LIBRARY OF THE MASSACHUSETTS INSTITUTE OF TECHNOLOGY mmsm A Model RETAIL OUTLET LOCATION: of the Distribution Netv;ork Aggregate Performance 642-73 Philippe A. Naert*, and Alain January, 1973 V. Bultez** MASS. tN'ST. -\fr.\C[ MAR 121973 DtWEY LIBRARY A Model RETAIL OUTLET LOCATION: of the Distribution Netv;ork Aggregate Performance 642-73 Philippe A. Naert*, and Alain V. Bultez** January, 1973 * Assistant Professor of Management Science, Sloan School of Manageirent, M.I.T. ** Charge de cours suppleant, Faculte Uni versi taire Catholique de Mons (F.U.CA.M.) IllC UUUiiUlO HiJii the data base. UU uitUiiiS iiwiv »J«-*i V. \j * t_MiMi^iti r»..w [--tx^V »'-'-«-. w..^ They are also indebted to the National Bureau of Economic Research (N.B.E.R.) and the Sloan School of Management for computer funds allocated. i)^- RECFIVFD JUN ' M. I. T. 19 1973 LIBRARIES 1 Locatjon of sales outlets Is of major concnrn to such orp, rTnl>!ations oil as companies, banks, etc. The problem ten approached in two steps. p, The total r.^irkct is divided into expand or to contract. Thus for example in period (i) 1 add '.•;ill 2 t outlets in area n. of- r<.>- stape amounts to decidinp in which repions to and the first ions, Ir, area 2, in n^ 1, the company t, etc. We will i 1 refer to the first step as the ap pre pate location prohlern. The second stape consists in chooLlnp specific sites for these new cut- these decisions are lets. In practice, dent, especially because they are orpani ;^a t i on . macle rep, arded as rather in da pen- at different levels in the The number of new outlets is a corporate decision, whereas specific sites are selected at the repional level, subject however to approval by the corporate headnucrtsrs. Many companies feel that, at least for the time beinp, aggregate and detailed this paper our sole crncern will be with aggi'cgs- location. Our procedure will be closely related loped by Hartung and Fisher ness, of neither worth the effort ncr the . In te linFing of the 1 ^ cost is problerris hiierarchical and suffered from the model parameters. a f/']. Hov-jcver, t'j a model deve- their work lacked robust- variety of deficiencies :n the estimation In section 2 ^.'e will review the Hartunp- Fisher [hereafter M-F] model and the various weaknesses associated v^/ith it. section In 1 vie section 3 the estimation problems wjll be examined. will propose various changes to the r'^dsl whicti 637560 v.' In ill I maKc it robust, a ond we will use data from a major oil corpany in European country to est i ma to the parr.mGtcrr. and validate cur ap- proach . 2 ^ • THE HARTDNH-FISHrR rODFL The market is to reductrd quasi duopoly, a we consider our brand.l, versus competitive brands taken c. Buyinr behavior is described as X. = probability that in period t-1, person a will usual top, ether, first order Markov chain. a transition prnbabilit, ins are defined as follows The that is, i-;ho 7- u : usual I y buys brand buy brand i i in period t. Adding the word "usually" broadens the definition used by H-F because it alio v; 3 for incidental purchases of -This important for usually buys brand v;ho on is a a a competitive brand. product such as gasoline. Take 1. On a given day he is running thruway and tanks at the next s'ervice area. not i, If a person cLit of gas the brand is we should not conclude yet that brand switching has occurred The other transition probabilities are similarly defined : probability of switching from probability of remaining a to buyer of pro ability of switching from [j c i to i a c. competitive bran r , Market share of bronri t Is m=x.*m_+o.-m i,t-1 i.t c,t-1 1) 1 In in period 1 1 steady state, rn i m (2) i,ei = a / . ,t (1 i - X 11 + . o . and thus , .t-1 = ) c^. 1 / fo ic The transition probabilities the decision variables of company i. .examp 1 a v-/ill ) be functions of such as advertising, cxpendi- e . = i s. (a., 1 p., I'^ic d., i a , d cc , p H-F consider only the number of sales outlets as determinants of the (3) transition probabilities. The following functions ^i = 1^1 • d^ / (d^ * d / [d + . 