II: Coss-doman ansfe va 2. eflec age aces fom age S 1. sample nal saes P execue shaed epesenaon aces fom age Doman + 4. ansfe ackng sgnal G gue 1: ansfe s spl no wo phases: I leanng he ne-sae mappng χs va manfold algnmen, and II nalzng he age polcy va mappng he souce ask polcy. By adopng an RL famewok whee polces ae saefeedback conolles, we show ha we can use opmal sae aecoes fom he souce ask o nellgenly nalze a conol polcy n he age ask, whou needng o explcly consuc an ne-acon mappng. We accomplsh hs by leanng a pseudo-nveble ne-sae mappng beween he sae spaces of a pa of asks usng manfold algnmen, whch can hen be used o ansfe opmal sequences of saes o he age. he fac ha ou algohm does no eque leanng an explc ne-acon mappng sgnfcanly educes s compuaonal complexy. Ou appoach consss of wo phases gue 1. s, usng aces gaheed n he souce and age asks, we lean an ne-sae mappng χs usng manfold algnmen Phase I n gue 1. o pefom hs sep, we adap he Unsupevsed Manfold Algnmen UMA algohm Wang & Mahadevan 2009, as dealed n he nex secon. Second, we use χs o poec sae aecoes fom he souce o he age ask Phase II n gue 1. hese poeced sae aecoes defne a se of a ackng aecoes fo he age ask ha allow us o pefom one sep of polcy gaden mpovemen n he age ask. hs polcy mpovemen sep nellgenly nalzes he age polcy, whch esuls n supeo leanng pefomance han sang fom a andomly nalzed polcy, as shown n ou expemens. Alhough we focus on polcy gaden mehods, ou appoach could easly be adaped o ohe polcy seach mehods e.g., PoWER, REPS, ec.; see Kobe e al Leanng an Ine-Sae Mappng Unsupevsed Manfold Algnmen UMA s a echnque ha effcenly dscoves an algnmen beween wo daases Wang & Mahadevan UMA was developed o algn daases fo knowledge ansfe beween wo supevsed leanng asks. Hee, we adap UMA o an RL seng by algnng souce and age ask sae spaces wh poenally dffeen dmensons ms and m. o lean χs elang S and n S, aecoes of o saes n he souce,, ns ask, τ = s1,..., shs, ae obaned by =1 followng π, and aecoes of saes n he age ask, τ = n on,, s1, ae obaned by ulz,..., sh =1 ng π, whch s nalzed usng andomly seleced polcy paamees. o smplcy of exposon, we assume ha aecoes n he souce doman have lengh HS and hose n he age doman have lengh H ; howeve, ou algohm s capable of handlng vaable-lengh aecoes. We ae neesed n he seng whee daa s scace n he age ask han n he souce ask.e., n ns. Gven aecoes fom boh he souce and age asks, we flaen he aecoes.e., we ea he saes as unodeed and hen apply he ask-specfc sae ansfomaon o oban wo ses of sae feaue vecos, one fo he souce ask and one fo he age ask. Specfcally, we ceae he followng ses of pons: 1 1 X = Φ s1,..., Φ shs, ns ns shs s1,,φ Φ 1 1 X = Φ s1,..., Φ sh, n n Φ s1,,φ sh. Gven X RmS HS ns, X Rm H n, we can apply he UMA algohm Wang & Mahadevan 2009 wh mnmal modfcaon, as descbed nex. Unsupevsed Manfold Algnmen UMA he fs sep of applyng UMA o lean he ne-sae mappng s o epesen each ansfomed sae n boh he souce and age asks n ems of s local geomey. We use he noaon Rx Rk+1 k+1 o efe o he max of pa wse Eucldean dsances among he k-neaes neghbos of x X. Smlaly, Rx efes o he equvalen ma x of dsances fo he k-neaes neghbos of x X. he elaons beween local geomees n X and X ae epesened RnS HS n H wh n byhe max Wo w, = exp ds Rx, Rx and dsance mec ds Rx, Rx = " mn mn orx oh γ1 Rx, 4 1 h k! Rx γ2 orx oh #. We use he noaon o oh o denoe he hh vaan of he k! pemuaons of he ows and columns of he npu max, s he obenus nom, and γ1 and γ2 ae defned as: R o R R o or oh h x x x x. γ1 = γ2 = Rx Rx orx oh orx oh

4 o algn he manfolds, UMA compues he on Laplacan LX + µγ L = 1 µγ 2 µγ 3 L X + µγ 4 5 wh dagonal maces Γ 1 Γ 4 R n H n H, whee Γ 1, R n S H S n S H S and = w, and Γ 4, = w,. he maces Γ 2 R n S H S n H and Γ 3 R n H n S H S on he wo manfolds wh Γ 2, = w, and Γ 3, = w,. Addonally, he non-nomalzed Laplacans L X and L X ae defned as: L X = D X W X and L X = D X W X, whee D X R n S H S n S H S s a dagonal max wh D, = X w, and, smlaly, D, = X w,. he maces W and W epesen he smlay n he souce and age ask sae spaces especvely and can be compued smla o W. o on he manfolds, UMA fs defnes wo maces: τ Z = 0 0 τ DS 0 D = 0 D S. 6 Gven Z and D, UMA compues opmal poecons o educe he dmensonaly of he on sucue by akng he d mnmum egenvecos ζ 1,..., ζ d of he genealzed egenvalue decomposon ZLZ ζ = λzdz ζ. he opmal poecons α and α ae hen gven as he fs d 1 and d 2 ows of [ζ 1,..., ζ d ], especvely. Gven he embeddng dscoveed by UMA, we can hen defne he ne-sae mappng as: χ S [ ] = α + α [ ]. 7 he nvese of he ne-sae mappng o poec age saes o he souce ask can be deemned by akng he pseudo-nvese of Eqn. 7, yeldng χ + S [ ] = α+ α [ ]. Inuvely, hs appoach algns he mpoan egons of he souce ask s sae space sampled based on opmal souce aecoes wh he sae space exploed so fa n he age ask. Alhough acons wee gnoed n consucng he manfolds, he algned epesenaon mplcly capues local sae anson dynamcs whn each ask snce he saes came fom aecoes, povdng a mechansm o ansfe aecoes beween asks, as we descbe nex. Polcy ansfe and Impovemen Nex, we dscuss he pocedue fo nalzng he age polcy, π. We consde a model-fee seng n whch he polcy s lnea ove a se of poenally non-lnea sae feaue funcons modulaed by Gaussan nose whee he magnude of he nose balances exploaon and exploaon. Specfcally, we can we he souce and age polces as: π π s s = Φ s = Φ s θ + ɛ θ + ɛ, whee ɛ N 0, Σ and ɛ N 0, Σ. τ =1 o nalze π, we fs sample m nal age aecoes D = fom he age ask usng a { } m andomly nalzed polcy hese can be newly sampled aecoes o smply he ones used o do he nal manfold algnmen sep. Nex, we map he se of nal saes n D o he souce ask usng χ + S. We hen un he opmal souce polcy sang fom each of hese mapped nal saes o poduce a se of m opmal sae aecoes n he souce ask. nally, he esulng sae aecoes ae mapped back o he age ask usng χ S o geneae a se of efleced sae- { aecoes n he age ask, D = τ } m =1. o clay, we assume ha all aecoes ae of lengh H; howeve, hs s no a fundamenal lmaon of ou algohm. We defne he followng ansfe cos funcon: J θ = m =1 p θ τ ˆR τ, τ whee ˆR s a cos funcon ha penalzes devaons beween he nal sampled aecoes n he age ask and he efleced opmal aecoes: ˆR τ, τ = 1 H H s s = Mnmzng, Eqn. 8 s equvalen o aanng a age polcy paameezaon θ such ha π follows he efleced aecoes D. uhe, Eqn. 8 s n exacly he fom equed o apply sandad off-he-shelf polcy gaden algohms o mnmze he ansfe cos. he Manfold Algnmen Coss-Doman ansfe fo Polcy Gadens MAXD-PG famewok s dealed 1 n Algohm 1. Specal Cases Ou wok can be seen as an exenson of he smple modelbased case wh a lnea-quadac egulao LQR Bempoad e al polcy, whch s deved and explaned n he onlne appendx 2 accompanyng hs pape. Alhough he assumpons made by he model-based case seem escve, he analyss n he appendx coves a wde ange of applcaons. hese, fo example, nclude: a he case n whch a dynamcal model s povded befoehand, o b he case n whch model-based RL algohms ae adoped see Buşonu e al In he man pape, howeve, we consde he moe geneal model-fee case. Expemens and Resuls o assess MAXD-PG s pefomance, we conduced expemens on ansfe boh beween asks n he same doman as well as beween asks n dffeen domans. Also, we suded 1 Lnes 9-11 of Algohm 1 eque neacon wh he age doman o a smulao fo acqung he opmal polcy. Such an assumpon s common o polcy gaden mehods, whee a each eaon, daa s gaheed and used o eavely mpove he polcy. 2 he onlne appendx s avalable on he auhos webses.

