Scribd Logo
Upload

Welcome to Scribd!

  • Krarup 1983

    Uploaded by

    Aleksandar Micic
    12 views46 pages
    rad 83

    Original Title

    krarup1983

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    Report this Document
    rad 83

    Copyright:

    © All Rights Reserved
    Download as PDF or read online from Scribd
    Flag for inappropriate content
    12 views46 pages

    Krarup 1983

    Uploaded by

    Aleksandar Micic
    rad 83

    Copyright:

    © All Rights Reserved
    Download as PDF or read online from Scribd
    Flag for inappropriate content
    of 46
    36 The simple plant location problem: Survey and synthesis Jakob KRARUP DIKU, Instisute of Datalogy, Univeristy of Copenhagen, DK-2200 Copenhagen N, Denmark Peter Mark PRUZAN Institute of Economics, Unicersity of Copenhagen, DK-145S Copenhagen K, Denmark Received May 1981 Revised February 1982 With emphasis on the simple plant location problem (SPLP), we consider an important family of discrete, deterministic, single-criterion, NP-hard, and widely applicable optimization problems. The introductory discussion on problem formulation aspects is followed by the establishment of relationships between SPLP and set packing, set covering and set partitioning problems which all are ‘among those structures in integer programming having the most wide-spread applications. An extensive discourse on solution properties and computational techniques, spanning from early heuristics to the presumably most novel exact methods is then provided. Other subjects of concern include a subfamily of SPLP’s solvable in polynomial time, analyses of approximate algorithms, transformability of p-CENTER and p-MEDIAN to SPLP, and structural properties of the SPLP polytope. Along the way we attempt to synthesize these findings and relate them to other areas of integer programming. Introduction The past two decades have witnessed an explosive growth in the literature on location problems. This is not at all surprising since locational decisions is one of the more profitable areas of applied O.R. and ample theoretical challenges are offered. However, among the myriads of formulations considered, only four of these: p-CENTER, p-MEDIAN, SIMPLE PLANT LOCATION, and QUADRATIC ASSIGNMENT - at times referred to as prototype location problems — have played a particularly dominant role. If seminal works such as Fermat’s I-MEDIAN from the early 1600's and Sylvester's 1-CENTER brain teaser from 1857 are disregarded, all four problems entered the stage in their present form in the period 1957-64. In contrast to p-CENTER and p-MEDIAN treated at length in textbooks including Francis and White (1974), Christofides (1975), Jacobsen and Pruzan (1978), Handler and Mirchandani (1979) and in the survey by Krarup and Pruzan (1979), we have long sought in vain for a comprehensive exposition with particular focus on the simple plant location problem (SPLP). This is surprising since, judging from a rough estimate of the number of papers devoted to each prototype problem and from their applicability to real-life decision-making, SPLP seems to have attracted most attention. In chronological order, early reviews and shorter summaries of the state of the art can be found in Balinski and Spielberg (1969), ReVelle et al. (1970), Eilon et al. (1971), Hansen (1972), Elshafei and Haley (1974), Francis and White (1974), Kaufman (1975), Salkin (1975), Jacobsen (1977), Guignard and Spielberg (1977), Jacobsen and Pruzan (1978) and Cornuégjols (1978). ‘The authors are indebted to Egon Balas, Jens Clausen, Monique Guignard-Spielberg, the late Jonathan Halpern, Antoon Kolen, Kurt Spielberg, Philip Wolfe, and Stanislaw Walukiewice for valuable suggestions on the preliminary version of this paper. We also ‘wish to express our gratitude to the referees, James G, Morris and Richard M. Soland, who were so kind as to divulge their identities. ‘The final version has benefited from their meticulous screening and constructive crit ‘North-Holland Publishing Company European Journal of Operational Research 12 (1983) 36-81 0377-2217/83 /0000-0000/$03.00 © 1983 North-Holland J. Krarup, P.M. Pruzan / The simple plant location problem 2 The present exposition is an attempt to fill the gap. To give adequate credit to all previous contributors would be a Herculian task, far beyond our level of ambition; instead, we shall consider some of the main findings, spanning the SPLP-literature from what appears to be the first formulation of the problem in the carly 60's to the most recent developments, represented by a series