1 c ^i^ d . 1 ) v;ere post u la- 5. For given value of a d both , and show decreasing returns. and a. X The increasing with ri, are limits of X. 11 are a. and k respectively. However, K in d c the and K and hence , d <» ->• and X. restricted to he betv^een zero and one. For example, H-F are not larger than one. and K for and of o. i Since is X probability, its dependence on a should be constrained in such interval. [0,1] "Unless The H-F model aware of that. H-F are well k = 1.0, = k values of d./(d 1 + c d,]. a d. way that its value will lie therefore loc^s robustness. They state that, X. and o, Ho v; ever, are not probabilities for all equations and '3) can be (4} i ted. More general functions can be substituted for equations (4) (3) and without invalidating later results. The authors found that 3 for their problem equations Section 4 v-jill be (4] were sufficiently accurate". devoted to examining ways through v-. ch the model can be mads robust. Let us now relate market share to the parameters k and k. Substituting {3} and (4] for X and o in ['^) gives 6. (5) i.e Lpt q - C1 sales of and i, Note that H-F replac::' industry sales. Market share = Q m^ in I.e m. q./ = We should observr; that by-^1 a./O. (51 this implies the assumption that observed market share is equal to steady state market share. will return to this issue in ssctiori V-'e terms, H-F obtain. q^/d^ (6] From (65 = k^ Q • (d^ / follows that for it piven value of a d '- , c and average sal as per outlet will increase with ri, and ty when p = (d^ -2.a0, + anc) y«d.) goes therefore, per outlet when d./d ^ 1 = c to ^ero. With = k f. v/ithi p < D, oes to infini- 4.44 and = k 0.64, the model>woulri predict infinite sales For values 1/2.8. d . 1 /d > c 1/2.8, the mo- The function pattern is depicted del would predict negative sales. inFigurol. Insert Fioure cbD'jt here 1 On the other hand pf 1 -^'itlet l.ambin will decrease when applied u'-" i tlic H-F model on a f p > nufitaer , of the model average sal^' outlets incrftastrG. brand of p;asoline in a European • 7. cruntrv [10, chanter 7^, and founci values of and 0.9 5 7. y = However, even with difficulties. With d. (for larp,c given a k one may still 0, > p d }, run into approaches q. k* Q/u sales exceed industry sales. The next rameters k issue deals with the estimation of the pa- and k„. Equation (B) is nonlinear In K k, and could not be estimated by linear regression. H-F applied of trans for rr.ations to [6) and ended up with the folio in i-j and a series rr function (7) H-F use linear regression to estimate a and k_ are then uniquely determined from and a B. and The paramete-s B. probleins absociated witii the Gj.LimdLion pioceduie These will be examined in section in an 1^ There usyd arf ij y H-F. 3. aggregate retail outlet location model in to maximize discounted return of firm (i). i^jhich k various the ob, . Let J = the number of regions (J=1,...,J1 T = the tine horizon (t-0....,T] T = the disco nt factor r;: = return per unit sold in ri'T LI q. , d. = ... , d., as existing r. rep, ion J, period but superscripted for regions nnd time q. upher of outlets at time D Jt and in period t Jt .Jt The return maximization model is then Max Z I J t .f ^i' .f subject to jt I it c~ J ' qf = K^ df = dJ° ^ jt - i Jt 'i n^ < ^ " b ' qJ^ . . I / [dJ^ nJ^ t . ..dfl Vj.t Vj.t Vj.t 9. The readRr wil] increasing '•it function of d specified as costs, i.e. notice that with . < p and therefore unle^js D, it L Z j t n^ c. \i^ * ^J the objective func- are deduced, , i. 1 constraints. It p, i\'en by also important to note th^t even when the is discounted coiistruction costs ore deduced, may still be monotone increoslnp, 3. carefully is r;: an is net return and unless the discounted construction a tion is monotcjne increasing and the solution is rrerely the n-j the objective function for d^ , |d < / \i\ • ISSUES IN ESTIMATING THE H-F MfinEL timated values of rif;;ht a B in equation (7) hand side of the equation while in fact the number of retail outletsis expected left and to have hand side of equation a causal effect on market share and not (7) is q./'J = 1 tion (7) can be rewritten From have ri./d 1 on c a causality point of + d c /d.