5 Algohm 1 Manfold Algnmen Coss-Doman ansfe fo Polcy Gadens MAXD- PG Inpus: Souce and age asks and, opmal souce polcy π, # souce and age aces ns and n, # neaes neghbos k, # age ollous z, nal # of age saes m. Lean χs : 1: Sample ns opmal souce aces, τ, and n andom age aces, τ 2: Usng he modfed UMA appoach, lean α and α o poduce χs = α+ α [ ] ansfe & Inalze Polcy: 3: Collec m nal age saes s1 P0 4: Poec hese m saes o he souce by applyng χ+ S [ ] 5: Apply he opmal souce polcy π on hese poeced n om saes o collec D = τ =1 6: Poec he samples n D o he age usng χs [ ] o poduce ackng age aces D 7: Compue ackng ewads usng Eqn : Use polcy gadens o mnmze Eqn. 8, yeldng θ Impove Polcy: 0 9: Sa wh θ and sample z age ollous 10: ollow polcy gadens e.g., epsodc REINORCE bu usng age ewads R 11: Reun opmal age polcy paamees θ he obusness of he leaned mappng by vayng he numbe of souce and age samples used fo ansfe and measung he esulan age ask pefomance. In all cases we compaed he pefomance of MAXD- PG o sandad polcy gaden leanes. Ou esuls show ha MAXD- PG was able o: a lean a vald ne-sae mappng wh elavely lle daa fom he age ask, and b effecvely ansfe beween asks fom ehe he same o dffeen domans. Dynamcal Sysem Domans We esed MAXD- PG and sandad polcy gaden leanng on fou dynamcal sysems gue 2. On all sysems, he ewad funcon was based on wo facos: a penalzng saes fa fom he goal sae, and b penalzng hgh foces acons o encouage smooh, low-enegy movemens. Smple Mass Spng Dampe SM: he goal wh he SM s o conol he mass a a specfed poson wh zeo velocy. he sysem dynamcs ae descbed by wo saevaables ha epesen he mass poson and velocy, and a sngle foce ha acs on he ca n he x decon. Ca Pole CP: he goal s o swng up and hen balance he pole vecally. he sysem dynamcs ae descbed va a fou-dmensonal sae veco hx, x, θ, θ, epesenng he poson, velocy of he ca, and he angle and angula velocy of he pole, especvely. he acons conss of a foce ha acs on he ca n he x decon. h, h, h, hx, x hx, x hx, x a Smple Mass hx, x h 3, h 33, 3 h 3, 3 h 3h, 3, h 2 22, 2, 2, 2 2 h 1, h 11, 1 h 2h h 1, h 11, 1 hx, x hx, hx, x x hx, x c hee-lnk Ca Pole h, 2 hx, x hx, x hx, x b Ca Pole x e11hx,e11 e2 e ee11 e ee31 e e 21oll o1ll e l ol e2b e 2B ee12be 2Be1 B e1 e B B 1pB c pc 4 ph 4 h e3be e3b ch p4 3B ch 4 yawyaw yaw e3b oll yaw d Quadoo gue 2: Dynamcal sysems used n he expemens. hee-lnk Ca Pole 3CP: he 3CP dynamcs ae descbed va an egh-dmensonal sae veco hx, x, θ1, θ 1, θ2, θ 2, θ3, θ 3, whee x and x descbe he poson and velocy of he ca and θ and θ epesen he angle and angula velocy of he h lnk. he sysem s conolled by applyng a foce o he ca n he x decon, wh he goal of balancng he hee poles upgh. Quadoo QR: he sysem dynamcs wee adoped fom a smulao valdaed on eal quadoos Bouabdallah 2007; Voos & Bou Amma 2010, and ae descbed va hee angles and hee angula veloces n he body fame.e., e1b, e2b, and e3b. he acons conss of fou oo oques {1, 2, 3, 4 }. Each ask coesponds o a dffeen quadoo confguaon e.g., dffeen amaue lenghs, ec., and he goal s o sablze he dffeen quadoos. Same-Doman ansfe We fs evaluae MAXD- PG on same-doman ansfe. Whn each doman, we can oban dffeen asks by vayng he sysem paamees e.g., fo he SM sysem we vaed mass M, spng consan K, and dampng consan b as well as he ewad funcons. We assessed he pefomance of usng he ansfeed polcy fom MAXD- PG vesus sandad polcy gadens by measung he aveage ewad on he age ask vs. he amoun of leanng eaons n he age. We also examned he obusness of MAXD- PG s pefomance based on he numbe of souce and age samples used o lean χs. Rewads wee aveaged ove 500 aces colleced fom 150 nal saes. Due o space consans, we epo same-doman ansfe esuls hee; deals of he asks and expemenal pocedue can be found n he appendx2. gue 3 shows MAXD- PG s pefomance usng vayng numbes of souce and age samples o lean χs. hese esuls eveal ha ansfe-nalzed polces oupefom sandad polcy gaden nalzaon. uhe, as he numbe of samples used o lean χs nceases, so does boh he nal and fnal pefomance n all domans. All nalzaons esul n equal pe-eaon compuaonal cos. heefoe, MAXD- PG boh mpoves sample complexy and educes wall-clock leanng me.