of papers currently under publication. Along the way we shall attempt to synthesize these findings, relate them to other areas of general integer programming and suggests likely courses of future developments, While most well-defined problems bear unambiguous names, SPLP has been dealt with in the literature under a wide variety of different titles usually composed of an adjective (uncapacited, simple, optimal) and a substantive (plant, warehouse, facility, site) followed by the word location, Including omission of the adjective, we have among the possible combinations seen ten or so of these employed so far. And new contributions to the name confusion arise if ‘economies-of-scale’ or ‘fixed-cost discrete space location-alloc- ation’ and the like are incorporated, not to speak of the flock of variations, based upon word-to-word translation into English ‘of non-English designations. Even some familiarity with certain non-English languages does not always suffice for identifying a SPLP-paper written in one of these via its title. SPLP derives its name from the analogy to decision problems concerning the location of plans or facilities (e.g, factories, warehouses, schools) so as to minimize the total cost of serving clients (e.g. depots, retail outlets, students). We postulate that the major reason that SPLP has been the subject of so much attention is that despite (or perhaps due to) its transparent structure, it has contributed to the formulation and solution of a multitude of complex planning problems. Unfortunately, to the best of our knowledge there does not exist a readily available list of such applications which could lend credence to our postulate; we can though draw upon our own experience as consultants where we have utilized SPLP formulations as the basis for providing decision inputs to real-world problems regarding the number, size, design, location, and service patterns for such widely varied ‘plants’ as high-schools, hospitals, silos, slaughterhouses, electronic components, warehouses, as well as traditional production plants. In contrast to the other prototype location problems, SPLP permits in a sense the broadest framework. Neither the number of plants to be located nor the transportation or communication pattern are predetermined, Furthermore, the basic formulation of SPLF lends itslef readily to sensitivity analyses. In addition, SPLP invites modifications which may permit more ‘realistic’ modelling. While SPLP is basically a discrete, static, deterministic, one-product, fixed-plus-linear costs minimization problem formulation, it can be modified to accommodate dynamic, stochastic, multi-product, nonlinear cost minimization formulations, It can also include prices as well as costs in its criteria and can also be used within the context of multicriteria optimization. Last, but not least, there exists a veritable arsenal of well documented exact algorithms, heuristics, relaxations, and simulation procedures, more or less tailored to provide solutions to SPLP’s. It appears thus that the attention paid to SPLP is due to both its relevance for sundry decision problems, its simple, easily grasped structure, and the availability of effective solution methods. There is however the reservation that while p-CENTER and p-MEDIAN usually are dealt with in both continuous (planar) and discrete (network) formulations, there is a dearth of literature as to formulations of SPLP in the plane. We shall therefore be mainly concernced with the SPLP-family within the context of networks but briefly comment the very rare exceptions in Section 13. Our presentation of SPLP and its connections to close and more distant relatives is organized as follows: Alternative symbolic formulations of SPLP are introduced and their properties are briefly discussed in Section 1; furthermore, the blurred historical origin and some notational discrepancies are accounted for. The relationships between SPLP and set covering, set packing, and set partitioning problems are established in Section 2, emphasizing once more the applicability of SPLP, while Section 3 provides a rudiment of computational complexity to demonstrate that SPLP (as could be expected) belongs to the class of notorious optimization problems which is known as NP-hard. The five following sections deal with solution methods and computational aspects. Like other difficult combinatorial problems, it is not surprising that SPLP was only amenable to heuristics during the first years of its lifetime; a collection covering the period up to 1966 is reviewed in Section 5. Numerous exact