}. Or. equ 1 on the viev^/ I'^ft of we would like to the equation. Scne 1 simple manipulations of equation equation, 1 as the ripht and m. "^ (q,/Q)(1 (6) result in the followinp, lineai 10, 1/m.=Y*<5'(cl/ri.) c (9) 1 where y (1 = the market In equation 1 and P)/nc, - 6 share function, (5) m, (i-/:th ned fron equations (7), Or in ternG 1/ci. = K, and k K= 1/ can estimate it directly as nbtainer; wt.' m. r-eplacinp, and (9], of } hv nonlinear er. tinatjnn (5K The estimations were pcrforme 5 on the TROLL systpn on r^arquardt's paramet ers [12 The estimation proceriurG in TROLL , is br-seri alporithni for least-squares estimation of nonlinear 1. Wg estimated eauations brand of gasoline in a (7), (91, and Furopcan coun'try- There arc for (5) 35 a major qu^Tterly observations, from the first quarter of 1962 to the third quarter of 1970. on its The statistical results for the nonlinear model linearized form around the ontimu'n, i.e. are based the minimum of the 6 residual sum of squares . Let Pt The general model is written as. = (0, ... C ... 0„) ^ vnctcr of Let e bG the final InrDst-squnr'RS er- tirpatG of Q close to Vlith , 3 9f (y, ,ei (in) E(\' f(y^,o] = } \—-~-A T p 3f with f - p 1 I - ^^ e=G ^ ^ (X^.O) pt f^ f = (x^ .0) - f j: t t , pt p p can be v.Titten as (10) jqufltion t wherB, z pt P f. - V f • following matrix X P) for the nonlinear model are The statistics a (T linear function of the parametBrs t similar to those of - a lineai- Thus, rep;ross'ion. if we definr; the : P1 12 /here T is the number of observations ^ariance-covariancc n>atrix of is ?? tbe residual mean sqLiare is V [ available, 0) = s -n 2-1 • i f-, (F'F) , the e s 1 3 in v-zhich ra t 7 : e r 1?. Note that a^ is n no ir, Icn^Gr an unhiased c.i timet of n thR disturbance variance and that even when the en- or tf;rm , normally diGtributnd, result, valid in they the u'. nal t-, therefore he The share of brand : rf; to Worse even, k is nep, a statistics are not these statistics will be reported here rep, arded as mere suits for tirand less riod of observations, f: normally dlntrihut. ed. As Innp, er and Durbin -Watson- F-, Howpver, psner.-^l. shoLilri no is than risen valur^s. are presented per cent. 5 its outlet otive. i com. pa Yet, in Taiile I. ovpr the i%'hole pe- share was hip her than 8 per cent. The values of the '"'urbin-VJatson sta- tistic indicates autncon- elation of the residuals, and hence problems with the model specification. There are several possifile explanations. One is that im, bles may have been left out. port ant additional explanatory variaA previous study of these data shows 8 that this is not the case . More liRely reasons are the misspecifir tion of the transition probability functions, and the assumption that observed market shares are enuilibrium values. In other cases one may have better luck. For exam9 pie, VIP. applied the H-F model on another brand (a) , and found sults similar to those in the H-F paper as illustrated in Table The Durbin- V/atson statistic, 10 autocorrelation ho v; over, ap, ain indicates reII significant Insert Tables In the next which thR model can formulations 4 , to \r I ?.nd aiiojt il s c c t i o n ' here we w j 1 1 rxp 1 n rb made robust. Wo will apply c t u' both w o ri vs in these rridrl the data for brand i. ALT E;_P h!_AT TVF FO RniJt -VrrO^N S In this section we will redefine the transition oxponential functions prohahility functions. First, as relative number nf outlets in section 4.1., funct i ons j section 4.2. n How to rn and next as of the lopistic ake use of these models will he further examined in section 4,3. 