Dynamic Pricing for Smart Grid with Reinforcement Learning

Size: px
Start display at page:

Download "Dynamic Pricing for Smart Grid with Reinforcement Learning"

Transcription

1 Dynamc Prcng for Smart Grd wth Renforcement Learnng Byung-Gook Km, Yu Zhang, Mhaela van der Schaar, and Jang-Won Lee Samsung Electroncs, Suwon, Korea Department of Electrcal Engneerng, UCLA, Los Angeles, USA Department of Electrcal and Electronc Engneerng, Yonse Unversty, Seoul, Korea Abstract In the smart grd system, dynamc prcng can be an effcent tool for the servce provder whch enables effcent and automated management of the grd. However, n practce, the lack of nformaton about the customers tme-varyng load demand and energy consumpton patterns and the volatlty of electrcty prce n the wholesale market make the mplementaton of dynamc prcng hghly challengng. In ths paper, we study a dynamc prcng problem n the smart grd system where the servce provder decdes the electrcty prce n the retal market. In order to overcome the challenges n mplementng dynamc prcng, we develop a renforcement learnng algorthm. To resolve the drawbacks of the conventonal renforcement learnng algorthm such as hgh computatonal complexty and low convergence speed, we propose an approxmate state defnton and adopt vrtual experence. Numercal results show that the proposed renforcement learnng algorthm can effectvely work wthout a pror nformaton of the system dynamcs. I. INTRODUCTION In the smart grd system, thanks to the real-tme nformaton exchange through communcaton networks, customers can schedule the operaton of ther applances accordng to the change of electrcty prce va the automated energy management system equpped n households, whch we refer to as demand response [1]. From the customers perspectve, prevous works on the load schedulng focus on drectly controllng the energy consumpton of the resdental applances. For example, n our prevous work [2], we proposed two dfferent load schedulng algorthms for a collaboratve and a non-collaboratve smart grd system by takng nto account the customers bdrectonal energy tradng capablty va electrc vehcles. Most of these works am at maxmzng the socal welfare of the smart grd system assumng that the prcng polces are predetermned n the electrcty market by the servce provder. Consequently, the servce provder s regarded as a passve and ndfferent entty whose role n the smart grd system s sgnfcantly lmted. On the contrary, from the servce provder s perspectve, dynamc prcng s an attractve tool that enables effcent grd operaton n terms of both effcent energy consumpton and automated management. Our paper s related to ths second strand of lterature. Specfcally, we focus on a scenaro where the servce provder can adaptvely decde the retal electrcty prce based on the customers load demand level and the wholesale prce such that t mnmzes ether the customers dsutlty (n the case of a benevolent servce provder) or ts own cost (n the case of a proft-makng servce provder). Although dynamc prcng does not drectly control each Ths work was supported n part by Md-career Researcher Program through NRF grant funded by the MSIP, Korea (213R1A2A2A693), and by USA Natonal Scence Foundaton (NSF) grant CCF customer s load schedulng, the approprate prcng can gve consderable benefts to the smart grd system by encouragng the customers to consume energy n a more effcent way. Recently, there have been several works on dynamc prcng for smart grd [3][4][][6][7][8]. In [3] and [4], dynamc prcng problems were studed amng at maxmzng the socal welfare. Consderng a smart grd system wth multple resdences and a sngle servce provder, optmal dynamc prcng schemes were proposed based on the dual decomposton approaches. The authors n [] focused on the smart grd system wth non-cooperatve customers where the conventonal optmzaton approach cannot be appled to as n [3]. To overcome the lack of cooperaton of customers, a smulated annealng-based dynamc prcng algorthm was developed. In a smlar context, the authors n [6] modeled a dynamc prcng problem as a Stackelberg game where the servce provder decdes the retal prce and each selfsh customer decdes the schedule for ts applances accordng to the prce. In [7], the authors developed an ncentve-based dynamc prcng scheme whch allows the servce provder to decde the ncentve for the customers who shft ther applances usage from peak hours to off-peak hours. In [8], the authors ntroduced a twotmescale dynamc prcng scheme to ncorporate both the customers wth day-ahead schedulng and the customers wth real-tme schedulng. To take nto account the uncertantes of energy supply and demand, the authors formulated a Markov decson process (MDP) problem and developed an onlne algorthm. Despte those prevous efforts, there stll exst several crtcal challenges n mplementng dynamc prcng for demand response. Frst, n the practcal smart grd system, t s not easy for the servce provder to obtan the customer-sde nformaton such as ther current load demand levels and the transton probablty of the demand levels, and the customer-specfc utlty models ncludng the wllngness to purchase electrc energy gven ther load demand level and retal prce. Second, even f the servce provder can obtan those nformaton, t wll surfer from varous system dynamcs and uncertantes. The servce provder whch les between the utlty company and the customers may not obtan the perfect nformaton of those system dynamcs a pror. Fnally, the servce provder s requred to have the ablty to estmate the mpact of ts current prcng decson on the customers future behavor. In fact, the current prce nfluences not only the customers current energy consumpton but also ther energy consumpton for the next several hours or the next day. However, t s not easy for the servce provder to calculate the optmal prce consderng the future nfluence of the current prce wthout the detal customer-sde nformaton. Thus, most of exstng /14/$ IEEE 646

2 2 works on dynamc prcng for smart grd have been studed n myopc approaches where the algorthms for dynamc prcng and demand sde load schedulng are conducted wthn a gven tme perod wthout consderng the long-term performance of the smart grd system. In order to overcome the aforementoned challenges of dynamc prcng, n ths paper, we use renforcement learnng to allow the servce provder to learn the behavors of customers and the change of wholesale prce to make an optmal prcng decson. We consder varous stochastc dynamcs of the smart grd system ncludng the customers dynamc demand generaton and energy consumptons, and wholesale prce changes. Based on the consdered system model, we formulate an MDP problem where the servce provder decdes the retal electrcty prce based on the observed system state transton to mnmze ts expected total cost or the customers dsutlty. Contrary to the prevous works [3]-[8] wth smplfed models, n ths paper, we consder a more realstc system model to ncorporate the customers demand generaton and load demand change as well as the wholesale market dynamc where the wholesale electrcty prce can be changed by the utlty company at each tme-slot. To solve the MDP problem wthout a pror nformaton about the change of the customers load demand level, we adopt the Q-learnng algorthm, and to resolve the exstng drawbacks of the conventonal Q-learnng algorthm, we propose the followng two mprovements. Frst, to reduce the complexty of the Q-learnng algorthm whch manly comes from the large number of customers, we propose an alternatve state defnton based on the observed total energy consumpton. Second, to mprove the learnng speed, we adopt vrtual experence n Q-learnng updates. The rest of ths paper s organzed as follows. In Secton II, the system model s presented. In Secton III, we defne the dynamc prcng problem and develop the renforcement learnng-based dynamc prcng algorthm. We provde numercal results n Secton IV and fnally conclude n Secton V. II. SYSTEM MODEL We consder a smart grd system whch conssts of one servce provder and a set of customers I as n Fg. 1. The smart grd system operates n a tme-slotted fashon, where each tme-slot has an equal duraton. At each tme-slot t, the servce provder buys electrc energy from the utlty company through a wholesale electrcty market and provdes t to the customers through a retal electrcty market. In the retal electrcty market, at each tme-slot t, the servce provder determnes the retal prcng functon a t : R + R + and charges each customer an electrcty bll a t (e t ), where et denotes customer s energy consumpton at tme-slot t. We defne the set of retal prcng functons as A and assume that the number of retal prcng functons, A, s fnte. At each tme-slot, each customer generates ts electrcty load demand and decdes the amount of energy consumpton based on ts current load demand level and the retal prcng functon. We assume that the customers average demand generaton rate and the wholesale prcng functon at a tmeslot can vary dependng on ts actual tme n a day. Moreover, ths tme-dependency of customers energy consumpton may nfluence the utlty company s decson on the wholesale prcng functon. To model ths tme-dependency of the demand generaton rate and the wholesale prcng functon, we ntroduce a set of perods H = {,1,,H 1} each of t c ( ) Utlty Company Wholesale prcng functon t t!eˆ ( d ) Servce Provder Flow of nformaton Retal prcng functon t a ( ) Flow of electrc energy Fg. 1. Smart grd system. eˆ ( d ) t t t t eˆ! ( d ) Customers whch represents an actual tme n a day. We map each tmeslot t to one perod h H denotng the perod at tme-slot t by h t. We assume that the sequence of perods h t,t =,1,2, s predetermned and repeated every day. For example, f one day conssts of H = 24 perods (.e., 24 hours), each tmeslot t s mapped to one perod n H = {,1,,23} and the mappng between tme-slots and perods can be represented as h t = mod (t,h), t. (1) A. Model of Customer s Response In each tme-slot, each customer has an accumulated load demand 1, whch s defned as the total amount of energy that t wants to consume for ts applances n that tme-slot. We denote the amount of the accumulated load demand of customer at tme-slot t by d t D, where D s the set of customer accumulated load demand levels. Once customer consumes energy e t at tme-slot t, t mples that the correspondng amount of customer s load demand s satsfed and the rest of the accumulated load demand d t et s not satsfed and we call t the remanng load demand. When the remanng load demand of a customer s greater than, t causes some degree of dssatsfacton to the customer at that tme-slot. To capture ths dssatsfacton from the remanng load demand, we ntroduce a dsutlty functon for u : R + R + for each customer. We assume that u ( ) s an ncreasng functon of the remanng load demandd t et. Based on the dsutlty u (d t et ) and the electrcty bll a t (e t ) that customer has to pay to the servce provder, we defne customer s cost at each tme-slot t as φ t (dt,et ) = u (d t et )+at (e t ). (2) We assume that each customer tres to mnmze ts cost at each tme-slot by decdng the amount of ts energy consumpton and let ê t (dt ) denote customer s energy consumpton decson that mnmzes ts cost 2,.e., ê t (dt ) = argmn φ t e t mn(emax,d t ) (dt,et ), (3) 1 For the convenence, accumulated load demand and load demand are used nterchangeably n the rest of ths paper. 2 Customer s energy consumpton decson s also a functon of the retal prcng functon a t. However, for the smple expresson, we represent t as ê t (dt ). 647