6 Aveage Rewad Souce 1000 age Samples 500 Souce 500 age Samples Souce 300 age Samples Souce 100 age Samples Sandad Polcy Gadens Ieaons Ieaons Ieaons Ieaons 3000 a Smple Mass b Ca Pole c hee-lnk Ca Pole d Quadoo gue 3: Same-doman ansfe esuls. All plos shae he same legend and vecal axs label. Aveage Rewad Souce 1000 age Samples Souce 500 age Samples 500 Souce 300 age Samples 100 Souce 100 age Samples 315 Sandad Polcy Gadens Ieaons Ieaons Ieaons a Smple Mass o Ca Pole 305 b Ca Pole o hee-lnk CP c Ca Pole o Quadoo θ θ * age: SM age: CP age: 3CP Souce: SM Souce: CP Souce: 3CP Pocuses Measue d Algnmen Qualy vs ansfe gue 4: Coss-doman ansfe esuls. Plos a c depc age ask pefomance, and shae he same legend and axs labels. Plo d shows he coelaon beween manfold algnmen qualy Pocuses mec and qualy of he ansfeed knowledge. Coss-Doman ansfe Nex, we consde he moe dffcul poblem of cossdoman ansfe. he expemenal seup s dencal o he same-doman case wh he cucal dffeence ha he sae and/o acon spaces wee dffeen fo he souce and he age ask snce he asks wee fom dffeen domans. We esed hee coss-doman ansfe scenaos: smple mass o ca pole, ca pole o hee-lnk ca pole, and ca pole o quadoo. In each case, he souce and age ask have dffeen numbes of sae vaables and sysem dynamcs. Deals of hese expemens ae avalable n he appendx 2. gue 4 shows he esuls of coss-doman ansfe, demonsang ha MAXD-PG can acheve successful ansfe beween dffeen ask domans. hese esuls enfoce he conclusons of he same-doman ansfe expemens, showng ha a ansfe-nalzed polces oupefom sandad polcy gadens, even beween dffeen ask domans and b nal and fnal pefomance mpoves as moe samples ae used o lean χ S. We also examned he coelaon beween he qualy of he manfold algnmen, as assessed by he Pocuses mec Goldbeg & Rov 2009, and he qualy of he ansfeed knowledge, as measued by he dsance beween he ansfeed θ and he opmal θ paamees gue 4d. On boh measues, smalle values ndcae bee qualy. Each daa pon epesens a ansfe scenao beween wo dffeen asks, fom ehe SM, CP, o 3CP; we dd no consde quadoo asks due o he equed smulao me. Alhough we show ha he Pocuses measue s posvely coelaed wh ansfe qualy, we hesae o ecommend as a pedcve measue of ansfe pefomance. In ou appoach, he coss-doman mappng s no guaaneed o be ohogonal, and heefoe he Pocuses measue s no heoecally guaaneed o accuaely measue he qualy of he global embeddng.e., Goldbeg and Rov s 2009 Coollay 1 s no guaaneed o hold, bu he Pocuses measue sll appeas coelaed wh ansfe qualy n pacce. We can conclude ha MAXD-PG s capable of: a auomacally leanng an ne-sae mappng, and b effecvely ansfeng beween dffeen doman sysems. Even when he souce and age asks ae hghly dssmla e.g., ca pole o quadoo, MAXD-PG s capable of successfully povdng age polcy nalzaons ha oupefom saeof-he-a polcy gaden echnques. Concluson We noduced MAXD-PG, a echnque fo auonomous ansfe beween polcy gaden RL algohms. MAXD-PG employs unsupevsed manfold algnmen o lean an nesae mappng, whch s hen used o ansfe samples and nalze he age ask polcy. MAXD-PG s pefomance was evaluaed on fou dynamcal sysems, demonsang ha MAXD-PG s capable of mpovng boh an agen s nal and fnal pefomance elave o usng polcy gaden algohms whou ansfe, even acoss dffeen domans.

7 Acknowledgemens hs eseach was suppoed by ONR gan #N , AOSR gan #A , and NS gan IIS We hank he anonymous evewes fo he helpful feedback. Refeences Agall, B. D.; Chenova, S.; Veloso, M.; and Bownng, B A suvey of obo leanng fom demonsaon. Robocs and Auonomous Sysems 575: Baneee, B., and Sone, P Geneal game leanng usng knowledge ansfe. In Poceedngs of he 20h Inenaonal Jon Confeence on Afcal Inellgence, Bempoad, A.; Moa, M.; Dua, V.; and Pskopoulos, E he explc lnea quadac egulao fo consaned sysems. Auomaca 381:3 20. Bócs, B.; Csao, L.; and Pees, J Algnmen-based ansfe leanng fo obo models. In Poceedngs of he Inenaonal Jon Confeence on Neual Newoks IJCNN. Bou Amma, H.; aylo, M.; uyls, K.; Dessens, K.; and Wess, G Renfocemen leanng ansfe va spase codng. In Poceedngs of he 11h Confeence on Auonomous Agens and Mulagen Sysems AAMAS. Bouabdallah, S Desgn and Conol of Quadoos wh Applcaon o Auonomous lyng. Ph.D. Dsseaon, École polyechnque fédéale de Lausanne. Buşonu, L.; Babuška, R.; De Schue, B.; and Ens, D Renfocemen Leanng and Dynamc Pogammng Usng uncon Appoxmaos. Boca Raon, loda: CRC Pess. Goldbeg, Y.; and Rov, Y Local Pocuses fo manfold embeddng: a measue of embeddng qualy and embeddng algohms. Machne Leanng 771: Kobe, J.; Bagnell, A.; and Pees, J Renfocemen leanng n obocs: a suvey. Inenaonal Jounal of Robocs Reseach 3211: Kondas, G., and Bao, A Buldng poable opons: Skll ansfe n enfocemen leanng. In Poceedngs of he 20h Inenaonal Jon Confeence on Afcal Inellgence, Lu, Y., and Sone, P Value-funcon-based ansfe fo enfocemen leanng usng sucue mappng. In Poceedngs of he 21s Naonal Confeence on Afcal Inellgence, Pees, J., and Schaal, S Polcy gaden mehods fo obocs. In Poceedngs of he IEEE/RSJ Inenaonal Confeence on Inellgen Robos and Sysems, Pees, J., and Schaal, S Naual aco-cc. Neuocompung 717-9: Pees, J.; Vayakuma, S.; and Schaal, S Naual aco-cc. In Poceedngs of he 16h Euopean Confeence on Machne Leanng ECML, Spnge. Suon, R. S.; McAllese, D. A.; Sngh, S. P.; and Mansou, Y Polcy gaden mehods fo enfocemen leanng wh funcon appoxmaon. Neual Infomaon Pocessng Sysems aylo, M. E., and Sone, P ansfe leanng fo enfocemen leanng domans: a suvey. Jounal of Machne Leanng Reseach 10: aylo, M. E.; Kuhlmann, G.; and Sone, P Auonomous ansfe fo enfocemen leanng. In Poceedngs of he 7h Inenaonal Jon Confeence on Auonomous Agens and Mulagen Sysems AAMAS, aylo, M. E.; Sone, P.; and Lu, Y ansfe leanng va ne-ask mappngs fo empoal dffeence leanng. Jounal of Machne Leanng Reseach 81: aylo, M. E.; Wheson, S.; and Sone, P ansfe va ne-ask mappngs n polcy seach enfocemen leanng. In Poceedngs of he 6h Inenaonal Jon Confeence on Auonomous Agens and Mulagen Sysems. oey, L.; Shavlk, J.; Walke,.; and Macln, R Relaonal macos fo ansfe n enfocemen leanng. In Blockeel, H.; Ramon, J.; Shavlk, J.; and adepall, P., eds., Inducve Logc Pogammng, volume 4894 of Lecue Noes n Compue Scence. Spnge Beln Hedelbeg Voos, H., and Bou Amma, H Nonlnea ackng and landng conolle fo quadoo aeal obos. In Poceedngs of he IEEE Inenaonal Confeence on Conol Applcaons CCA, Wang, C., and Mahadevan, S Manfold algnmen whou coespondence. In Poceedngs of he 21s Inenaonal Jon Confeence on Afcal Inellgence IJCAI, Mogan Kaufmann.

CHAPTER 10: LINEAR DISCRIMINATION

CHAPTER 10: LINEAR DISCRIMINATION HAPER : LINEAR DISRIMINAION Dscmnan-based lassfcaon 3 In classfcaon h K classes ( k ) We defned dsmnan funcon g () = K hen gven an es eample e chose (pedced) s class label as f g () as he mamum among g

More information

Modern Energy Functional for Nuclei and Nuclear Matter. By: Alberto Hinojosa, Texas A&M University REU Cyclotron 2008 Mentor: Dr.