4.1. Exponential model Define A as . i d./d approaches infinity, A. approaches one. Fnuation (1?) thus relates the relative number of outlets D., bility X . in , a Similarly, (13} = \ 1 c 0. - the transition protia- to robust way. EXP(-a • c D defined as is X ] c And therefore. o = c EXPC-a. 'D.) 1 1 , a _, i = EXPf-a c 14 TarKet share at at tirrie related to market share ic t . timp t-1. (14) rn = i.t . . [1 E - XP ( a - n • . 1 EXP(- + J D • c i.t . , E - ] XP [ - a • P c , c.t 1 ) m • , , i,t-1 . ^ c.t Assume now for a moment, as H-F did, that ohsorvcd markot share values pre Gquilibriurt values, for piven values oT D. . I.t (15) and D ^ c.t . 1. ^ i.t.e That is, EXP{-a = • c Applying a linear mod e (1G) 1 D ^] c.t / fEXP(-a, • D. ,) a logjt transformation to - m. EXP(-a c + I.t i D • ^}1 c.t results in (15) . l0R[m. ^ '^'x.t.G / (1 )] i.t.e-* . a ,• = i D. , I.t a n • c . c.t The results of the estimation of ted in Table III. The estimated a. is negative which would seum to indicate that the assumption on enuilibrium It is not at should be c]Gar that many assumptions equilitjriurn form For ex a n^ pie, (16). - the ccrsumption patterns arc adequately stable. b) - the unit-tirnj period is ouffiniently of a not tf- e provided that a) -J - have to be made in order to accept satisfied in this case - tive effect all Wr-irrcjntec marketing campaign long so that the disrup- launched in period [t-1,t] - on the B teady- st nt <; narknt shares can resorh In new equilibrJum achieved within thq same period a c) - ft-l.tl, the unit-tirnc period is short ennuph so that no compntitive reaction can interfere witfi retain equation (15). we may The this new ccuilihrium, restrictive and somewhat contra- dictory chnr-acter of this no n -exhaustive ret of assumptions cxplciin why vm should turn to dynamic forms. This does not imply that we have to disrecnrd the steady-state aspects w/hen we are about to make decisions on where to add new outlets. The results in Tahle III merely indicate the market dynamics should be taken into ac- count in estimatinp, the parameters. Insert Table Equation (14) I I I about her3 is .estimated in two different ways. we applied the Senuentia] 1 (SUMT) intrinsically nonlinear, and was First, we used TF^DLL. Secondly, Unconstrained MinJmizotinn Technique 2 . rjonlincar programming can bn used for nonlinear estima- tion in the following way. Let the model be. r-i is i n i mi z : achieved by solving the nonlinear programming problem belcw (17) min i:J^![ cj(e) t = 1 , - ;rt TV Tnhl'.' >quation (Ih] usinj; Table IV and stiov-.'S Fi twe re5u]t;; of the TROLL and SU^iT. Figure 2 n;; 1 1 illu?. ma t i tratcs on of and \. X c 1 VJith A tht: currRnt number of outlets, . ThGSG valuR'5 are very dieted equilibrium market share in d./Cdj + ri c i 1 ], and assuminr in share would be share hip, h, is ri c ' one per cent a needed to incr-easc outlet one percent. 5 look at the problem regional basis. For this particular product, we have informa- tion on four different regions. from unit worth v-jile depends on industry sales volume, More interesting of course is to on increase remains constant, equilibrium mar- and the number of new outlets tjy A . which is to be Rxpectcd per cent. Whether such an increase Ket share would increase by profit, D^ = a high of about per cent in ID sho'.'vn Adding outlets in region It differ r ^ c ros s per cent in one region to another region. per outlet added are in a about low of Incremental year'ly regional sales Table V for each of the four regions. contributes thehighest marginal return. quite possible that the response parameters is regions. gulshed from met rope The company's outlet share varies ] i For example, ru r- a 1 areas might be d i :; t In our particular application t an areas. i n in- 17. formation by region wos .. vailable only on an annual basis over period of six years. This was insufficient for the purpose of a es- timating responsR coefficients by re ion. f, Insert