3 3 where e max s the maxmum amount of energy that customer can consume at each tme-slot due to physcal lmtatons of the grd. We assume that a porton of each customer s remanng load demand at a tme-slot s carred forward to the next tme-slot. We call the correspondng load demand the demand backlog and represent t as λ (d t êt (dt )), where λ 1 s the backlog rate of load demand that determnes the amount of demand backlog of customer. At each tme-slot t, each customer randomly generates ts new load demand, D t(ht ), and ts dstrbuton s assumed to be dependent on the current perodh t. At the begnnng of each tme-slott+1, accordng to the demand backlog at the prevous tme-slott,λ (d t êt (dt )), and the newly generated load demand, D t+1 (h t+1 ), customer s accumulated load demand d t+1 s updated as d t+1 = λ (d t êt (dt ))+Dt+1 (h t+1 ). (4) Here, f λ =, no remanng demand at a tme-slot s carred forward generatng no demand backlog, whereas f λ = 1, all the remanng demand at a tme-slot s carred forward to the next tme-slot. It s worth notng that the transton probablty of the load demand from d t to dt+1 depends only on the accumulated load demand d t, the perod ht, and the retal prcng functon a t at tme-slot t, and we represent t as p d (d t+1 d t,ht,a t ). B. Wholesale Electrcty Market At each tme-slot t, the servce provder buys electrc energy, whch corresponds to the total amount of energy consumpton of customers, I êt (dt ), from the utlty company n the wholesale electrcty market as llustrated n Fg. 1. The utlty company charges the servce provder an wholesale electrcty cost based on a wholesale prcng functon c t : R + R +, where c t s a functon of the total amount of energy consumpton I êt (dt ). We assume that ct s selected among a fnte number of wholesale prcng functons n set C and ts transton probablty from c t to c t+1 depends on the current wholesale prcng functon, c t, and the current perod, h t, and thus t can be represented as p c (c t+1 c t,h t ). We defne the servce provder s cost functon at each tmeslot t as a functon of the customers load demand vector d t = [d t ] I, the wholesale electrcty prcng functon c t, and the retal prcng functon a t,.e., ψ t ( d t,c t,a t ) = c t( ) ê t (d t ) a t (ê t (d t )), () I I where the frst term denotes the total wholesale electrcty cost that the servce provder has to pay to the utlty company and the second term denotes the servce provder s revenue from sellng energy to the customers. III. REINFORCEMENT LEARNING ALGORITHM In ths secton, based on the smart grd system ntroduced n the prevous secton, we frst formulate a dynamc prcng problem n the framework of MDP. Then, by usng renforcement learnng, we develop an effcent and fast dynamc prcng algorthm whch does not requre the nformaton about the system dynamcs and uncertantes. A. Problem Formulaton We formulate the dynamc prcng problem n the smart grd system as an MDP problem, whch s defned by a set of decson maker s actons, a set of system states and ther transton probabltes, and a system cost functon for the decson maker. In our MDP problem, the decson maker s the servce provder whose acton s choosng a retal prcng functon a t A at each tme-slot t. We defne the state of our smart grd system at tme-slot t as the combnaton of the accumulated load demands vector, dt, the current perod h t, and the wholesale prcng functon, c t,.e., s t = ( d t,h t,c t ) S, (6) where S = I D H C. Snce the transton of each customer s load demand (from d t to dt+1 ), that of the perod (from h t to h t+1 ), and that of the wholesale prcng functon (from c t to c t+1 ) depend only on the state s t and acton a t at tme-slot t, the sequence of states {s t,t =,1,2, } follows a Markov decson process wth acton a t. The transton probablty from state s t = ( d t,h t,c t ) to state s t+1 = ( d t+1,h t+1,c t+1 ) wth gven acton a t can be represented as p s (s t+1 s t,a t ) =p h (h t+1 h t )p c (c t+1 c t,h t ) (7) I p d (d t+1 d t,h t,a t ), where p h (h t+1 h t ) denotes the transton probablty of the perod from h t to h t+1. We defne the system cost for the servce provder at each tme-slot t as the weghted sum of the servce provder s cost and the customers cost at the tme-slot: r t (s t,a t ) = (1 ρ)ψ t ( d t,c t,a t )+ρ I φ t (d t,ê t ), (8) where ρ denotes the weghtng factor that determnes the relatve mportance between the servce provder s cost and the customers cost. We denote the statonary polcy that maps states to actons (retal prcng functons) by π : S A,.e, a t = π(s t ). The objectve of our dynamc prcng problem s to fnd an optmal polcy π for each state s S that mnmzes the expected dscounted system cost of the servce provder as n the followng MDP problem (P): [ ] (P) : mn E (γ) t r t (s t,π(s t )), (9) π:s A t= where γ < 1 s the dscount factor whch represents the relatve mportance of the future system cost compared wth the present system cost. The optmal statonary polcy π can be well defned by usng the optmal acton-value functon Q : S A R whch satsfes the followng Bellman optmalty equaton: Q (s,a) = r(s,a)+γ s Sp(s s,a)v (s ), () where V (s ) s the optmal state-value functon [9], whch s defned as V (s ) = mn a A Q (s,a), s S. (11) 648