methods have appeared since 1966, culminating with the dual-based algorithms, rooted in Lagrangean relaxation, and claimed to be ‘close to the ultimate in efficiency’. The extensive selection presented in Sections 6-8 demonstrates the 38 J. Krarup, P.M. Prazan / The simple plant location problem diversification of integer programming techniques adapted for SPLP over the years. We proceed in Section 9 with a class of specially structured SPLP’s solvable in polynomial time, and capable of modelling certain ‘classical’ production planning problems. Analyses of heuristics and relaxations for a composite formula- tion having both p-MEDIAN and SPLP as special cases were initiated by Cornuéjols et al. (1977a) under the heading ACCOUNT LOCATION and have since then enjoyed much fruitful research. Some of the main results are summarized in Section 10 together with an observation due to Krarup and Pruzan (1981b) stressing the need for further refinements of error analysis. Based on Krarup and Pruzan (1981c), where convex combinations of p-CENTER, p-MEDIAN, and SPLP are integrated in a single model, it is shown in Section 11 that p-CENTER and p-MEDIAN are transformable to SPLP. Structural properties of the SPLP polytope are considered in Section 12 where certain families of facets are identified. Section 13, a mixture between an appetizer and an annotated bibliography, deals with selected variants and extensions of SPLP; pertinent keywords are capacitated, piecewise linear, dynamic, stochastic, continuous, multicom- modity, multicriteria, investment and pricing. 1. Problem formulation Given a finite set of possible locations for establishing new facilities or redimensioning already existing facilities, SPLP deals with the supply of a single commodity (or standard productmix) from a subset of these to a set of clients with a prescribed demand for the commodity. Facilities are assumed to have unlimited capacity such that in principle any facility can satsify all demands. For given costs associated with the facilities and with the direct transportation routes from facilities to clients, we seek a minimum cost production/transportation plan (in terms of the number of facilities established, their locations, and the amount shipped from each facility to each client) satisfying all demands. The constituents of SPLP are: mi n The number of potential facilities indexed by i, 1€ = {1,...,m), ‘The number of clients indexed by j, EJ = (1,...57), The fixed cost of establishing facility i, P;: The per unit cost of operating facility i (including variable production and administrative costs etc.), b;: The number of units demanded by client j, 4,;: The transportation cost of shipping one unit from facility i to client j. It is customary to use the adjectives ‘open’ and ‘closed’ for designating the state of a facility. The cost of sending no units from a facility is zero (ie. the facility is closed) while any positive shipment from the ith facility incurs a fixed cost f (the facility is open) plus costs p,+f,, per unit produced at facility i and transported to the client j. We introduce the m+ mn variables : y= Lif facility # is open and 0 otherwise, Number of units produced at facility ¢ and shipped to client j. The full-blooded SPLP is the mixed-integer prograi min = 2 (ate) t Lhe a) ierje. ier 2 5b), Jed, (2) ier ky 3 5,20, ie, () je 54>, ieLjes, (4) wEO1}, ie. (5) The m restrictions (3) are devices to ensure that the total fixed cost for a facility is incurred whenever J. Krarip, P.M, Prucan / The simple plant location problem 9 positive shipments are made from it, The k’s are positive constants, not less than the maximal outflow from the corresponding facilities. If all p> 0, t,,2 0, no facility need ever ship more than the total amount demanded, and each k, may be replaced by 3)..,b,. Similarly, since under the assumption ( p, + f,;)=0 it will never pay to ship a larger number of units to a client than demanded, the inequalities (2) can be replaced by equations, Should negative (p,-+¢,,) occur for a given data instance, we can by means of appropriately selected constants perform a transformation so that the assumption above of nonnegativity holds: replace ( p, + t,;) by (p)+4,;) + a; all i, j where a = {a} is an n-vector of constants such that fa;|> max,e, {| p;+ 4,l|(pi + 1,,)<0}. This transformation will not affect the optimal solution to the given instance since the only effect will be that the objective function is increased by the constant term Bic). Similarly, should any f, be negative which, from a practical viewpoint, is rather absurd, it