Table V about h^re 4.2. L^O£ist_i c__modRj Market share as a function of relative number of outlets is often thouc^t of as having an oil in companies, for example, S shaped form. Various have been able to observe such curvsHS S plots of sales or market share as functions of the number or 1 3 the relative number of outlets The theoretical arguments in favor of such ped relationship at the market share the form of equation bility level. Thus, [15) at - level - a S sha- already introduced in may also hold at the transition proba- one extreme, if the oil company has too few filling-stations consumers will notice them too infrequently and will often be obliged to tank up at other companies stations; es a consequence, their loyalty will be very low. As the number of stations increases consumers will be able to tank up at the company petrol pumps locatod in various geof.r aphi c a 1 areas and their loyalt will be enhanced accordingly. At the other extreme, continues to extend its distribution network, tructed wil] of i,::"r to if the company each new station conr attract consumers from the remaining hard cn:- competitors customers. Estimatinc thg pararr.ctors of equation (20) If thR hir. toricel data come from shares and markrt v/ill raise on stable marknt, a i.e. isr-ue. p. our cutlet relatively little variatjility. F^hares'-. hovv will be severe nulticnllinee'rity probl^uos. As mnritioned in 3, a the independent vr. first orrinr Taylor ted at the thei-e sp':;tion riahles in the linf;ar equation derived t>/:pansiori current solution. first are Let order derivatives evalua- current solution be the frorr. o a. = 1 3m [ a . = tween ^.t-1 b ^^i.t^^i=^l . = b 1 /3a. and 3m An / i.t . 1 state, i. ./3b, important issue in nonlinear estimation din^ good initial values. used the follov/inj' procedure. VJe is In fin- stead the elasticity of market share with respect to the relative number of outlets is. (24) °i,t 'i.t-1 8rn. , 1 a. (22) (23) a. A . 1 • m c , t , e t 13. For thpsp rnasons, a curve relating, tran- slu^peci S sition probabilitlDs to the ralative num'npr of retrtil outlets '^poms -D. 11 (18) 11 /(a. +n. A. 3 approaches onr, according to Similarly b (19) with \ a - C and b ^ D defined as is X / [o C CO * D ^ positive constant b shaped pattern S a ciii r, iccc Market share at %.t -' ff"i,t I. i If D state market share (21) 1, ^ i.t.e a. tirr^e =a/(a is t +D , °c.t • . t- . 1 + ^i ^ ^^ - [a c / (a remains equal to v/oulri ). then ^c bi ^^°^ are b^ ), =a./(a. +D. defined as pmb abilities The s^'^itchinp, . b. ' '-\ + D c D. c. . ^) ^c °i.t^-^% ' ^c.t ^ 1 and to D . steady be, =a(a. ci +D.i,t^)/[a.(a ic +D c,t^} + a(a. ci +D.i.t.) To estimate initial values for the parameters, we assumed a value value for m in a = c 0.62S and b ?.46. = c With these initio Divergence optimal solution. values TROLL failcid to find an 1 ocf. urrori initial values which we tried. for these and for all other The SUMT program performed hetter. The main reason is pro ably that the method for minimizinp, the ti unconstrained penalty function second order procedure, the Nowton-Rephson method, is whereas f^larouard 1 t ' ' only uses first order derivatives . Furthermore, all our SUMT "computations were done in double precision. ly This may be pai-ticular- important in view of the factthat there is rity problem. The SUf-IT tistics ere very poor, estimates as have the correct sign, all all, are> shov;n in a and the magnitudes real multicollinea- Table VI. The expected. Weverthelcss, the estimated values are a estimation method s t sta- the coefficients are reasonable. After not too different from our initial subjective estimates. -t to an Table actual market share of sales for adding a V, about . Vc'^le V/II shows the incremental now outJet in each of the r8i:;lens. 21 sert Tabl- VI 4.3. _no_do_l_ £P£lj_f^a.t les lop, ab?ut her i Oj2 TflblRS tinl and I \1 and VIT rpspcctively how the shov-; istic morials can be applied to