4 4 Snce Q (s,a) s the expected dscounted system cost wth acton a n state s, we can obtan the optmal statonary polcy as π (s) = argmnq (s,a). (12) a A In ths paper, we use the well-known Q-learnng algorthm to solve our MDP problem (P) wthout nformaton of state transton probabltes. We refer the readers to [] for more detal on the Q-learnng algorthm. In the followng subsectons, n order to resolve the crtcal ssues whch make t dffcult to apply the conventonal Q-learnng algorthm to our smart grd system, we propose an alternatve state defnton based on the observed total energy consumpton as well as adopt vrtual experence n Q-learnng updates. B. Energy Consumpton-Based Approxmate State (EAS) When the sze of the state space s large, the Q-learnng algorthm requres not only a large memory space to store the state-acton functon Q(s, a), but also a long tme to converge. Moreover, n the practcal smart grd system, t s dffcult for the servce provder to acqure or use the nformaton about the customers current load demands due to prvacy. In order to resolve these dffcultes, n ths secton, we propose an alternatve defnton of the system state, whch s based on the observed total energy consumpton, I êt 1, and the prevously chosen acton a t 1. For notatonal convenence, we wll omt d t n êt (dt ) n the rest of ths secton. The man dea of ths alternatve state defnton comes from the fact that, snce each customer s dsutlty functon u s a decreasng functon of ê t, from the load demand update process n (4), the retal prcng functon and each customer s energy consumpton at tme-slot t 1 n a tuple (ê t 1,a t 1 ) characterzes the past accumulated load demand d t 1. Hence, f the servce provder knows new load demand D t(ht ) at tme-slot t, t can regard a dfferent tuple (ê t 1,a t 1,D t(ht )) as a dfferent actual accumulated load demand of customer at tme-slot t. Smlarly, once a tuple ( I êt 1,a t 1, I Dt (ht )) s observed by the servce provder, t approxmately reflects the customers overall load demands at tme-slot t. It s worth notng that snce D t(ht ) s ndependent random varable for each customer, by the law of the large number, the average of the sum of new load demands, I Dt (ht )/ I, goes to ts expected value as the number of customers gets larger. Ths mples that n the practcal smart grd system wth a large number of customers, a tuple( I êt 1,a t 1 ) provdes enough nformaton for the servce provder to nfer the customers overall load demand level at tme-slot t. Hence, nstead of the orgnal state s t n (6), we can use a new state defnton based on the observed energy consumpton by whch the servce provder does not need to know ether each customer s load demand or ts dsutlty functon. To reduce the number of system states, we dscretze the observed energy consumpton I êt 1 nto a fnte number of energy levels ne 3 The set of observed energy consumpton level by usng a quantzaton operaton q E ( ). Then, we refer 3 In our system model, the method of dscretzaton of the observed energy consumpton s not lmted to a specfc method. A smple example s to use an equally dvded levels between the maxmum and mnmum amounts of the energy consumpton. Algorthm 1 Q-Learnng Algorthm wth Vrtual Experence 1: Intalze Q arbtrarly, t = 2: for each tme-slot t 3: Choose a t accordng to polcy π(x t ) 4: Take acton a t, observe system cost r(x t,a t ) and next state x t+1 : Obtan experence tuple σ t+1 = (x t,a t,r t,x t+1 ) 6: Generate set of vrtual experence tuples θ(σ t+1 ) 7: for each vrtual experence tuple σ t+1 θ(σ t+1 ) 8: v = r(x t,a t )+γmax a AQ(x t+1,a ) Q(x t,a t ) 9: Q(x t,a t ) Q(x t,a t )+α t v : end 11: end to tuple (q E ( I êt 1 ),a t 1 ) as the approxmate demand at tme-slot t and represent t as ( ) d t app = (q E ê t 1,a t 1). (13) I Based on the approxmate demand, we now defne the energy consumpton-based approxmate state (EAS) of the smart grd system as x t = (d t app,h t,c t ) X, (14) where X = E A H C denotes the set of the EASs. Note that EAS extremely reduces the number of states from S = I D H C to X = E A H C, whle allowng the servce provder to easly nfer the customers current load demands level wthout usng drect sgnalng from the customers. Now, we can smply substtute the orgnal state defnton s t by EAS x t n the Q-learnng algorthm. C. Accelerated Learnng usng Vrtual Experence Although the EAS x t sgnfcantly reduces the state space, the learnng speed of the Q-learnng algorthm mght be serously lmted by ts nherent structure n whch only one state-acton par s updated at each tme-slot. In ths subsecton, n order to mprove the speed of the Q-learnng algorthm, we adopt vrtual experence whch was ntroduced n [11]. The Q-learnng algorthm wth vrtual experence enables the servce provder to update multple state-acton pars at each tme-slot by explotng a pror known partal nformaton of the state transton probablty p x (x t+1 x t,a t ). In ths subsecton, we consder the case where the servce provder knows the transton probablty of the wholesale prcng functon p c (c t+1 c t,h t ) a pror. Note that ths s not a too restrctve assumpton because the servce provder can gather suffcent data for the transton probablty of the wholesale prcng functon whle partcpatng n the wholesale electrcty market. We frst defne the experence tuple (ET) observed by the servce provder at tme-slot t+1 as σ t+1 = (x t,a t,r t,x t+1 ), where r t s the observed system cost. Then, gven the actual ET σ t+1, we defne a set of vrtual experence tuples (vrtual ETs) θ(σ t+1 ), whch are statstcally equvalent 4 to the actual ET σ t+1,.e., θ(σ t+1 ) = σt+1 d t app = dt app, h t = h t, ã t = a t, r t = r( c t ), p c ( c t+1 c t, h t ) = p c (c t+1 c t,h t ), (1) where r(c) represents the system cost that s vrtually calculated by usng an arbtrary wholesale prcng functonc. In our 4 An ET σ t+1 = ( x t,ã t, r t, x t+1 ) s sad to be statstcally equvalent to ET σ t+1 = (x t,a t,r t,x t+1 ) f p x( x t+1 x t,ã t ) = p x(x t+1 x t,a t ) and the system cost r t can be calculated by usng σ t

5 TABLE I COMPLEXITY COMPARISON OF THREE DIFFERENT ALGORITHMS. Q-learnng wth orgnal state Q-learnng wth EAS Q-learnng wth EAS and vrtual experence 3 Learnng update complexty per teraton O( A ) Memory complexty O( I D H C A ) O( A ) O( E H C A 2 ) O( θ(ˆσ) A ) O( E H C A 2 ) Average system cost Renforcement learnng Myopc optmzaton kwh Tme slot Fg. 2. Load demand profle. smart grd system, f d t app and ht are fxed, the system cost r t can be easly calculated for an arbtrary wholesale prcng functons c t C by applyng the same energy consumpton I êt to (8). Moreover, snce the transton probablty of the wholesale prcng functon p c (c t+1 c t,h t ) s ndependent of approxmate demand d t app and retal electrcty prcng functon a t, we can easly generate the set of vrtual ETs θ(σ t+1 ) n (1) from the observed actual ET σ t+1. Whle the observed ET σ t+1 s used to update only one state-acton functon Q(x t,a t ) n the conventonal Q-learnng, by usng the vrtual ETs, the Q-learnng algorthm can update multple state-acton pars at each tme-slot. The Q-learnng algorthm wth vrtual experence s outlned n Algorthm 1. Lnes 3- descrbe the operaton of the conventonal Q-learnng where the servce provder obtans the experence tuple σ t+1. Then, n lne 6, based on σ t+1, a set of ts vrtual experences s generated. In lnes 7-, the acton-value functon Q(x t,a t ) s updated for all vrtual experences n σ t+1. The complexty of the proposed renforcement learnng algorthms are summarzed n Table I. Although the Q-learnng algorthm wth vrtual experence has a hgher update complexty than the Q-learnng algorthm wthout vrtual experence, as we wll show n Secton IV, t sgnfcantly reduces the number of tme-slots needed to converge, whch s regarded as a more mportant aspect than the computatonal complexty at each tme-slot n renforcement learnng algorthms. IV. NUMERICAL RESULTS In ths secton, we provde numercal results to evaluate the performance of our dynamc prcng algorthm. One day conssts of 24 tme-slots each of whch lasts for one hour. Hence, the set of perods s gven as H = {,1,,23} and the mappng between tme-slots and perods s gven as h t = mod (t,24) λ Fg. 3. Performance comparson of our renforcement learnng algorthm and the myopc optmzaton algorthm varyng λ. We consder a smart grd system wth 2 customers. The newly generated load demand of customer, D t(ht ), follows a Posson dstrbuton wth expected value ω,h t, whch s proportonal to the hourly average load shapes of resdental electrcty servces n Calforna [12] as shown n Fg. 2. All customers have the same backlog rate,.e., λ = λ, I. Each customer s dsutlty functon u (d t et ) s gven as u (d t e t ) = κ (d t e t ) 2, (16) where κ s a constant that represents customer s dsutlty senstvty to ts remanng demand. Here, we let κ = κ =.1, I. We model the wholesale prcng functon c t as a quadratc functon of the total energy consumpton I êt as n [3],[4], and [6] : c t( ) ê t = µ t ( 2. ê t +νh t ê) t (17) t I I We set µ t =.2, t andν t h t to be a random varable whose expected value, v h t, changes accordng to the correspondng perod h t based on the hourly average load shape n Fg. 2. Wth a gven perodh t,ν t h t s unform randomly chosen among values n {.2v h t,.v h t,,1.7v h t}. The dscount factor γ n problem (P) s fxed to.9. The retal prcng functon a t s a lnear functon of the energy consumpton ê t,.e., I a t (ê t ) = χt ê t, (18) where the coeffcent χ t can be chosen among set {.2,.4,,1.} each element of whch s drectly mapped to one retal prcng functon n A. We frst evaluate the performance of the dynamc prcng algorthm by comparng t wth that of the myopc optmzaton algorthm. In the myopc optmzaton algorthm, the servce provder chooses an acton wth the lowest expected nstantaneous system cost, by updatng the state-acton functon Q(s, a) smlarly to the Q-learnng update wth a dscount factor γ =. Ths mples that the myopc optmzaton algorthm focuses only on the mmedate system cost wthout consderng the mpact of the current acton on the future system cost. In Fg. 3, we show the average system costs of those two dynamc prcng algorthms by changng the backlog rate, λ, from to 1. We set ρ =. whch corresponds to the case where the servce provder ams at mnmzng the sum of the total dsutlty and the wholesale cost. We can observe that the average system costs ncrease as λ ncreases 6