Modern Energy Functional for Nuclei and Nuclear Matter. By: Alberto Hinojosa, Texas A&M University REU Cyclotron 2008 Mentor: Dr. Moden Enegy Funconal fo Nucle and Nuclea Mae By: lbeo noosa Teas &M Unvesy REU Cycloon 008 Meno: D. Shalom Shlomo Oulne. Inoducon.. The many-body poblem and he aee-fock mehod. 3. Skyme neacon. 4. aee-fock

More information

5-1. We apply Newton s second law (specifically, Eq. 5-2). F = ma = ma sin 20.0 = 1.0 kg 2.00 m/s sin 20.0 = 0.684N. ( ) ( )

5-1. We apply Newton s second law (specifically, Eq. 5-2). F = ma = ma sin 20.0 = 1.0 kg 2.00 m/s sin 20.0 = 0.684N. ( ) ( ) 5-1. We apply Newon s second law (specfcally, Eq. 5-). (a) We fnd he componen of he foce s ( ) ( ) F = ma = ma cos 0.0 = 1.00kg.00m/s cos 0.0 = 1.88N. (b) The y componen of he foce s ( ) ( ) F = ma = ma

More information

Name of the Student:

Name of the Student: Engneeng Mahemacs 05 SUBJEC NAME : Pobably & Random Pocess SUBJEC CODE : MA645 MAERIAL NAME : Fomula Maeal MAERIAL CODE : JM08AM007 REGULAION : R03 UPDAED ON : Febuay 05 (Scan he above QR code fo he dec

More information

1 Constant Real Rate C 1

1 Constant Real Rate C 1 Consan Real Rae. Real Rae of Inees Suppose you ae equally happy wh uns of he consumpon good oday o 5 uns of he consumpon good n peod s me. C 5 Tha means you ll be pepaed o gve up uns oday n eun fo 5 uns

More information

Unsupervised Cross-Domain Transfer in Policy Gradient Reinforcement Learning via Manifold Alignment

Unsupervised Cross-Domain Transfer in Policy Gradient Reinforcement Learning via Manifold Alignment Proceedngs of he Tweny-Nnh AAAI Conference on Arfcal Inellgence Unsupervsed Cross-Doman Transfer n Polcy Graden Renforcemen Learnng va Manfold Algnmen Haham Bou Ammar Unv. of Pennsylvana hahamb@seas.upenn.edu

More information

CptS 570 Machine Learning School of EECS Washington State University. CptS Machine Learning 1

CptS 570 Machine Learning School of EECS Washington State University. CptS Machine Learning 1 ps 57 Machne Leann School of EES Washnon Sae Unves ps 57 - Machne Leann Assume nsances of classes ae lneal sepaable Esmae paamees of lnea dscmnan If ( - -) > hen + Else - ps 57 - Machne Leann lassfcaon

More information

( ) ( )) ' j, k. These restrictions in turn imply a corresponding set of sample moment conditions:

( ) ( )) ' j, k. These restrictions in turn imply a corresponding set of sample moment conditions: esng he Random Walk Hypohess If changes n a sees P ae uncoelaed, hen he followng escons hold: va + va ( cov, 0 k 0 whee P P. k hese escons n un mply a coespondng se of sample momen condons: g µ + µ (,,

More information

I-POLYA PROCESS AND APPLICATIONS Leda D. Minkova

I-POLYA PROCESS AND APPLICATIONS Leda D. Minkova The XIII Inenaonal Confeence Appled Sochasc Models and Daa Analyss (ASMDA-009) Jne 30-Jly 3, 009, Vlns, LITHUANIA ISBN 978-9955-8-463-5 L Sakalaskas, C Skadas and E K Zavadskas (Eds): ASMDA-009 Seleced

More information

to Assess Climate Change Mitigation International Energy Workshop, Paris, June 2013

to Assess Climate Change Mitigation International Energy Workshop, Paris, June 2013 Decomposng he Global TIAM-Maco Maco Model o Assess Clmae Change Mgaon Inenaonal Enegy Wokshop Pas June 2013 Socaes Kypeos (PSI) & An Lehla (VTT) 2 Pesenaon Oulne The global ETSAP-TIAM PE model and he Maco

More information

Chapter 3: Vectors and Two-Dimensional Motion

Chapter 3: Vectors and Two-Dimensional Motion Chape 3: Vecos and Two-Dmensonal Moon Vecos: magnude and decon Negae o a eco: eese s decon Mulplng o ddng a eco b a scala Vecos n he same decon (eaed lke numbes) Geneal Veco Addon: Tangle mehod o addon

More information

Handling Fuzzy Constraints in Flow Shop Problem

Handling Fuzzy Constraints in Flow Shop Problem Handlng Fuzzy Consans n Flow Shop Poblem Xueyan Song and Sanja Peovc School of Compue Scence & IT, Unvesy of Nongham, UK E-mal: {s sp}@cs.no.ac.uk Absac In hs pape, we pesen an appoach o deal wh fuzzy

More information

FIRMS IN THE TWO-PERIOD FRAMEWORK (CONTINUED)

FIRMS IN THE TWO-PERIOD FRAMEWORK (CONTINUED) FIRMS IN THE TWO-ERIO FRAMEWORK (CONTINUE) OCTOBER 26, 2 Model Sucue BASICS Tmelne of evens Sa of economc plannng hozon End of economc plannng hozon Noaon : capal used fo poducon n peod (decded upon n

More information

s = rθ Chapter 10: Rotation 10.1: What is physics?

s = rθ Chapter 10: Rotation 10.1: What is physics? Chape : oaon Angula poson, velocy, acceleaon Consan angula acceleaon Angula and lnea quanes oaonal knec enegy oaonal nea Toque Newon s nd law o oaon Wok and oaonal knec enegy.: Wha s physcs? In pevous

More information

Fast Calibration for Robot Welding System with Laser Vision

Fast Calibration for Robot Welding System with Laser Vision Fas Calbaon fo Robo Weldng Ssem h Lase Vson Lu Su Mechancal & Eleccal Engneeng Nanchang Unves Nanchang, Chna Wang Guoong Mechancal Engneeng Souh Chna Unves of echnolog Guanghou, Chna Absac Camea calbaon

More information

CHAPTER 3 DETECTION TECHNIQUES FOR MIMO SYSTEMS

CHAPTER 3 DETECTION TECHNIQUES FOR MIMO SYSTEMS 4 CAPTER 3 DETECTION TECNIQUES FOR MIMO SYSTEMS 3. INTRODUCTION The man challenge n he paccal ealzaon of MIMO weless sysems les n he effcen mplemenaon of he deeco whch needs o sepaae he spaally mulplexed

More information

ajanuary't I11 F or,'.

ajanuary't I11 F or,'. ',f,". ; q - c. ^. L.+T,..LJ.\ ; - ~,.,.,.,,,E k }."...,'s Y l.+ : '. " = /.. :4.,Y., _.,,. "-.. - '// ' 7< s k," ;< - " fn 07 265.-.-,... - ma/ \/ e 3 p~~f v-acecu ean d a e.eng nee ng sn ~yoo y namcs

More information

Simulation of Non-normal Autocorrelated Variables

Simulation of Non-normal Autocorrelated Variables Jounal of Moden Appled Sascal Mehods Volume 5 Issue Acle 5 --005 Smulaon of Non-nomal Auocoelaed Vaables HT Holgesson Jönöpng Inenaonal Busness School Sweden homasholgesson@bshse Follow hs and addonal

More information

Machine Learning. Spectral Clustering. Lecture 23, April 14, Reading: Eric Xing 1

Machine Learning. Spectral Clustering. Lecture 23, April 14, Reading: Eric Xing 1 Machne Leanng -7/5 7/5-78, 78, Spng 8 Spectal Clusteng Ec Xng Lectue 3, pl 4, 8 Readng: Ec Xng Data Clusteng wo dffeent ctea Compactness, e.g., k-means, mxtue models Connectvty, e.g., spectal clusteng

More information

Real-coded Quantum Evolutionary Algorithm for Global Numerical Optimization with Continuous Variables

Real-coded Quantum Evolutionary Algorithm for Global Numerical Optimization with Continuous Variables Chnese Jounal of Eleconcs Vol.20, No.3, July 2011 Real-coded Quanum Evoluonay Algohm fo Global Numecal Opmzaon wh Connuous Vaables GAO Hu 1 and ZHANG Ru 2 (1.School of Taffc and Tanspoaon, Souhwes Jaoong

More information

Field due to a collection of N discrete point charges: r is in the direction from

Field due to a collection of N discrete point charges: r is in the direction from Physcs 46 Fomula Shee Exam Coulomb s Law qq Felec = k ˆ (Fo example, f F s he elecc foce ha q exes on q, hen ˆ s a un veco n he decon fom q o q.) Elecc Feld elaed o he elecc foce by: Felec = qe (elecc

More information

Reinforcement learning

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

More information

Lecture 5. Plane Wave Reflection and Transmission

Lecture 5. Plane Wave Reflection and Transmission Lecue 5 Plane Wave Reflecon and Tansmsson Incden wave: 1z E ( z) xˆ E (0) e 1 H ( z) yˆ E (0) e 1 Nomal Incdence (Revew) z 1 (,, ) E H S y (,, ) 1 1 1 Refleced wave: 1z E ( z) xˆ E E (0) e S H 1 1z H (