compare in crp mental sa- from invGstmRnt in outlets in various regions. The ultimate^ measure of performance is the 1 profits rather son of rofional may differ from cne costs, and cost of rep, tl-ian sales ion to another. """^^ • • ^i 16* ced by q constraint in ri"'' • ^ q cost structure The a i: trans [lortation cross regions. proposed by H-F could (R] = the objertive k^-Q-^^ '^ t . For example, transition [probability functions be adjusted for the exponential section 4.1. We replace nompari- 5 land will vary purchasing, The optimization model 0''^ e>;ponf?n- / (d^^ + y o- function by ' ^'[^ ] is repla- : c.t c i , _j c c.t i,t-1 t or if cur interest is only in steady state results without conceri for the transient behavior. J.n- -aJ.D^ , I'^th in the optimiz'tlon, ''i the "i.t loeistic model, complications would ai-i;:t because the transition probability functions 77 convex for some ronrc of are D. [or D ), c 1 and concave clEewhRrn. NeverthG]o3S the model will remain useful. Instead find the optima] of trying' to be 5. allocation over time, the model could applied to civaluate various outlet expansion plans. CONCLUSION In of the ag^eref, and Fisher. this paper we have presented ate retail V.'e a detailed study outlet location model deve]opcd hy Hartung found that their E^'^eral procedure was sound, but that the specifics suffered from the estimation of thij a variety of problems rep, ordinp, model parameters, and the i-obustness of the response functions. Rem.edial action was proposed, and was applied to a brand of gasoline in a European country. .^^ c c C •rH Q 3 1 c c O •H JD 1.1 a 3 25 TABLE III Steady State Estimation of Exponential rodel for brand i Coefficient Value - t statistic 26.241 3.25 . 700 F(2 >33) = 3i DW = 0.49 26, TABLE IV Exponential Modnl for Brand i Coef f Icinnt F{? .32) Statistic 59. OG 2. 75 .70 68 2.76 70 69 SUMT 0. 73 7 . 63 37.3 ?'3 TABLE SUM Estimates VI for Logistic ModGl CoRfficient Value t statisti .019 8394 2.663 .061 . 30 FOOTNOTES Based on discussions with various major oil companies. The cGnciir;t linked hierarchical models has hpen examined hy CrL.vJstcn and of Scott-Horton \ 1 ] and , has been applied in the area of operatior-ril planning and control by Green FB], New son [15], and Shwimer A possible application in the public sector fias T'^Bl. been proposer' hy Hausman and Naert [6]. Our notation will differ from H-F. See tl-F tation At \ 7 , p. R 234, footnote ?]. The quote is given in our nn- . least within the limits discussed above. TROLL is an interactive estimation and simulation p-jckage develop 'ic by the Computer Research Center, National Bureau of Economic Re- search, For a Inc. For a description of the estimation riethod see [3]* further discussicn on this method, to Draper and Smith f?, [5. ''' pp. pp. 267 et sq.l '.-.'o may refer for cxa'^'.pl" and GOIDFELD and QUA^DT 49-57] Provided that the linearized form (11) of the model is valid 6, the final estimate of 6. test of hypothesis conducted hy LAhLilN ^ A ^ For brand a, ^ At observations 38 v.' Fill. ere available. this stof^c an additional qualification ought to be made about estimation procedure. The [l-F r. imilarity of the results oritainsd from regression analysis apnlied on equations (fi) and as (5) oppo- sed to the relative dissimilarity observed betvveon those derived from equations tion [ 7 ) (7) contains and (5) m. a , evidenced by equation of and a 6 should cause no surprise, stochastic variable, on both sides (fi)]. fie from small samples [as nee we sfiould expect biased estimates ; furtherrnorc; this in inccnslstent since the residuals indicate are since equa- a case a and 6 strong autocorrela- tion among the error terms. All in estimatir'g K. this adds up to pointing out the bias created and k„ via equation (7). Notice that no error, term was introduced in equatior (5], had one be added the linearization of (5) would have been impos- sible. ^ For additional discussion on Naert