6 6 Cost of customers $/kwh Average cost of customers Average cost of servce provder ρ 1. Average retal prce ρ Fg. 4. Impact of the weghtng factor ρ on the performances of customers and servce provder. Average system cost Tme slot x 4 1 Cost of servce provder wth vrtual experence wthout vrtual experence Fg.. Impact of vrtual experence on the learnng speed of the Q-learnng algorthm. n both dynamc prcng algorthms because wth a hgher backlog rate, the accumulated load demand causes a hgher dsutlty. We also observe that the performance gap between two algorthms ncreases as λ ncreases. Wth a low backlog rate, the remanng backlog s not carred forward to the next tme-slot. In ths case, the soluton of our renforcement learnng algorthm s the same as that of the myopc optmzaton algorthm. On the contrary, n the case wth a hgher backlog rate, the remanng backlog s carred forward to the next tmeslot. Hence, the servce provder s prcng decson at a tmeslot nfluences the accumulated load demand n the future, and thus ts future system cost. Due to ths dfference, especally when λ s large, our algorthm acheves better performance than the myopc optmzaton problem whch consders only the current system cost. To study the mpact of the weghtng factor ρ, n Fg. 4, we show the cost of customers, that of the servce provder, and the average retal prce wth varyng ρ from to 1. We set λ =.. We can observe that as ρ ncreases, the servce provder reduces the average retal prce, the cost of customers decreases, and the cost of the servce provder ncreases. For example, n the case wth ρ =, the servce provder ams at mnmzng ts own cost. Hence, the servce provder does not consder the customers dsutlty and chooses relatvely hgh prces to reduce the wholesale cost whch contrbutes to most of ts own cost. In the case wth ρ = 1, the servce provder ams at mnmzng the customers cost. Hence, the servce provder chooses relatvely low prces to provde electrc energy to the customers at a low retal prce as possble. In Fg., we compare the learnng speed of our renforcement learnng algorthm wth vrtual experence to the conventonal Q-learnng algorthm wthout vrtual experence. We set λ =. and ρ =.. We can observe that our algorthm wth vrtual experence acheves the near optmal average system cost after about 3, tme-slots, whle the conventonal Q-learnng algorthm shows a worse learnng speed. Ths means that even f the stochastc characterstc of the system dynamcs vary n tme, the proposed renforcement learnng algorthm can quckly adapt to the tme-varyng envronment by explotng vrtual experence update. V. CONCLUSION In ths paper, we studed a dynamc prcng problem for the smart grd system where the servce provder can adaptvely decde the electrcty prce accordng to the customers load demand levels and the wholesale prce. We developed a renforcement learnng-based dynamc prcng algorthm that enables effcent dynamc prcng wthout requrng the perfect nformaton about the system dynamcs a pror. To resolve the exstng drawbacks of the conventonal renforcement learnng algorthm, we proposed two mprovements: energy consumpton-based approxmate state defnton and the adopton of vrtual experence update n the conventonal Q-learnng algorthm. Numercal results show that the renforcement learnng-based dynamc prcng acheves a hgher long-term performance compared to the myopc optmzaton approach especally n the system where the customers have a hgh demand backlog rate. The results also show that our algorthm results n an mproved learnng speed due to the alternatve state defnton and vrtual experence mplyng that our dynamc prcng algorthm can be appled to the practcal smart grd system. REFERENCES [1] M. H. Albad and E. El-Saadany, Demand response n electrcty markets: An overvew, n IEEE Power Engneerng Socety General Meetng, 27. [2] B.-G. Km, S. Ren, M. van der Schaar, and J.-W. Lee, Bdrectonal energy tradng and resdental load schedulng wth electrc vehcles n the smart grd, IEEE Journal on Selected Areas n Communcatons, vol. 31, no. 7, pp , 213. [3] P. Samad, A.-H. Mohsenan-Rad, R. Schober, V. Wong, and J. Jatskevch, Optmal real-tme prcng algorthm based on utlty maxmzaton for smart grd, n IEEE SmartGrdComm, 2. [4] P. Tarasak, Optmal real-tme prcng under load uncertanty based on utlty maxmzaton for smart grd, n IEEE SmartGrdComm, 211. [] L. P. Qan, Y. Zhang, J. Huang, and Y. Wu, Demand response management va real-tme electrcty prce control n smart grds, IEEE Journal on Selected Areas n Communcatons, vol. 31, no. 7, pp , 213. [6] C. Chen, S. Kshore, and L. Snyder, An nnovatve RTP-based resdental power schedulng scheme for smart grds, n IEEE ICASSP, 211. [7] C. Joe-Wong, S. Sen, S. Ha, and M. Chang, Optmzed day-ahead prcng for smart grds wth devce-specfc schedulng flexblty, IEEE Journal on Selected Areas n Communcatons, vol. 3, no. 6, pp. 7 8, 212. [8] M. He, S. Murugesan, and J. Zhang, Multple tmescale dspatch and schedulng for stochastc relablty n smart grds wth wnd generaton ntegraton, n IEEE INFOCOM, 211, pp [9] L. P. Kaelblng, M. L. Lttman, and A. W. Moore, Renforcement learnng: A survey, Journal of Artfcal Intellgence Research, vol. 4, pp , May [] C. Watkns, Learnng from delayed rewards, Ph.D. dssertaton, Cambrdge Unversty, [11] N. Mastronarde and M. van der Schaar, Jont physcal-layer and systemlevel power management for delay-senstve wreless communcatons, IEEE Transactons on Moble Computng, vol. 12, no. 4, pp , 213. [12] Dynamc Load Profles n Calforna. Pacfc Gas & Electrc. [Onlne]. Avalable: use prces.shtml 61

Foresighted Demand Side Management

Foresighted Demand Side Management Foresghted Demand Sde Management 1 Yuanzhang Xao and Mhaela van der Schaar, Fellow, IEEE Department of Electrcal Engneerng, UCLA. {yxao,mhaela}@ee.ucla.edu. Abstract arxv:1401.2185v1 [cs.ma] 9 Jan 2014

More information

Winter 2008 CS567 Stochastic Linear/Integer Programming Guest Lecturer: Xu, Huan

Winter 2008 CS567 Stochastic Linear/Integer Programming Guest Lecturer: Xu, Huan Wnter 2008 CS567 Stochastc Lnear/Integer Programmng Guest Lecturer: Xu, Huan Class 2: More Modelng Examples 1 Capacty Expanson Capacty expanson models optmal choces of the tmng and levels of nvestments

More information

Online Appendix. t=1 (p t w)q t. Then the first order condition shows that

Online Appendix. t=1 (p t w)q t. Then the first order condition shows that Artcle forthcomng to ; manuscrpt no (Please, provde the manuscrpt number!) 1 Onlne Appendx Appendx E: Proofs Proof of Proposton 1 Frst we derve the equlbrum when the manufacturer does not vertcally ntegrate

More information

COMPARISON OF SOME RELIABILITY CHARACTERISTICS BETWEEN REDUNDANT SYSTEMS REQUIRING SUPPORTING UNITS FOR THEIR OPERATIONS

COMPARISON OF SOME RELIABILITY CHARACTERISTICS BETWEEN REDUNDANT SYSTEMS REQUIRING SUPPORTING UNITS FOR THEIR OPERATIONS Avalable onlne at http://sck.org J. Math. Comput. Sc. 3 (3), No., 6-3 ISSN: 97-537 COMPARISON OF SOME RELIABILITY CHARACTERISTICS BETWEEN REDUNDANT SYSTEMS REQUIRING SUPPORTING UNITS FOR THEIR OPERATIONS

More information

Resource Allocation with a Budget Constraint for Computing Independent Tasks in the Cloud

Resource Allocation with a Budget Constraint for Computing Independent Tasks in the Cloud Resource Allocaton wth a Budget Constrant for Computng Independent Tasks n the Cloud Wemng Sh and Bo Hong School of Electrcal and Computer Engneerng Georga Insttute of Technology, USA 2nd IEEE Internatonal

More information

An Online Learning Algorithm for Demand Response in Smart Grid

An Online Learning Algorithm for Demand Response in Smart Grid An Onlne Learnng Algorthm for Demand Response n Smart Grd Shahab Bahram, Student Member, IEEE, Vncent W.S. Wong, Fellow, IEEE, and Janwe Huang, Fellow, IEEE Abstract Demand response program wth real-tme

More information

Generalized Linear Methods

Generalized Linear Methods Generalzed Lnear Methods 1 Introducton In the Ensemble Methods the general dea s that usng a combnaton of several weak learner one could make a better learner. More formally, assume that we have a set

More information

ECE559VV Project Report

ECE559VV Project Report ECE559VV Project Report (Supplementary Notes Loc Xuan Bu I. MAX SUM-RATE SCHEDULING: THE UPLINK CASE We have seen (n the presentaton that, for downlnk (broadcast channels, the strategy maxmzng the sum-rate

More information

Lecture Notes on Linear Regression

Lecture Notes on Linear Regression Lecture Notes on Lnear Regresson Feng L fl@sdueducn Shandong Unversty, Chna Lnear Regresson Problem In regresson problem, we am at predct a contnuous target value gven an nput feature vector We assume

More information

Assortment Optimization under MNL

Assortment Optimization under MNL Assortment Optmzaton under MNL Haotan Song Aprl 30, 2017 1 Introducton The assortment optmzaton problem ams to fnd the revenue-maxmzng assortment of products to offer when the prces of products are fxed.