More information

ScienceDirect. Behavior of Integral Curves of the Quasilinear Second Order Differential Equations. Alma Omerspahic *

ScienceDirect. Behavior of Integral Curves of the Quasilinear Second Order Differential Equations. Alma Omerspahic * Avalable onlne a wwwscencedeccom ScenceDec oceda Engneeng 69 4 85 86 4h DAAAM Inenaonal Smposum on Inellgen Manufacung and Auomaon Behavo of Inegal Cuves of he uaslnea Second Ode Dffeenal Equaons Alma

More information

A hybrid method to find cumulative distribution function of completion time of GERT networks

A hybrid method to find cumulative distribution function of completion time of GERT networks Jounal of Indusal Engneeng Inenaonal Sepembe 2005, Vol., No., - 9 Islamc Azad Uvesy, Tehan Souh Banch A hybd mehod o fnd cumulave dsbuon funcon of compleon me of GERT newos S. S. Hashemn * Depamen of Indusal

More information

Journal of Engineering Science and Technology Review 7 (1) (2014) Research Article

Journal of Engineering Science and Technology Review 7 (1) (2014) Research Article Jes Jounal o Engneeng Scence and echnology Revew 7 5 5 Reseach Acle JOURNAL OF Engneeng Scence and echnology Revew www.jes.og Sudy on Pedcve Conol o ajecoy ackng o Roboc Manpulao Yang Zhao Dep. o Eleconc

More information

Modeling Background from Compressed Video

Modeling Background from Compressed Video Modelng acgound fom Compessed Vdeo Weqang Wang Daong Chen Wen Gao Je Yang School of Compue Scence Canege Mellon Unvesy Psbugh 53 USA Insue of Compung echnology &Gaduae School Chnese Academy of Scences

More information

p E p E d ( ) , we have: [ ] [ ] [ ] Using the law of iterated expectations, we have:

p E p E d ( ) , we have: [ ] [ ] [ ] Using the law of iterated expectations, we have: Poblem Se #3 Soluons Couse 4.454 Maco IV TA: Todd Gomley, gomley@m.edu sbued: Novembe 23, 2004 Ths poblem se does no need o be uned n Queson #: Sock Pces, vdends and Bubbles Assume you ae n an economy

More information

Variants of Pegasos. December 11, 2009

Variants of Pegasos. December 11, 2009 Inroducon Varans of Pegasos SooWoong Ryu bshboy@sanford.edu December, 009 Youngsoo Cho yc344@sanford.edu Developng a new SVM algorhm s ongong research opc. Among many exng SVM algorhms, we wll focus on

More information

) from i = 0, instead of i = 1, we have =

) from i = 0, instead of i = 1, we have = Chape 3: Adjusmen Coss n he abou Make I Movaonal Quesons and Execses: Execse 3 (p 6): Illusae he devaon of equaon (35) of he exbook Soluon: The neempoal magnal poduc of labou s epesened by (3) = = E λ

More information

L4:4. motion from the accelerometer. to recover the simple flutter. Later, we will work out how. readings L4:3

L4:4. motion from the accelerometer. to recover the simple flutter. Later, we will work out how. readings L4:3 elave moon L4:1 To appl Newon's laws we need measuemens made fom a 'fed,' neal efeence fame (unacceleaed, non-oang) n man applcaons, measuemens ae made moe smpl fom movng efeence fames We hen need a wa

More information

Chapter Finite Difference Method for Ordinary Differential Equations

Chapter Finite Difference Method for Ordinary Differential Equations Chape 8.7 Fne Dffeence Mehod fo Odnay Dffeenal Eqaons Afe eadng hs chape, yo shold be able o. Undesand wha he fne dffeence mehod s and how o se o solve poblems. Wha s he fne dffeence mehod? The fne dffeence

More information

The Unique Solution of Stochastic Differential Equations. Dietrich Ryter. Midartweg 3 CH-4500 Solothurn Switzerland

The Unique Solution of Stochastic Differential Equations. Dietrich Ryter. Midartweg 3 CH-4500 Solothurn Switzerland The Unque Soluon of Sochasc Dffeenal Equaons Dech Rye RyeDM@gawne.ch Mdaweg 3 CH-4500 Solohun Swzeland Phone +4132 621 13 07 Tme evesal n sysems whou an exenal df sngles ou he an-iô negal. Key wods: Sochasc

More information

When to Treat Prostate Cancer Patients Based on their PSA Dynamics

When to Treat Prostate Cancer Patients Based on their PSA Dynamics When o Tea Posae Cance Paens Based on he PSA Dynamcs CLARA day on opeaons eseach n cance eamen & opeaons managemen Novembe 7 00 Mael S. Lave PhD Man L. Pueman PhD Sco Tyldesley M.D. Wllam J. Mos M.D CIHR

More information

Variability Aware Network Utility Maximization

Variability Aware Network Utility Maximization aably Awae Newok ly Maxmzaon nay Joseph and Gusavo de ecana Depamen of Eleccal and Compue Engneeng, he nvesy of exas a Ausn axv:378v3 [cssy] 3 Ap 0 Absac Newok ly Maxmzaon NM povdes he key concepual famewok

More information

CHAPTER 10: LINEAR DISCRIMINATION

CHAPTER 10: LINEAR DISCRIMINATION CHAPER : LINEAR DISCRIMINAION Dscrmnan-based Classfcaon 3 In classfcaon h K classes (C,C,, C k ) We defned dscrmnan funcon g j (), j=,,,k hen gven an es eample, e chose (predced) s class label as C f g

More information

V.Abramov - FURTHER ANALYSIS OF CONFIDENCE INTERVALS FOR LARGE CLIENT/SERVER COMPUTER NETWORKS

V.Abramov - FURTHER ANALYSIS OF CONFIDENCE INTERVALS FOR LARGE CLIENT/SERVER COMPUTER NETWORKS R&RATA # Vol.) 8, March FURTHER AALYSIS OF COFIDECE ITERVALS FOR LARGE CLIET/SERVER COMPUTER ETWORKS Vyacheslav Abramov School of Mahemacal Scences, Monash Unversy, Buldng 8, Level 4, Clayon Campus, Wellngon

More information

APPROXIMATIONS FOR AND CONVEXITY OF PROBABILISTICALLY CONSTRAINED PROBLEMS WITH RANDOM RIGHT-HAND SIDES

APPROXIMATIONS FOR AND CONVEXITY OF PROBABILISTICALLY CONSTRAINED PROBLEMS WITH RANDOM RIGHT-HAND SIDES R U C O R R E S E A R C H R E P O R APPROXIMAIONS FOR AND CONVEXIY OF PROBABILISICALLY CONSRAINED PROBLEMS WIH RANDOM RIGH-HAND SIDES M.A. Lejeune a A. PREKOPA b RRR 7-005, JUNE 005 RUCOR Ruges Cene fo

More information

Integer Programming Models for Decision Making of. Order Entry Stage in Make to Order Companies 1. INTRODUCTION

Integer Programming Models for Decision Making of. Order Entry Stage in Make to Order Companies 1. INTRODUCTION Inege Pogammng Models fo Decson Makng of Ode Eny Sage n Make o Ode Companes Mahendawah ER, Rully Soelaman and Rzal Safan Depamen of Infomaon Sysems Depamen of Infomacs Engneeng Insu eknolog Sepuluh Nopembe,

More information

Solution of Non-homogeneous bulk arrival Two-node Tandem Queuing Model using Intervention Poisson distribution

Solution of Non-homogeneous bulk arrival Two-node Tandem Queuing Model using Intervention Poisson distribution Volume-03 Issue-09 Sepembe-08 ISSN: 455-3085 (Onlne) RESEARCH REVIEW Inenaonal Jounal of Muldscplnay www.jounals.com [UGC Lsed Jounal] Soluon of Non-homogeneous bulk aval Two-node Tandem Queung Model usng