pays to open the ith facility irrespective of positive outflow or not. In such cases, f; is added to the objective function (which hereby is reduced by a constant) and f, is replaced in (1) by zero. Without loss of generality (and with the real-world applications in mind) we shall therefore assume all f, to be nonnegative. Note that the formulation remains unaltered if we modify the definitions of f, and b, by the words ‘per time period’. In this case the fixed costs will include per time period investment costs and fixed operating costs corrésponding to the minimal costs required to maintain an open facility. Given the cost structure for SPLP, the decision problem is only considered to involve the flows (the s,,’s) instead of both flows and facility design. The costs are assumed to adequately reflect the type, size, and location of the facilities and to only depend upon whether or not flow occurs from a facility. Were the costs also dependent upon the type and size of the facility to be located, the total costs could not be determined from the flow alone but would have to involve decision variables corresponding to the type and size of each facility as well, This subject is reconsidered in Section 13 where piecewise linear cost functions are discussed. For any given binary y-vector an optimal set of ,,’s is easily determined. Let P= {i|y,=1} be a nonempty subset of open facilities and let P(j)= {iE P|(p, +4) = min, p{p, + %j})- If for all j, 5, for some i € P(j) (:= means ‘is set equal to’ and ‘some’ indicates that ties are resolved arbitrarily) and s,j:=0 otherwise, we have evidently found an optimal production transportation plan with respect to the given set P simply by assigning each client to the ‘cheapest’ open facility. In this sense, the y’s can be claimed to constitute the essential variables of the problem, and for that reason some authors prefer to distinguish between the strategic variables (y,) and the tactical variables (s,,). Accordingly, the number of distinct solutions (in terms of the strategic variables) amounts to 2”— 1, the number of combinations of open and closed facilities, excluding the case where all facilities are closed. Note for a comparison that the total number of distinct assignments of clients to facilities (each client is served by, or assigned to exactly one facility) is m”. How trivial it may seem, the fact that we only have to consider solutions where every client is supplied by a single facility, plays a significant role for several of the computational methods to be presented. To stress its importance, we state it formall: Proposition 1. SPLP has the Single Assignment Property. Based on this observation, some simplification of the formulation almost suggests itself. Let c= 5p, +4,,) be the total cost for supplying all of client j’s demand from the ith facility, and let x,, denote the Sraction of client j’s demand supplied from that facility. As discussed previously, all ¢,, can be assumed nonnegative, For a given subset P CJ of open facilities, the total costs incurred for an optimal solution with respect to Pis 3jcyminje pt, + S:cp/; SPLP is then the problem of identifying a subset P minimizing the total costs. The following version may thus be referred to as: SPLP: Combinatorial formulation min { > ming, + F }- (6) pcr | je,ier ! (Ep 40 J. Krarup, P.M. Pruzan / The simple plant location problem Upon inclusion of the additional constraint | P| = p, (6) is transformed to a model which for 7=J and f=0 is the p-median problem in a network with vertex set J: p-MEDIAN: Combinatorial formulation min { > maine} pohitice (2, ier where ¢;, usually is interpreted as the (weighted) distance between vertices / and j. Both SPLP and p-MEDIAN are special cases of a composite model known as Account Location which plays a key role for the algorithmic analyses exposed in Section 10. For a given SPLP at least one facility has to be open in any feasible solution (y,.x); feasibility is assured by the condition [I,¢,(1 —};)=0. By introducing m?n auxiliary, binary variables u,,; defined for all i, KEL jE by oS te if (ey 0, i € J. In such cases it is important to be aware of the effect of replacing the mX n ‘disaggregated’ inequalities (9a) by the ‘aggregated’ set of the m inequalities (9b). Advantages and drawbacks are discussed when we focus on the computational aspects of SPLP. 1.1. The origins of SPLP: Who was first? The ‘who was first - question’ may seem of minor importance only, but the somewhat obscure history and the incorrect references made now and then by various authors justify comment. Disregarding