More information

Pricing and Resource Allocation Game Theoretic Models

Pricing and Resource Allocation Game Theoretic Models Prcng and Resource Allocaton Game Theoretc Models Zhy Huang Changbn Lu Q Zhang Computer and Informaton Scence December 8, 2009 Z. Huang, C. Lu, and Q. Zhang (CIS) Game Theoretc Models December 8, 2009

More information

Markov Chain Monte Carlo Lecture 6

Markov Chain Monte Carlo Lecture 6 where (x 1,..., x N ) X N, N s called the populaton sze, f(x) f (x) for at least one {1, 2,..., N}, and those dfferent from f(x) are called the tral dstrbutons n terms of mportance samplng. Dfferent ways

More information

Lecture 14: Bandits with Budget Constraints

Lecture 14: Bandits with Budget Constraints IEOR 8100-001: Learnng and Optmzaton for Sequental Decson Makng 03/07/16 Lecture 14: andts wth udget Constrants Instructor: Shpra Agrawal Scrbed by: Zhpeng Lu 1 Problem defnton In the regular Mult-armed

More information

Errors for Linear Systems

Errors for Linear Systems Errors for Lnear Systems When we solve a lnear system Ax b we often do not know A and b exactly, but have only approxmatons  and ˆb avalable. Then the best thng we can do s to solve ˆx ˆb exactly whch

More information

EEE 241: Linear Systems

EEE 241: Linear Systems EEE : Lnear Systems Summary #: Backpropagaton BACKPROPAGATION The perceptron rule as well as the Wdrow Hoff learnng were desgned to tran sngle layer networks. They suffer from the same dsadvantage: they

More information

For now, let us focus on a specific model of neurons. These are simplified from reality but can achieve remarkable results.

For now, let us focus on a specific model of neurons. These are simplified from reality but can achieve remarkable results. Neural Networks : Dervaton compled by Alvn Wan from Professor Jtendra Malk s lecture Ths type of computaton s called deep learnng and s the most popular method for many problems, such as computer vson

More information

Econ107 Applied Econometrics Topic 3: Classical Model (Studenmund, Chapter 4)

Econ107 Applied Econometrics Topic 3: Classical Model (Studenmund, Chapter 4) I. Classcal Assumptons Econ7 Appled Econometrcs Topc 3: Classcal Model (Studenmund, Chapter 4) We have defned OLS and studed some algebrac propertes of OLS. In ths topc we wll study statstcal propertes

More information

EEL 6266 Power System Operation and Control. Chapter 3 Economic Dispatch Using Dynamic Programming

EEL 6266 Power System Operation and Control. Chapter 3 Economic Dispatch Using Dynamic Programming EEL 6266 Power System Operaton and Control Chapter 3 Economc Dspatch Usng Dynamc Programmng Pecewse Lnear Cost Functons Common practce many utltes prefer to represent ther generator cost functons as sngle-

More information

A Robust Method for Calculating the Correlation Coefficient

A Robust Method for Calculating the Correlation Coefficient A Robust Method for Calculatng the Correlaton Coeffcent E.B. Nven and C. V. Deutsch Relatonshps between prmary and secondary data are frequently quantfed usng the correlaton coeffcent; however, the tradtonal

More information

NUMERICAL DIFFERENTIATION

NUMERICAL DIFFERENTIATION NUMERICAL DIFFERENTIATION 1 Introducton Dfferentaton s a method to compute the rate at whch a dependent output y changes wth respect to the change n the ndependent nput x. Ths rate of change s called the

More information

Kernel Methods and SVMs Extension

Kernel Methods and SVMs Extension Kernel Methods and SVMs Extenson The purpose of ths document s to revew materal covered n Machne Learnng 1 Supervsed Learnng regardng support vector machnes (SVMs). Ths document also provdes a general

More information

Simultaneous Optimization of Berth Allocation, Quay Crane Assignment and Quay Crane Scheduling Problems in Container Terminals

Simultaneous Optimization of Berth Allocation, Quay Crane Assignment and Quay Crane Scheduling Problems in Container Terminals Smultaneous Optmzaton of Berth Allocaton, Quay Crane Assgnment and Quay Crane Schedulng Problems n Contaner Termnals Necat Aras, Yavuz Türkoğulları, Z. Caner Taşkın, Kuban Altınel Abstract In ths work,

More information

Lecture 4. Instructor: Haipeng Luo

Lecture 4. Instructor: Haipeng Luo Lecture 4 Instructor: Hapeng Luo In the followng lectures, we focus on the expert problem and study more adaptve algorthms. Although Hedge s proven to be worst-case optmal, one may wonder how well t would

More information

Numerical Heat and Mass Transfer

Numerical Heat and Mass Transfer Master degree n Mechancal Engneerng Numercal Heat and Mass Transfer 06-Fnte-Dfference Method (One-dmensonal, steady state heat conducton) Fausto Arpno f.arpno@uncas.t Introducton Why we use models and

More information

Global Sensitivity. Tuesday 20 th February, 2018

Global Sensitivity. Tuesday 20 th February, 2018 Global Senstvty Tuesday 2 th February, 28 ) Local Senstvty Most senstvty analyses [] are based on local estmates of senstvty, typcally by expandng the response n a Taylor seres about some specfc values

More information

Modeling and Design of Real-Time Pricing Systems Based on Markov Decision Processes

Modeling and Design of Real-Time Pricing Systems Based on Markov Decision Processes Appled Mathematcs, 04, 5, 485-495 Publshed Onlne June 04 n ScRes. http://www.scrp.org/journal/am http://dx.do.org/0.436/am.04.504 Modelng and Desgn of Real-Tme Prcng Systems Based on Markov Decson Processes

More information

Demand Response with Linear Bidding: Efficiency vs. Risk. Munther A. Dahleh MIT Institute for Data, Systems, and Society

Demand Response with Linear Bidding: Efficiency vs. Risk. Munther A. Dahleh MIT Institute for Data, Systems, and Society Demand Response wth Lnear Bddng: Effcency vs. Rsk Munther A. Dahleh MIT Insttute for Data, Systems, and Socety Collaoraton Na L : Harvard Unversty Ljun Chen: Unversty of Colorado at Boulder Qngqng Huang:

More information

A PROBABILITY-DRIVEN SEARCH ALGORITHM FOR SOLVING MULTI-OBJECTIVE OPTIMIZATION PROBLEMS

A PROBABILITY-DRIVEN SEARCH ALGORITHM FOR SOLVING MULTI-OBJECTIVE OPTIMIZATION PROBLEMS HCMC Unversty of Pedagogy Thong Nguyen Huu et al. A PROBABILITY-DRIVEN SEARCH ALGORITHM FOR SOLVING MULTI-OBJECTIVE OPTIMIZATION PROBLEMS Thong Nguyen Huu and Hao Tran Van Department of mathematcs-nformaton,

More information

Linear Approximation with Regularization and Moving Least Squares

Linear Approximation with Regularization and Moving Least Squares Lnear Approxmaton wth Regularzaton and Movng Least Squares Igor Grešovn May 007 Revson 4.6 (Revson : March 004). 5 4 3 0.5 3 3.5 4 Contents: Lnear Fttng...4. Weghted Least Squares n Functon Approxmaton...

More information

(1 ) (1 ) 0 (1 ) (1 ) 0

(1 ) (1 ) 0 (1 ) (1 ) 0 Appendx A Appendx A contans proofs for resubmsson "Contractng Informaton Securty n the Presence of Double oral Hazard" Proof of Lemma 1: Assume that, to the contrary, BS efforts are achevable under a blateral

More information

Lecture 10 Support Vector Machines II

Lecture 10 Support Vector Machines II Lecture 10 Support Vector Machnes II 22 February 2016 Taylor B. Arnold Yale Statstcs STAT 365/665 1/28 Notes: Problem 3 s posted and due ths upcomng Frday There was an early bug n the fake-test data; fxed

More information

Module 3 LOSSY IMAGE COMPRESSION SYSTEMS. Version 2 ECE IIT, Kharagpur

Module 3 LOSSY IMAGE COMPRESSION SYSTEMS. Version 2 ECE IIT, Kharagpur Module 3 LOSSY IMAGE COMPRESSION SYSTEMS Verson ECE IIT, Kharagpur Lesson 6 Theory of Quantzaton Verson ECE IIT, Kharagpur Instructonal Objectves At the end of ths lesson, the students should be able to:

More information

2E Pattern Recognition Solutions to Introduction to Pattern Recognition, Chapter 2: Bayesian pattern classification

2E Pattern Recognition Solutions to Introduction to Pattern Recognition, Chapter 2: Bayesian pattern classification E395 - Pattern Recognton Solutons to Introducton to Pattern Recognton, Chapter : Bayesan pattern classfcaton Preface Ths document s a soluton manual for selected exercses from Introducton to Pattern Recognton

More information

Feature Selection: Part 1

Feature Selection: Part 1 CSE 546: Machne Learnng Lecture 5 Feature Selecton: Part 1 Instructor: Sham Kakade 1 Regresson n the hgh dmensonal settng How do we learn when the number of features d s greater than the sample sze n?