More information

High-level Hierarchical Semantic Processing Framework for Smart Sensor Networks

High-level Hierarchical Semantic Processing Framework for Smart Sensor Networks HSI 2008 Kakow, Poland, May 25-27, 2008 Hgh-level Heachcal Semanc Pocessng Famewok fo Sma Senso Newoks Dema Buckne, Membe, IEEE, Jamal Kasb Rosemae Velk, Membe, IEEE, and Wolfgang Hezne, Membe, IEEE Venna

More information

Outline. Probabilistic Model Learning. Probabilistic Model Learning. Probabilistic Model for Time-series Data: Hidden Markov Model

Outline. Probabilistic Model Learning. Probabilistic Model Learning. Probabilistic Model for Time-series Data: Hidden Markov Model Probablsc Model for Tme-seres Daa: Hdden Markov Model Hrosh Mamsuka Bonformacs Cener Kyoo Unversy Oulne Three Problems for probablsc models n machne learnng. Compung lkelhood 2. Learnng 3. Parsng (predcon

More information

John Geweke a and Gianni Amisano b a Departments of Economics and Statistics, University of Iowa, USA b European Central Bank, Frankfurt, Germany

John Geweke a and Gianni Amisano b a Departments of Economics and Statistics, University of Iowa, USA b European Central Bank, Frankfurt, Germany Herarchcal Markov Normal Mxure models wh Applcaons o Fnancal Asse Reurns Appendx: Proofs of Theorems and Condonal Poseror Dsrbuons John Geweke a and Gann Amsano b a Deparmens of Economcs and Sascs, Unversy

More information

I-Hsuan Hong Hsi-Mei Hsu Yi-Mu Wu Chun-Shao Yeh

I-Hsuan Hong Hsi-Mei Hsu Yi-Mu Wu Chun-Shao Yeh Poceedngs of he 8 Wne Smulaon Confeence S. J. ason, R. R. Hll, L. önch, O. Rose, T. Jeffeson, J. W. Fowle eds. PRICING DECISION AND LEAD TIE SETTING IN A DUOPOL SEICONDUCTOR INDUSTR I-Hsuan Hong Hs-e Hsu

More information

SCIENCE CHINA Technological Sciences

SCIENCE CHINA Technological Sciences SIENE HINA Technologcal Scences Acle Apl 4 Vol.57 No.4: 84 8 do:.7/s43-3-5448- The andom walkng mehod fo he seady lnea convecondffuson equaon wh axsymmec dsc bounday HEN Ka, SONG MengXuan & ZHANG Xng *

More information

Delay-Dependent Control for Time-Delayed T-S Fuzzy Systems Using Descriptor Representation

Delay-Dependent Control for Time-Delayed T-S Fuzzy Systems Using Descriptor Representation 82 Inenaonal Jounal of Conol Auomaon and Sysems Vol 2 No 2 June 2004 Delay-Dependen Conol fo me-delayed -S Fuzzy Sysems Usng Descpo Repesenaon Eun ae Jeung Do Chang Oh and Hong Bae ak Absac: hs pape pesens

More information

A VISCOPLASTIC MODEL OF ASYMMETRICAL COLD ROLLING

A VISCOPLASTIC MODEL OF ASYMMETRICAL COLD ROLLING SISOM 4, BUCHAEST, - May A VISCOPLASTIC MODEL OF ASYMMETICAL COLD OLLING odca IOAN Spu Hae Unvesy Buchaes, odcaoan7@homal.com Absac: In hs pape s gven a soluon of asymmecal sp ollng poblem usng a Bngham

More information

An axisymmetric incompressible lattice BGK model for simulation of the pulsatile ow in a circular pipe

An axisymmetric incompressible lattice BGK model for simulation of the pulsatile ow in a circular pipe INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN FLUIDS In. J. Nume. Meh. Fluds 005; 49:99 116 Publshed onlne 3 June 005 n Wley IneScence www.nescence.wley.com). DOI: 10.100/d.997 An axsymmec ncompessble

More information

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

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

More information

N 1. Time points are determined by the

N 1. Time points are determined by the upplemena Mehods Geneaon of scan sgnals In hs secon we descbe n deal how scan sgnals fo 3D scannng wee geneaed. can geneaon was done n hee seps: Fs, he dve sgnal fo he peo-focusng elemen was geneaed o

More information

Suppose we have observed values t 1, t 2, t n of a random variable T.

Suppose we have observed values t 1, t 2, t n of a random variable T. Sppose we have obseved vales, 2, of a adom vaable T. The dsbo of T s ow o belog o a cea ype (e.g., expoeal, omal, ec.) b he veco θ ( θ, θ2, θp ) of ow paamees assocaed wh s ow (whee p s he mbe of ow paamees).

More information

Optimal Sensor Placement for Cooperative Distributed Vision

Optimal Sensor Placement for Cooperative Distributed Vision Opmal Senso Placemen fo Coopeave Dsbued Vson Lus E. Navao-Semen, John M. Dolan, and Padeep K. Khosla, Depamen of Eleccal and Compue Engneeng he Robocs Insue Canege Mellon Unves Psbugh, PA, USA {lenscmu,

More information

Lecture VI Regression

Lecture VI Regression Lecure VI Regresson (Lnear Mehods for Regresson) Conens: Lnear Mehods for Regresson Leas Squares, Gauss Markov heorem Recursve Leas Squares Lecure VI: MLSC - Dr. Sehu Vjayakumar Lnear Regresson Model M

More information

Molecular dynamics modeling of thermal and mechanical properties

Molecular dynamics modeling of thermal and mechanical properties Molecula dynamcs modelng of hemal and mechancal popees Alejando Sachan School of Maeals Engneeng Pudue Unvesy sachan@pudue.edu Maeals a molecula scales Molecula maeals Ceamcs Meals Maeals popees chas Maeals

More information

CS434a/541a: Pattern Recognition Prof. Olga Veksler. Lecture 4

CS434a/541a: Pattern Recognition Prof. Olga Veksler. Lecture 4 CS434a/54a: Paern Recognon Prof. Olga Veksler Lecure 4 Oulne Normal Random Varable Properes Dscrmnan funcons Why Normal Random Varables? Analycally racable Works well when observaon comes form a corruped

More information

A Methodology for Detecting the Change of Customer Behavior based on Association Rule Mining

A Methodology for Detecting the Change of Customer Behavior based on Association Rule Mining A Mehodology fo Deecng he Change of Cusome Behavo based on Assocaon Rule Mnng Hee Seo Song, Soung He Km KAIST Gaduae School of Managemen Jae Kyeong Km KyungHee Unvesy Absac Undesandng and adapng o changes

More information

Lecture 6: Learning for Control (Generalised Linear Regression)

Lecture 6: Learning for Control (Generalised Linear Regression) Lecure 6: Learnng for Conrol (Generalsed Lnear Regresson) Conens: Lnear Mehods for Regresson Leas Squares, Gauss Markov heorem Recursve Leas Squares Lecure 6: RLSC - Prof. Sehu Vjayakumar Lnear Regresson

More information

University of California, Davis Date: June xx, PRELIMINARY EXAMINATION FOR THE Ph.D. DEGREE ANSWER KEY

University of California, Davis Date: June xx, PRELIMINARY EXAMINATION FOR THE Ph.D. DEGREE ANSWER KEY Unvesy of Calfona, Davs Dae: June xx, 009 Depamen of Economcs Tme: 5 hous Mcoeconomcs Readng Tme: 0 mnues PRELIMIARY EXAMIATIO FOR THE Ph.D. DEGREE Pa I ASWER KEY Ia) Thee ae goods. Good s lesue, measued

More information

Adaptive complex modified hybrid function projective synchronization of different dimensional complex chaos with uncertain complex parameters

Adaptive complex modified hybrid function projective synchronization of different dimensional complex chaos with uncertain complex parameters Nonlnea Dyn (26) 83:9 2 DOI.7/s7--239-8 ORIGINAL PAPER Adapve complex modfed hybd funcon pojecve synchonzaon of dffeen dmensonal complex chaos wh uncean complex paamees Jan Lu Shuang Lu Julen Clnon Spo