and Bultez [14 ^ For [41. a li. nearizinc, such nonlinear models, 1. theoretical exposition of SUMT, see Fiacco and Mc CormicK The computer program is described in MylancJer et al. [13]. set 32 ' ^ Many factors making the S are reported by Kotler [9, ^ '* shaped curve plausible in this industry pp. See Wilde and Bcightler fl7, 96-971. pp. 22-24] for a discussion of the Newton-Raphson method. The SUMT routine has various options with regard to the technique for minimizing the penalty function. One of these is the Newton-Raphson method. ^^ Assuming that orofit maximization is the objective. ^^ Different types of regions, e.g. metropolitan, suburban, rural, have dif f ei by regions m.igh . 33. REFFRENCFS Crowston, W.O., and r.S. [1] S co 1 1 - P^nr Decision S^'stnms", £r ncer;c1inp;B of Fall of- TIP'S. John [31 N.R., nraper, \2] V'.h" 1 ey an 7 Smith, and H. ri t n "Tdc , 11th Dor,if',n of n t rrna1 1 Hi o rarch, i ca ona 1 M pr Repression Analyair. . New York he I 1 t i n£_ . Sons and P. M,, Eisnoi', as 19 to 1 , 9 B G /^'PP^J-Si' . Pindyck, Generalized Approach to Estimation "A Implemented in the TROLL/I System", Working Paper, National Bureau of Economic Research, Computer Research Center for Economics and Panapement Science, [4 Fiacco, 1 ti a 1 A . V . llnconst and G.P. , r aineri P^in P-'c PI arch 1972. CormicK, Pionlinear Programmin g imization Techniques . Mew York : : Soo uon - John Wiley _ and Sons [51 , - Ho 1 1 Plethoris and Publishing Company, i n Econom etrics 1972. "heuristic Coupling of Aggregate and Detailed riodels Factory Scheduling", School 1 PJort h : Green. R.S., in [7 93G Goldfeld, S.P1., and R.E. Ouandt, PJonlinea r Amsterdam [PI 1 of r.anagement, Unpublished Ph. Pi.I.T., Hartung, P.W.. and J.L. P. Dissertation, Sloan 1971. Fisher. "Brand Switching and r at h em Programming in Plarket Expansion". Management Scien ce, vol. gust 1965, pp. 15-^9. t t i r. 2, Au j 34. [ e] Hausman, W.H., and P. A. "Optimal RuGOurce AllocaMon in WaGrt. Family Planning ScrvicGS DelivRry", Unpublished Working Paper, Alfred [9l Kotlnr Ph., New York Fin] Lambin, Uni ve rs [11T Sloan School of ManaRGment, F^.I.T., August 1972. P. arKe t i n ;;. Decision Making Holt. Rinchart and Wistnn, : J.J., r^odeles et Pro Rramns i t ai res Lambin, J.J., of Business [I2l fl , de "Is vol. Marquardt, D., s rie 1971. Harketinp. nylander, 45, n° 4, sum -Version 4 11, R.L. : Presses : n° T he Jo urnal October 1972. : W.C. Paris Gasoline Advertising Justified ?", linear Parameters", Jou rnal of the Society [131 , France, 1970. "An Algorithm for Le as t -Sq are Mathemat ics, vol. Bu il dinp, A pproach, riod el ;_ A of Estimation of s In dur, trial _5 iJon- n d _A_p p l_i n 2, June 1963, pp. 431-4 41. Holmes, and G.P. Mc Cormick. "A Guid^- to The Computer Program Implementing the Sc:qucntial Unconstrained Minimization Technique for Nonlinear Programming", Paper RAC-P-63, Research Analysis Corporation, October 1971. [141 Naert. P. A., and A.V. Bultez, "Logically Consistent r^arket Share Models", Working Paper n° 607-77, Alfred gement, M.I.T., June 1972 (forthcoming) . ) P. also in Journa l Sloan School of Manao f _^ Ma rket ing Research [15] Newson, E.F.P., blished Ph. n [iFl . I . T . , 19 Shwimpr, linp, in 7 D. Dissertation, Alfred "Interaction Between Job Shop", P. Wilde. D.J., A p, p. Unpublished Ph. Sloan School of nanapomen t ri7l to Finite Capacity", UnpuSloan School of ranap;er:ent 1. J., a "Lot Size Scheriulinp and T.S. Englcwood Cliffs, N.J. , f'l . I . T . , rep ate and Hctaileri ScheduD. Dissertation, Alfred 1'J72. Beiphtler, Foundationr, of Optimization, : Prentice Hall, 19B7. P. mmm ^SiSfSii Date Due A 0G15'89 MAV 2 715 } I 0^ OCT SEP /\PR 3 •t1 28 '77 11198' ZOi^l 65^-72 / 3 ^Ofl ^^^-"^ a||iiii'|?|jp|»f||||| ; 3 ^laao dd3 7D2 stm 3 TDaO 003 b71 507 ^^^ o9.l3 llllli™ll'"11'f'"!!fl!llil 3 TDfiD DD3 b71 C^:>S Mfil 3 TOaO 003 b71 4M0 3 TO 60 003 b71i 41t> 6^? -73 3 TDfiO 003 70E Mbfl ^^-5-73 3 TOfiO 003 702 M7b