More information

Module 9. Lecture 6. Duality in Assignment Problems

Module 9. Lecture 6. Duality in Assignment Problems Module 9 1 Lecture 6 Dualty n Assgnment Problems In ths lecture we attempt to answer few other mportant questons posed n earler lecture for (AP) and see how some of them can be explaned through the concept

More information

The Second Anti-Mathima on Game Theory

The Second Anti-Mathima on Game Theory The Second Ant-Mathma on Game Theory Ath. Kehagas December 1 2006 1 Introducton In ths note we wll examne the noton of game equlbrum for three types of games 1. 2-player 2-acton zero-sum games 2. 2-player

More information

The Multiple Classical Linear Regression Model (CLRM): Specification and Assumptions. 1. Introduction

The Multiple Classical Linear Regression Model (CLRM): Specification and Assumptions. 1. Introduction ECONOMICS 5* -- NOTE (Summary) ECON 5* -- NOTE The Multple Classcal Lnear Regresson Model (CLRM): Specfcaton and Assumptons. Introducton CLRM stands for the Classcal Lnear Regresson Model. The CLRM s also

More information

MMA and GCMMA two methods for nonlinear optimization

MMA and GCMMA two methods for nonlinear optimization MMA and GCMMA two methods for nonlnear optmzaton Krster Svanberg Optmzaton and Systems Theory, KTH, Stockholm, Sweden. krlle@math.kth.se Ths note descrbes the algorthms used n the author s 2007 mplementatons

More information

Hila Etzion. Min-Seok Pang

Hila Etzion. Min-Seok Pang RESERCH RTICLE COPLEENTRY ONLINE SERVICES IN COPETITIVE RKETS: INTINING PROFITILITY IN THE PRESENCE OF NETWORK EFFECTS Hla Etzon Department of Technology and Operatons, Stephen. Ross School of usness,

More information

Welfare Properties of General Equilibrium. What can be said about optimality properties of resource allocation implied by general equilibrium?

Welfare Properties of General Equilibrium. What can be said about optimality properties of resource allocation implied by general equilibrium? APPLIED WELFARE ECONOMICS AND POLICY ANALYSIS Welfare Propertes of General Equlbrum What can be sad about optmalty propertes of resource allocaton mpled by general equlbrum? Any crteron used to compare

More information

Canonical transformations

Canonical transformations Canoncal transformatons November 23, 2014 Recall that we have defned a symplectc transformaton to be any lnear transformaton M A B leavng the symplectc form nvarant, Ω AB M A CM B DΩ CD Coordnate transformatons,

More information

Technical Note: Capacity Constraints Across Nests in Assortment Optimization Under the Nested Logit Model

Technical Note: Capacity Constraints Across Nests in Assortment Optimization Under the Nested Logit Model Techncal Note: Capacty Constrants Across Nests n Assortment Optmzaton Under the Nested Logt Model Jacob B. Feldman, Huseyn Topaloglu School of Operatons Research and Informaton Engneerng, Cornell Unversty,

More information

Lecture 12: Discrete Laplacian

Lecture 12: Discrete Laplacian Lecture 12: Dscrete Laplacan Scrbe: Tanye Lu Our goal s to come up wth a dscrete verson of Laplacan operator for trangulated surfaces, so that we can use t n practce to solve related problems We are mostly

More information

arxiv: v2 [cs.gt] 7 Aug 2017

arxiv: v2 [cs.gt] 7 Aug 2017 Stochastc Games for Smart Grd Energy Management wth Prospect Prosumers S. Rasoul Etesam, Wald Saad, Narayan Mandayam, and H. Vncent Poor arxv:6.67v [cs.gt] 7 Aug 7 Abstract In ths paper, the problem of

More information

LOW BIAS INTEGRATED PATH ESTIMATORS. James M. Calvin

LOW BIAS INTEGRATED PATH ESTIMATORS. James M. Calvin Proceedngs of the 007 Wnter Smulaton Conference S G Henderson, B Bller, M-H Hseh, J Shortle, J D Tew, and R R Barton, eds LOW BIAS INTEGRATED PATH ESTIMATORS James M Calvn Department of Computer Scence

More information

The Study of Teaching-learning-based Optimization Algorithm

The Study of Teaching-learning-based Optimization Algorithm Advanced Scence and Technology Letters Vol. (AST 06), pp.05- http://dx.do.org/0.57/astl.06. The Study of Teachng-learnng-based Optmzaton Algorthm u Sun, Yan fu, Lele Kong, Haolang Q,, Helongang Insttute

More information

Annexes. EC.1. Cycle-base move illustration. EC.2. Problem Instances

Annexes. EC.1. Cycle-base move illustration. EC.2. Problem Instances ec Annexes Ths Annex frst llustrates a cycle-based move n the dynamc-block generaton tabu search. It then dsplays the characterstcs of the nstance sets, followed by detaled results of the parametercalbraton

More information

Simulated Power of the Discrete Cramér-von Mises Goodness-of-Fit Tests

Simulated Power of the Discrete Cramér-von Mises Goodness-of-Fit Tests Smulated of the Cramér-von Mses Goodness-of-Ft Tests Steele, M., Chaselng, J. and 3 Hurst, C. School of Mathematcal and Physcal Scences, James Cook Unversty, Australan School of Envronmental Studes, Grffth

More information

Dynamic Programming. Lecture 13 (5/31/2017)

Dynamic Programming. Lecture 13 (5/31/2017) Dynamc Programmng Lecture 13 (5/31/2017) - A Forest Thnnng Example - Projected yeld (m3/ha) at age 20 as functon of acton taken at age 10 Age 10 Begnnng Volume Resdual Ten-year Volume volume thnned volume

More information

An Admission Control Algorithm in Cloud Computing Systems

An Admission Control Algorithm in Cloud Computing Systems An Admsson Control Algorthm n Cloud Computng Systems Authors: Frank Yeong-Sung Ln Department of Informaton Management Natonal Tawan Unversty Tape, Tawan, R.O.C. ysln@m.ntu.edu.tw Yngje Lan Management Scence

More information

Structure and Drive Paul A. Jensen Copyright July 20, 2003

Structure and Drive Paul A. Jensen Copyright July 20, 2003 Structure and Drve Paul A. Jensen Copyrght July 20, 2003 A system s made up of several operatons wth flow passng between them. The structure of the system descrbes the flow paths from nputs to outputs.

More information

Experience with Automatic Generation Control (AGC) Dynamic Simulation in PSS E

Experience with Automatic Generation Control (AGC) Dynamic Simulation in PSS E Semens Industry, Inc. Power Technology Issue 113 Experence wth Automatc Generaton Control (AGC) Dynamc Smulaton n PSS E Lu Wang, Ph.D. Staff Software Engneer lu_wang@semens.com Dngguo Chen, Ph.D. Staff

More information

3.1 Expectation of Functions of Several Random Variables. )' be a k-dimensional discrete or continuous random vector, with joint PMF p (, E X E X1 E X

3.1 Expectation of Functions of Several Random Variables. )' be a k-dimensional discrete or continuous random vector, with joint PMF p (, E X E X1 E X Statstcs 1: Probablty Theory II 37 3 EPECTATION OF SEVERAL RANDOM VARIABLES As n Probablty Theory I, the nterest n most stuatons les not on the actual dstrbuton of a random vector, but rather on a number

More information

Lecture 21: Numerical methods for pricing American type derivatives

Lecture 21: Numerical methods for pricing American type derivatives Lecture 21: Numercal methods for prcng Amercan type dervatves Xaoguang Wang STAT 598W Aprl 10th, 2014 (STAT 598W) Lecture 21 1 / 26 Outlne 1 Fnte Dfference Method Explct Method Penalty Method (STAT 598W)

More information

The Minimum Universal Cost Flow in an Infeasible Flow Network

The Minimum Universal Cost Flow in an Infeasible Flow Network Journal of Scences, Islamc Republc of Iran 17(2): 175-180 (2006) Unversty of Tehran, ISSN 1016-1104 http://jscencesutacr The Mnmum Unversal Cost Flow n an Infeasble Flow Network H Saleh Fathabad * M Bagheran

More information

Lecture 12: Classification

Lecture 12: Classification Lecture : Classfcaton g Dscrmnant functons g The optmal Bayes classfer g Quadratc classfers g Eucldean and Mahalanobs metrcs g K Nearest Neghbor Classfers Intellgent Sensor Systems Rcardo Guterrez-Osuna

More information

Indian Buffet Game With Negative Network Externality and Non-Bayesian Social Learning

Indian Buffet Game With Negative Network Externality and Non-Bayesian Social Learning IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 45, NO. 4, APRIL 2015 609 Indan Buffet Game Wth Negatve Network Externalty and Non-Bayesan Socal Learnng Chunxao Jang, Member, IEEE, Yan

More information

VQ widely used in coding speech, image, and video

VQ widely used in coding speech, image, and video at Scalar quantzers are specal cases of vector quantzers (VQ): they are constraned to look at one sample at a tme (memoryless) VQ does not have such constrant better RD perfomance expected Source codng

More information

Psychology 282 Lecture #24 Outline Regression Diagnostics: Outliers

Psychology 282 Lecture #24 Outline Regression Diagnostics: Outliers Psychology 282 Lecture #24 Outlne Regresson Dagnostcs: Outlers In an earler lecture we studed the statstcal assumptons underlyng the regresson model, ncludng the followng ponts: Formal statement of assumptons.