More information

Nanoparticles. Educts. Nucleus formation. Nucleus. Growth. Primary particle. Agglomeration Deagglomeration. Agglomerate

Nanoparticles. Educts. Nucleus formation. Nucleus. Growth. Primary particle. Agglomeration Deagglomeration. Agglomerate ucs Nucleus Nucleus omaon cal supesauaon Mng o eucs, empeaue, ec. Pmay pacle Gowh Inegaon o uson-lme pacle gowh Nanopacles Agglomeaon eagglomeaon Agglomeae Sablsaon o he nanopacles agans agglomeaon! anspo

More information

A multiple-relaxation-time lattice Boltzmann model for simulating. incompressible axisymmetric thermal flows in porous media

A multiple-relaxation-time lattice Boltzmann model for simulating. incompressible axisymmetric thermal flows in porous media A mulple-elaxaon-me lace Bolmann model fo smulang ncompessble axsymmec hemal flows n poous meda Qng Lu a, Ya-Lng He a, Qng L b a Key Laboaoy of Themo-Flud Scence and Engneeng of Mnsy of Educaon, School

More information

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

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

More information

Density Matrix Description of NMR BCMB/CHEM 8190

Density Matrix Description of NMR BCMB/CHEM 8190 Densy Marx Descrpon of NMR BCMBCHEM 89 Operaors n Marx Noaon Alernae approach o second order specra: ask abou x magnezaon nsead of energes and ranson probables. If we say wh one bass se, properes vary

More information

Solution in semi infinite diffusion couples (error function analysis)

Solution in semi infinite diffusion couples (error function analysis) Soluon n sem nfne dffuson couples (error funcon analyss) Le us consder now he sem nfne dffuson couple of wo blocks wh concenraon of and I means ha, n a A- bnary sysem, s bondng beween wo blocks made of

More information

Efficient Bayesian Network Learning for System Optimization in Reliability Engineering

Efficient Bayesian Network Learning for System Optimization in Reliability Engineering Qualy Technology & Quanave Managemen Vol. 9, No., pp. 97-, 202 QTQM ICAQM 202 Effcen Bayesan Newok Leanng fo Sysem Opmzaon n Relably Engneeng A. Gube and I. Ben-Gal Depamen of Indusal Engneeng, Faculy

More information

Computer Propagation Analysis Tools

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

More information

Cubic Bezier Homotopy Function for Solving Exponential Equations

Cubic Bezier Homotopy Function for Solving Exponential Equations Penerb Journal of Advanced Research n Compung and Applcaons ISSN (onlne: 46-97 Vol. 4, No.. Pages -8, 6 omoopy Funcon for Solvng Eponenal Equaons S. S. Raml *,,. Mohamad Nor,a, N. S. Saharzan,b and M.

More information

Basic molecular dynamics

Basic molecular dynamics 1.1, 3.1, 1.333,. Inoducon o Modelng and Smulaon Spng 11 Pa I Connuum and pacle mehods Basc molecula dynamcs Lecue Makus J. Buehle Laboaoy fo Aomsc and Molecula Mechancs Depamen of Cvl and Envonmenal Engneeng

More information

New Stability Condition of T-S Fuzzy Systems and Design of Robust Flight Control Principle

New Stability Condition of T-S Fuzzy Systems and Design of Robust Flight Control Principle 96 JOURNAL O ELECRONIC SCIENCE AND ECHNOLOGY, VOL., NO., MARCH 3 New Sably Conon of -S uzzy Sysems an Desgn of Robus lgh Conol Pncple Chun-Nng Yang, Ya-Zhou Yue, an Hu L Absac Unlke he pevous eseach woks

More information

Robust and Accurate Cancer Classification with Gene Expression Profiling

Robust and Accurate Cancer Classification with Gene Expression Profiling Robus and Accurae Cancer Classfcaon wh Gene Expresson Proflng (Compuaonal ysems Bology, 2005) Auhor: Hafeng L, Keshu Zhang, ao Jang Oulne Background LDA (lnear dscrmnan analyss) and small sample sze problem

More information

An introduction to Support Vector Machine

An introduction to Support Vector Machine An nroducon o Suppor Vecor Machne 報告者 : 黃立德 References: Smon Haykn, "Neural Neworks: a comprehensve foundaon, second edon, 999, Chaper 2,6 Nello Chrsann, John Shawe-Tayer, An Inroducon o Suppor Vecor Machnes,

More information

Physics 201 Lecture 15

Physics 201 Lecture 15 Phscs 0 Lecue 5 l Goals Lecue 5 v Elo consevaon of oenu n D & D v Inouce oenu an Iulse Coens on oenu Consevaon l oe geneal han consevaon of echancal eneg l oenu Consevaon occus n sses wh no ne eenal foces

More information

STABILITY CRITERIA FOR A CLASS OF NEUTRAL SYSTEMS VIA THE LMI APPROACH

STABILITY CRITERIA FOR A CLASS OF NEUTRAL SYSTEMS VIA THE LMI APPROACH Asan Jounal of Conol, Vol. 6, No., pp. 3-9, Mach 00 3 Bef Pape SABILIY CRIERIA FOR A CLASS OF NEURAL SYSEMS VIA HE LMI APPROACH Chang-Hua Len and Jen-De Chen ABSRAC In hs pape, he asypoc sably fo a class

More information

Hierarchical Production Planning in Make to Order System Based on Work Load Control Method

Hierarchical Production Planning in Make to Order System Based on Work Load Control Method Unvesal Jounal of Indusal and Busness Managemen 3(): -20, 205 DOI: 0.389/ujbm.205.0300 hp://www.hpub.og Heachcal Poducon Plannng n Make o Ode Sysem Based on Wok Load Conol Mehod Ehsan Faah,*, Maha Khodadad

More information

Backcalculation Analysis of Pavement-layer Moduli Using Pattern Search Algorithms

Backcalculation Analysis of Pavement-layer Moduli Using Pattern Search Algorithms Bakallaon Analyss of Pavemen-laye Modl Usng Paen Seah Algohms Poje Repo fo ENCE 74 Feqan Lo May 7 005 Bakallaon Analyss of Pavemen-laye Modl Usng Paen Seah Algohms. Inodon. Ovevew of he Poje 3. Objeve

More information

Maximum Likelihood Estimation

Maximum Likelihood Estimation Mau Lkelhood aon Beln Chen Depaen of Copue Scence & Infoaon ngneeng aonal Tawan oal Unvey Refeence:. he Alpaydn, Inoducon o Machne Leanng, Chape 4, MIT Pe, 4 Saple Sac and Populaon Paaee A Scheac Depcon

More information

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

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

More information

MCTDH Approach to Strong Field Dynamics

MCTDH Approach to Strong Field Dynamics MCTDH ppoach o Song Feld Dynamcs Suen Sukasyan Thomas Babec and Msha Ivanov Unvesy o Oawa Canada Impeal College ondon UK KITP Sana Babaa. May 8 009 Movaon Song eld dynamcs Role o elecon coelaon Tunnel

More information

Graduate Macroeconomics 2 Problem set 5. - Solutions

Graduate Macroeconomics 2 Problem set 5. - Solutions Graduae Macroeconomcs 2 Problem se. - Soluons Queson 1 To answer hs queson we need he frms frs order condons and he equaon ha deermnes he number of frms n equlbrum. The frms frs order condons are: F K

More information

Tecnologia e Inovação, Lisboa, Portugal. ABB Corporate Research Center, Wallstadter Str. 59, Ladenburg, Germany,

Tecnologia e Inovação, Lisboa, Portugal. ABB Corporate Research Center, Wallstadter Str. 59, Ladenburg, Germany, A New Connuous-Tme Schedulng Fomulaon fo Connuous Plans unde Vaable Eleccy Cos Pedo M. Caso * Io Hajunkosk and Ignaco E. Gossmann Depaameno de Modelação e Smulação de Pocessos Insuo Naconal de Engenhaa