More information

COS 521: Advanced Algorithms Game Theory and Linear Programming

COS 521: Advanced Algorithms Game Theory and Linear Programming COS 521: Advanced Algorthms Game Theory and Lnear Programmng Moses Charkar February 27, 2013 In these notes, we ntroduce some basc concepts n game theory and lnear programmng (LP). We show a connecton

More information

Joint Scheduling and Power-Allocation for Interference Management in Wireless Networks

Joint Scheduling and Power-Allocation for Interference Management in Wireless Networks Jont Schedulng and Power-Allocaton for Interference Management n Wreless Networks Xn Lu *, Edwn K. P. Chong, and Ness B. Shroff * * School of Electrcal and Computer Engneerng Purdue Unversty West Lafayette,

More information

Credit Card Pricing and Impact of Adverse Selection

Credit Card Pricing and Impact of Adverse Selection Credt Card Prcng and Impact of Adverse Selecton Bo Huang and Lyn C. Thomas Unversty of Southampton Contents Background Aucton model of credt card solctaton - Errors n probablty of beng Good - Errors n

More information

CHAPTER 5 NUMERICAL EVALUATION OF DYNAMIC RESPONSE

CHAPTER 5 NUMERICAL EVALUATION OF DYNAMIC RESPONSE CHAPTER 5 NUMERICAL EVALUATION OF DYNAMIC RESPONSE Analytcal soluton s usually not possble when exctaton vares arbtrarly wth tme or f the system s nonlnear. Such problems can be solved by numercal tmesteppng

More information

A Distributed Spectrum Sharing Algorithm in Cognitive Radio Networks

A Distributed Spectrum Sharing Algorithm in Cognitive Radio Networks A Dstrbuted Spectrum Sharng Algorthm n Cogntve Rado Networks We Sun, Tong Lu, Yanmn Zhu, Bo L Shangha Jao Tong Unversty Hong Kong Unversty of Scence and Technology Emal: {sunwt,lutong2569164,yzhu}@sjtu.edu.cn,

More information

Some modelling aspects for the Matlab implementation of MMA

Some modelling aspects for the Matlab implementation of MMA Some modellng aspects for the Matlab mplementaton of MMA Krster Svanberg krlle@math.kth.se Optmzaton and Systems Theory Department of Mathematcs KTH, SE 10044 Stockholm September 2004 1. Consdered optmzaton

More information

Lecture 4: November 17, Part 1 Single Buffer Management

Lecture 4: November 17, Part 1 Single Buffer Management Lecturer: Ad Rosén Algorthms for the anagement of Networs Fall 2003-2004 Lecture 4: November 7, 2003 Scrbe: Guy Grebla Part Sngle Buffer anagement In the prevous lecture we taled about the Combned Input

More information

Computing Correlated Equilibria in Multi-Player Games

Computing Correlated Equilibria in Multi-Player Games Computng Correlated Equlbra n Mult-Player Games Chrstos H. Papadmtrou Presented by Zhanxang Huang December 7th, 2005 1 The Author Dr. Chrstos H. Papadmtrou CS professor at UC Berkley (taught at Harvard,

More information

Pop-Click Noise Detection Using Inter-Frame Correlation for Improved Portable Auditory Sensing

Pop-Click Noise Detection Using Inter-Frame Correlation for Improved Portable Auditory Sensing Advanced Scence and Technology Letters, pp.164-168 http://dx.do.org/10.14257/astl.2013 Pop-Clc Nose Detecton Usng Inter-Frame Correlaton for Improved Portable Audtory Sensng Dong Yun Lee, Kwang Myung Jeon,

More information

Outline. Communication. Bellman Ford Algorithm. Bellman Ford Example. Bellman Ford Shortest Path [1]

Outline. Communication. Bellman Ford Algorithm. Bellman Ford Example. Bellman Ford Shortest Path [1] DYNAMIC SHORTEST PATH SEARCH AND SYNCHRONIZED TASK SWITCHING Jay Wagenpfel, Adran Trachte 2 Outlne Shortest Communcaton Path Searchng Bellmann Ford algorthm Algorthm for dynamc case Modfcatons to our algorthm

More information

Proceedings of the 10th WSEAS International Confenrence on APPLIED MATHEMATICS, Dallas, Texas, USA, November 1-3,

Proceedings of the 10th WSEAS International Confenrence on APPLIED MATHEMATICS, Dallas, Texas, USA, November 1-3, roceedngs of the 0th WSEAS Internatonal Confenrence on ALIED MATHEMATICS, Dallas, Texas, USA, November -3, 2006 365 Impact of Statc Load Modelng on Industral Load Nodal rces G. REZA YOUSEFI M. MOHSEN EDRAM

More information

Chapter 5. Solution of System of Linear Equations. Module No. 6. Solution of Inconsistent and Ill Conditioned Systems

Chapter 5. Solution of System of Linear Equations. Module No. 6. Solution of Inconsistent and Ill Conditioned Systems Numercal Analyss by Dr. Anta Pal Assstant Professor Department of Mathematcs Natonal Insttute of Technology Durgapur Durgapur-713209 emal: anta.bue@gmal.com 1 . Chapter 5 Soluton of System of Lnear Equatons

More information

Appendix B: Resampling Algorithms

Appendix B: Resampling Algorithms 407 Appendx B: Resamplng Algorthms A common problem of all partcle flters s the degeneracy of weghts, whch conssts of the unbounded ncrease of the varance of the mportance weghts ω [ ] of the partcles

More information

The optimal delay of the second test is therefore approximately 210 hours earlier than =2.

The optimal delay of the second test is therefore approximately 210 hours earlier than =2. THE IEC 61508 FORMULAS 223 The optmal delay of the second test s therefore approxmately 210 hours earler than =2. 8.4 The IEC 61508 Formulas IEC 61508-6 provdes approxmaton formulas for the PF for smple

More information

Ryan (2009)- regulating a concentrated industry (cement) Firms play Cournot in the stage. Make lumpy investment decisions

Ryan (2009)- regulating a concentrated industry (cement) Firms play Cournot in the stage. Make lumpy investment decisions 1 Motvaton Next we consder dynamc games where the choce varables are contnuous and/or dscrete. Example 1: Ryan (2009)- regulatng a concentrated ndustry (cement) Frms play Cournot n the stage Make lumpy

More information

CS286r Assign One. Answer Key

CS286r Assign One. Answer Key CS286r Assgn One Answer Key 1 Game theory 1.1 1.1.1 Let off-equlbrum strateges also be that people contnue to play n Nash equlbrum. Devatng from any Nash equlbrum s a weakly domnated strategy. That s,

More information

Uncertainty in measurements of power and energy on power networks

Uncertainty in measurements of power and energy on power networks Uncertanty n measurements of power and energy on power networks E. Manov, N. Kolev Department of Measurement and Instrumentaton, Techncal Unversty Sofa, bul. Klment Ohrdsk No8, bl., 000 Sofa, Bulgara Tel./fax:

More information

On an Extension of Stochastic Approximation EM Algorithm for Incomplete Data Problems. Vahid Tadayon 1

On an Extension of Stochastic Approximation EM Algorithm for Incomplete Data Problems. Vahid Tadayon 1 On an Extenson of Stochastc Approxmaton EM Algorthm for Incomplete Data Problems Vahd Tadayon Abstract: The Stochastc Approxmaton EM (SAEM algorthm, a varant stochastc approxmaton of EM, s a versatle tool

More information

k t+1 + c t A t k t, t=0

k t+1 + c t A t k t, t=0 Macro II (UC3M, MA/PhD Econ) Professor: Matthas Kredler Fnal Exam 6 May 208 You have 50 mnutes to complete the exam There are 80 ponts n total The exam has 4 pages If somethng n the queston s unclear,

More information

Markov Chain Monte Carlo (MCMC), Gibbs Sampling, Metropolis Algorithms, and Simulated Annealing Bioinformatics Course Supplement