More information

Multistage Median Ranked Set Sampling for Estimating the Population Median

Multistage Median Ranked Set Sampling for Estimating the Population Median Jounal of Mathematcs and Statstcs 3 (: 58-64 007 ISSN 549-3644 007 Scence Publcatons Multstage Medan Ranked Set Samplng fo Estmatng the Populaton Medan Abdul Azz Jeman Ame Al-Oma and Kamaulzaman Ibahm

More information

Mechanics Physics 151

Mechanics Physics 151 Mechancs Physcs 5 Lecure 0 Canoncal Transformaons (Chaper 9) Wha We Dd Las Tme Hamlon s Prncple n he Hamlonan formalsm Dervaon was smple δi δ Addonal end-pon consrans pq H( q, p, ) d 0 δ q ( ) δq ( ) δ

More information

Reflection and Refraction

Reflection and Refraction Chape 1 Reflecon and Refacon We ae now neesed n eplong wha happens when a plane wave avelng n one medum encounes an neface (bounday) wh anohe medum. Undesandng hs phenomenon allows us o undesand hngs lke:

More information

Clustering (Bishop ch 9)

Clustering (Bishop ch 9) Cluserng (Bshop ch 9) Reference: Daa Mnng by Margare Dunham (a slde source) 1 Cluserng Cluserng s unsupervsed learnng, here are no class labels Wan o fnd groups of smlar nsances Ofen use a dsance measure

More information

Dynamic Team Decision Theory. EECS 558 Project Shrutivandana Sharma and David Shuman December 10, 2005

Dynamic Team Decision Theory. EECS 558 Project Shrutivandana Sharma and David Shuman December 10, 2005 Dynamc Team Decson Theory EECS 558 Proec Shruvandana Sharma and Davd Shuman December 0, 005 Oulne Inroducon o Team Decson Theory Decomposon of he Dynamc Team Decson Problem Equvalence of Sac and Dynamc

More information

Accelerated Sequen.al Probability Ra.o Test (SPRT) for Ongoing Reliability Tes.ng (ORT)

Accelerated Sequen.al Probability Ra.o Test (SPRT) for Ongoing Reliability Tes.ng (ORT) cceleaed Sequen.al Pobably Ra.o Tes (SPRT) fo Ongong Relably Tes.ng (ORT) Mlena Kasch Rayheon, IDS Copygh 25 Rayheon Company. ll ghs eseved. Cusome Success Is Ou Msson s a egseed adema of Rayheon Company

More information

calculating electromagnetic

calculating electromagnetic Theoeal mehods fo alulang eleomagne felds fom lghnng dshage ajeev Thoapplll oyal Insue of Tehnology KTH Sweden ajeev.thoapplll@ee.kh.se Oulne Despon of he poblem Thee dffeen mehods fo feld alulaons - Dpole

More information

Density Matrix Description of NMR BCMB/CHEM 8190

Density Matrix Description of NMR BCMB/CHEM 8190 Densy Marx Descrpon of NMR BCMBCHEM 89 Operaors n Marx Noaon If we say wh one bass se, properes vary only because of changes n he coeffcens weghng each bass se funcon x = h< Ix > - hs s how we calculae

More information

Introduction ( Week 1-2) Course introduction A brief introduction to molecular biology A brief introduction to sequence comparison Part I: Algorithms

Introduction ( Week 1-2) Course introduction A brief introduction to molecular biology A brief introduction to sequence comparison Part I: Algorithms Course organzaon Inroducon Wee -2) Course nroducon A bref nroducon o molecular bology A bref nroducon o sequence comparson Par I: Algorhms for Sequence Analyss Wee 3-8) Chaper -3, Models and heores» Probably

More information

( ) () we define the interaction representation by the unitary transformation () = ()

( ) () we define the interaction representation by the unitary transformation () = () Hgher Order Perurbaon Theory Mchael Fowler 3/7/6 The neracon Represenaon Recall ha n he frs par of hs course sequence, we dscussed he chrödnger and Hesenberg represenaons of quanum mechancs here n he chrödnger

More information

Digital Wiener s Filtering in Seismic Data Processing in Trans-Ramos Prospect of Rivers State

Digital Wiener s Filtering in Seismic Data Processing in Trans-Ramos Prospect of Rivers State Jounal of Emegng Tends n Engneeng and Appled Scences (JETEAS) (1): 43-49 Scholaln eseach Insue Jounals, 11 (ISSN: 141-716) jeeas.scholalneseach.og Jounal of Emegng Tends n Engneeng and Appled Scences (JETEAS)

More information

SUBDIFFUSION SUPPORTS JOINING OF CORRECT ENDS DURING REPAIR OF

SUBDIFFUSION SUPPORTS JOINING OF CORRECT ENDS DURING REPAIR OF SUBDIFFUSIO SUPPORTS JOIIG OF CORRECT EDS DURIG REPAIR OF DA DOUBLE-STRAD BREAKS S. Gs *, V. Hable, G.A. Dexle, C. Geubel, C. Sebenwh,. Haum, A.A. Fedl, G. Dollnge Angewande Physk und essechnk LRT, Unvesä

More information

Linear Response Theory: The connection between QFT and experiments

Linear Response Theory: The connection between QFT and experiments Phys540.nb 39 3 Lnear Response Theory: The connecon beween QFT and expermens 3.1. Basc conceps and deas Q: ow do we measure he conducvy of a meal? A: we frs nroduce a weak elecrc feld E, and hen measure

More information

Monetary policy and models

Monetary policy and models Moneay polcy and odels Kes Næss and Kes Haae Moka Noges Bank Moneay Polcy Unvesy of Copenhagen, 8 May 8 Consue pces and oney supply Annual pecenage gowh. -yea ovng aveage Gowh n oney supply Inflaon - 9

More information

A Novel Fast Otsu Digital Image Segmentation Method

A Novel Fast Otsu Digital Image Segmentation Method The Inenaonal Aab Jounal of Infomaon Technology, Vol. 3, No. 4, July 06 47 A Novel Fas Osu Dgal Image Segmenaon Mehod Duaa AlSaeed,, Ahmed oudane,, and Al El-Zaa 3 Depamen of Compue Scence and Dgal Technologes,

More information

Appendix H: Rarefaction and extrapolation of Hill numbers for incidence data

Appendix H: Rarefaction and extrapolation of Hill numbers for incidence data Anne Chao Ncholas J Goell C seh lzabeh L ander K Ma Rober K Colwell and Aaron M llson 03 Rarefacon and erapolaon wh ll numbers: a framewor for samplng and esmaon n speces dversy sudes cology Monographs

More information

8. HAMILTONIAN MECHANICS

8. HAMILTONIAN MECHANICS 8. HAMILTONIAN MECHANICS In ode o poceed fom he classcal fomulaon of Maxwell's elecodynamcs o he quanum mechancal descpon a new mahemacal language wll be needed. In he pevous secons he elecomagnec feld

More information

Design of Optimal PID, Fuzzy and New Fuzzy-PID Controller for CANSAT Carrier System Thrust Vector

Design of Optimal PID, Fuzzy and New Fuzzy-PID Controller for CANSAT Carrier System Thrust Vector In J Advanced esgn and Manufacung Technology, Vol 8/ No 2/ June - 2015 1 esgn of Opmal PI, Fuzzy and New Fuzzy-PI Conolle fo CANSAT Cae Sysem Thus Veco A Kosa * epamen of New Scences and Technologes, Unvesy

More information

Existence and Uniqueness Results for Random Impulsive Integro-Differential Equation

Existence and Uniqueness Results for Random Impulsive Integro-Differential Equation Global Journal of Pure and Appled Mahemacs. ISSN 973-768 Volume 4, Number 6 (8), pp. 89-87 Research Inda Publcaons hp://www.rpublcaon.com Exsence and Unqueness Resuls for Random Impulsve Inegro-Dfferenal

More information

Volatility Interpolation

Volatility Interpolation Volaly Inerpolaon Prelmnary Verson March 00 Jesper Andreasen and Bran Huge Danse Mares, Copenhagen wan.daddy@danseban.com brno@danseban.com Elecronc copy avalable a: hp://ssrn.com/absrac=69497 Inro Local

More information