Markov Chain Monte Carlo (MCMC), Gibbs Sampling, Metropolis Algorithms, and Simulated Annealing Bioinformatics Course Supplement Markov Chan Monte Carlo MCMC, Gbbs Samplng, Metropols Algorthms, and Smulated Annealng 2001 Bonformatcs Course Supplement SNU Bontellgence Lab http://bsnuackr/ Outlne! Markov Chan Monte Carlo MCMC! Metropols-Hastngs

More information

Parameter Estimation for Dynamic System using Unscented Kalman filter

Parameter Estimation for Dynamic System using Unscented Kalman filter Parameter Estmaton for Dynamc System usng Unscented Kalman flter Jhoon Seung 1,a, Amr Atya F. 2,b, Alexander G.Parlos 3,c, and Klto Chong 1,4,d* 1 Dvson of Electroncs Engneerng, Chonbuk Natonal Unversty,

More information

A Particle Filter Algorithm based on Mixing of Prior probability density and UKF as Generate Importance Function

A Particle Filter Algorithm based on Mixing of Prior probability density and UKF as Generate Importance Function Advanced Scence and Technology Letters, pp.83-87 http://dx.do.org/10.14257/astl.2014.53.20 A Partcle Flter Algorthm based on Mxng of Pror probablty densty and UKF as Generate Importance Functon Lu Lu 1,1,

More information

Adaptive Consensus Control of Multi-Agent Systems with Large Uncertainty and Time Delays *

Adaptive Consensus Control of Multi-Agent Systems with Large Uncertainty and Time Delays * Journal of Robotcs, etworkng and Artfcal Lfe, Vol., o. (September 04), 5-9 Adaptve Consensus Control of Mult-Agent Systems wth Large Uncertanty and me Delays * L Lu School of Mechancal Engneerng Unversty

More information

4 Analysis of Variance (ANOVA) 5 ANOVA. 5.1 Introduction. 5.2 Fixed Effects ANOVA

4 Analysis of Variance (ANOVA) 5 ANOVA. 5.1 Introduction. 5.2 Fixed Effects ANOVA 4 Analyss of Varance (ANOVA) 5 ANOVA 51 Introducton ANOVA ANOVA s a way to estmate and test the means of multple populatons We wll start wth one-way ANOVA If the populatons ncluded n the study are selected

More information

Appendix B. The Finite Difference Scheme

Appendix B. The Finite Difference Scheme 140 APPENDIXES Appendx B. The Fnte Dfference Scheme In ths appendx we present numercal technques whch are used to approxmate solutons of system 3.1 3.3. A comprehensve treatment of theoretcal and mplementaton

More information

ELASTIC WAVE PROPAGATION IN A CONTINUOUS MEDIUM

ELASTIC WAVE PROPAGATION IN A CONTINUOUS MEDIUM ELASTIC WAVE PROPAGATION IN A CONTINUOUS MEDIUM An elastc wave s a deformaton of the body that travels throughout the body n all drectons. We can examne the deformaton over a perod of tme by fxng our look

More information

Portfolios with Trading Constraints and Payout Restrictions

Portfolios with Trading Constraints and Payout Restrictions Portfolos wth Tradng Constrants and Payout Restrctons John R. Brge Northwestern Unversty (ont wor wth Chrs Donohue Xaodong Xu and Gongyun Zhao) 1 General Problem (Very) long-term nvestor (eample: unversty

More information

Min Cut, Fast Cut, Polynomial Identities

Min Cut, Fast Cut, Polynomial Identities Randomzed Algorthms, Summer 016 Mn Cut, Fast Cut, Polynomal Identtes Instructor: Thomas Kesselhem and Kurt Mehlhorn 1 Mn Cuts n Graphs Lecture (5 pages) Throughout ths secton, G = (V, E) s a mult-graph.

More information

JAB Chain. Long-tail claims development. ASTIN - September 2005 B.Verdier A. Klinger

JAB Chain. Long-tail claims development. ASTIN - September 2005 B.Verdier A. Klinger JAB Chan Long-tal clams development ASTIN - September 2005 B.Verder A. Klnger Outlne Chan Ladder : comments A frst soluton: Munch Chan Ladder JAB Chan Chan Ladder: Comments Black lne: average pad to ncurred

More information

Pricing Problems under the Nested Logit Model with a Quality Consistency Constraint

Pricing Problems under the Nested Logit Model with a Quality Consistency Constraint Prcng Problems under the Nested Logt Model wth a Qualty Consstency Constrant James M. Davs, Huseyn Topaloglu, Davd P. Wllamson 1 Aprl 28, 2015 Abstract We consder prcng problems when customers choose among

More information

Chapter 3 Describing Data Using Numerical Measures

Chapter 3 Describing Data Using Numerical Measures Chapter 3 Student Lecture Notes 3-1 Chapter 3 Descrbng Data Usng Numercal Measures Fall 2006 Fundamentals of Busness Statstcs 1 Chapter Goals To establsh the usefulness of summary measures of data. The

More information

Real-Time Systems. Multiprocessor scheduling. Multiprocessor scheduling. Multiprocessor scheduling

Real-Time Systems. Multiprocessor scheduling. Multiprocessor scheduling. Multiprocessor scheduling Real-Tme Systems Multprocessor schedulng Specfcaton Implementaton Verfcaton Multprocessor schedulng -- -- Global schedulng How are tasks assgned to processors? Statc assgnment The processor(s) used for

More information

Ensemble Methods: Boosting

Ensemble Methods: Boosting Ensemble Methods: Boostng Ncholas Ruozz Unversty of Texas at Dallas Based on the sldes of Vbhav Gogate and Rob Schapre Last Tme Varance reducton va baggng Generate new tranng data sets by samplng wth replacement

More information

Example: (13320, 22140) =? Solution #1: The divisors of are 1, 2, 3, 4, 5, 6, 9, 10, 12, 15, 18, 20, 27, 30, 36, 41,

Example: (13320, 22140) =? Solution #1: The divisors of are 1, 2, 3, 4, 5, 6, 9, 10, 12, 15, 18, 20, 27, 30, 36, 41, The greatest common dvsor of two ntegers a and b (not both zero) s the largest nteger whch s a common factor of both a and b. We denote ths number by gcd(a, b), or smply (a, b) when there s no confuson

More information

Difference Equations

Difference Equations Dfference Equatons c Jan Vrbk 1 Bascs Suppose a sequence of numbers, say a 0,a 1,a,a 3,... s defned by a certan general relatonshp between, say, three consecutve values of the sequence, e.g. a + +3a +1

More information

University of California, Davis Date: June 22, 2009 Department of Agricultural and Resource Economics. PRELIMINARY EXAMINATION FOR THE Ph.D.

University of California, Davis Date: June 22, 2009 Department of Agricultural and Resource Economics. PRELIMINARY EXAMINATION FOR THE Ph.D. Unversty of Calforna, Davs Date: June 22, 29 Department of Agrcultural and Resource Economcs Department of Economcs Tme: 5 hours Mcroeconomcs Readng Tme: 2 mnutes PRELIMIARY EXAMIATIO FOR THE Ph.D. DEGREE

More information

Capacity Constraints Across Nests in Assortment Optimization Under the Nested Logit Model

Capacity Constraints Across Nests in Assortment Optimization Under the Nested Logit Model Capacty Constrants Across Nests n Assortment Optmzaton Under the Nested Logt Model Jacob B. Feldman School of Operatons Research and Informaton Engneerng, Cornell Unversty, Ithaca, New York 14853, USA

More information

Lecture Space-Bounded Derandomization

Lecture Space-Bounded Derandomization Notes on Complexty Theory Last updated: October, 2008 Jonathan Katz Lecture Space-Bounded Derandomzaton 1 Space-Bounded Derandomzaton We now dscuss derandomzaton of space-bounded algorthms. Here non-trval

More information

The Feynman path integral

The Feynman path integral The Feynman path ntegral Aprl 3, 205 Hesenberg and Schrödnger pctures The Schrödnger wave functon places the tme dependence of a physcal system n the state, ψ, t, where the state s a vector n Hlbert space

More information

Inductance Calculation for Conductors of Arbitrary Shape

Inductance Calculation for Conductors of Arbitrary Shape CRYO/02/028 Aprl 5, 2002 Inductance Calculaton for Conductors of Arbtrary Shape L. Bottura Dstrbuton: Internal Summary In ths note we descrbe a method for the numercal calculaton of nductances among conductors

More information

1 The Mistake Bound Model

1 The Mistake Bound Model 5-850: Advanced Algorthms CMU, Sprng 07 Lecture #: Onlne Learnng and Multplcatve Weghts February 7, 07 Lecturer: Anupam Gupta Scrbe: Bryan Lee,Albert Gu, Eugene Cho he Mstake Bound Model Suppose there

More information