problem formulations such as that in Baumol and Wolfe (1958) which upon appropriate redefinition of terminology can be interpreted as including SPLP as a special case, the first explicit formulation of SPLP is frequently attributed to Balinski (1966) whose expository article on integer programming includes the mixed-integer formulation (7)-(11) with the strong constraints (9a). The paper was presented at the The IBM Scientific Symposium on Combinatorial Problems in March 1964 but remained unpublished until 1966. After some search, a copy was finally made available to us by Kolen (1978) who drew our attention to an even earlier source, a Mathematica report by Balinski and Wolfe (1963) referred to in Balinski (1966), Wolfe was then approached for a copy; his reply (Wolfe (1980)) includes the following remarks: “The paper by Balinski and myself, ‘On Benders decomposition and a plant location problem’, has disappeared. We indeed wrote it. We do not remember too well what was in it. Several years ago Balinski, noticing that it was not in his files, asked me for a copy; I didn’t have it either. According to him, it also turned out not to be in Mathematica’s official archives”. However, SPLP’s are also dealt with in the pioneering papers by Kuehn and Hamburger (1963) and Manne (1964). Both are often cited and will deservedly be discussed in Section 5. The Journal of Farm Economics is not exactly a traditional forum for locationists and discrete optimizers; accordingly, apart from researchers in the field of agricultural management, only a very few authors have noticed the seminal paper by Stollsteimer (1963). Based on the unpublished Ph.D. thesis, Stollsteimer (1961), he proposes four models of which the second actually is a SPLP. In a historical context, it should not detract from his contribution that the solution method devised is mere complete enumeration, thus allowing any cost structure whatsoever. On the contrary, judging from the annotated bibliography Lea (1973), in spite of its obvious weaknesses, Stollsteimer’s work managed to generate a flock of Ph.D. theses and published papers on agricultural economics; see e.g. Ladd and Halvorson (1970), Chern and Polopolus (1970), and Warrack and Fletcher (1970), all discussed in some detail in the monograph, Elshafei and Haley (1974). Thus, Hansen et al. (1981) seem pretty close to the truth by owing credit to Balinski, Manne and Stollsteimer for the first formulation of SPLP. Is the full truth then that SPLP was formulated indepen- dently by Stollsteimer, Balinski (and possibly Wolfe), Kuehn and Hamburger, and Manne? The lack of cross references between their works bears evidence to this hypothesis. 1.2, A comment on the terminology and notation employed. SPLP can be viewed upon as a covering problem: by selecting an entry (i,j) corresponding to the transportation route from facility i to client j, we have in a sense covered the jth column by the ith row. However, it is customary in formulating mathematical programming problems that the right-hand side literally is on the right-hand side of the coefficient matrix; consequently, SPLP should rather be a problem of covering rows by means of columns, To conform to this commonly accepted rule, / and J should have been interchanged in the formulation. There seems however, to be a discrepancy between two different traditions: of all formulations of SPLP found in the literature, only a few are consistent with the notion of covering rows while the vast majority in principle apply the same notation as was adopted in (7)-(11). We felt tempted to join the first group (authors of future SPLP-papers are strongly urged to do so) but decided not to because of the confusion this would cause readers who will consult earlier works referred to here. To avoid perpetual redefinitions of the most common concepts employed, we assume wherever appropriate that the three groups of identifiers (1, i, 1, fs Pis Qis Mis Yad» Cay Size Sip» Xi) AN (Jj, My by) are a2 J. Krarup, P.M, Prasan / The simple plant location problem I: set of Js set of facilities clients racine f a \ "ition sovess tom) \ + fixed cost ; E \eyjs unde transp. cos unit prod. cost Se ere aaeF 1, all j. I* C1 is called a partition (exact cover) of J if equality holds for all j, ie. if Yj<)+a,,= 1. For all i, let y,=1 if row / is in the cover and 0 otherwise, and let, be the cost incurred for y, = 1. The (weighted) set covering problem is to find a cover of minimum cost min DAI i Oorl, allé Lay, JES, y, 7 or, in a more compact notation SET COVER min{ fp| yA =e, = 0 or 1, all} where ¢ is an n-vector of ones. Accordingly, the (weighted) set partitioning problem reads SET PARTITION min{ fy] yA =e,y,=0 or 1, all}. By reversing the inequality sign of the constraint set yA =e in SET COVER and by changing the objective from minimization to maximization, we arrive at the (weighted) set packing problem SET PACKING max{ fy| yA Se, y,=0 or 1, all i} J. Krarap, P.M. Pruzan / The simple plant location problem 43 The seven transformations T1-T7 relating SPLP to the three problems formulated above or converting SET PARTITION to either SET COVER or SET PACKING are visualized in Fig. 2. The transformations will later be summarized in Propositions 2-4 and illustrated by a numerical example. Fig. 2. Overview of transformations T1-T7. In all casés involving SPLP, we will demonstrate the transformations based upon SPLP-S (SPLP with strong constraints). We shall first show how SPLP-S via simple transformations TI and T2 can be brought to the form of SET PARTITION and SET PACKING respectively such that optimal solutions to these will also solve SPLP optimally. Let ¥;:=1—y,, all i, and let r,,, all i, j be nonnegative slack variables used to replace the disaggregated inequalities (9a) by equations. Now, SPLP-S becomes min XBapry B Lhe Daeyah alli, xy tH AM =1, alls, 7 Xytj20, alli,j, F,=Oorl, alli. Without affecting the optimal solution, we can explicitly require all x,, to be 0 or | since SPLP as noted in Proposition 1 has the Single Assignment Property.Thus, all 7,, will be either 0 or | as well. The SPLP =» SET PARTITION transformation follows directly from the definitions of j, and 7,,: Tl: SPLP => SET PARTITION min 3 Zex,- Dan (+34) xy=l, allj, xy tHty=1, allay, a Vi ip TI ty JA Kyte =O00r1, alli. The term 3, f; in the objective function is shown enclosed in parentheses to indicate that it is constant. As to transformation of SPLP-S to SET PACKING, we retain the definition of Fas j,:=1—y,, all i. Furthermore, for all j, let w, be a nonnegative ‘artificial’ slack variable used to convert the » equations (8) into inequalities, ie. w,+3,x,,= 1, w=0 or 3,x,,<, all j. To assure that all w, become zeros in an optimal solution to SPLP, we associate with each of these slack variables a ‘sufficiently large’ weight L in the objective function. To obtain the desired effect, it suffices to choose L> max, {min,{ f,+¢,,}}. Since a solution x minimizing some function g(x) will also maximize — g(x), and since we, as before, can restrict all x,,’s to be either 0 or 1, the objective function for SPLP becomes min Deis LAFF LS + Dh or, for wj=1—3,x,,, all j T2: SPLP SET PACKING max DE(L-ey)x + Dh (-mL- Bh). Yayj3,f. We replace w, by —1 + 3,3,a,; in the objective function and obtain oi — ta = ( fit 23a, Dray? lh, aus} 7 7 : or (-Ln) + min{( + AeL)y| yA =e} which is SET COVER. Once again, as in T1 where a constant term 3, f, was included in the objective function, the objective function in T4 also includes a constant, here — Ln. Another equivalent formulation of min(3, f,y4]3,J,a1,= 1, all j) is niin Zr LD wt Soya = Low >o,any), i 7 7 for L>,f, as before. Proceeding in the same manner as above, we obtain rin Ln z (s- L3e,)¥ Bays 1, au} J. Krarup, P.M, Pruzan / The simple plant location problem 45 or (-Ln) = max((—/+ AeL)y| yA $, f, in both cases, By showing the number of variables and constraints for both the original problem and its transformed version in each case, the five transformations developed so far can be summarized as Proposition 2. Number of Number of var. constr. var. constr, : SPLP-S m+mn n+mn=SET PARTITION m+2mn_ n+mn SPLP-S m+mmn nt+mn=SET PACKING =o m+mn— ntmn SET COVER m n = SPLP-S m nen SET PARTITION m n > SET COVER m a : SET PARTITION m n = SET PACKING =m a The relationships established via TI-TS have, in some form or other appeared in the literature. The earliest publication year found for each is: TI: According to Guignard and Spielberg (1977), this transformation was discussed among Hoffman, Johnson and Padberg and suggested to the authors by Hoffman. T2: Padberg (1979); T3: Krarup (1967), T4: Lemke et al. (1971); TS: Balas and Padberg (1975b). As to the transformation T4 in the inverse order, Balas and Padberg (1976) note that SET COVER cannot be brought to the form of SET PARTITION. Apparently, they tacitly assume that the number of variables and constraints should be preserved as was the case for T4 and TS; otherwise, if a (substantially) larger number of variables and constraints is permitted, it is well known (see Section 3) that the so-called decision problems underlying SET COVER and SET PARTITION can be transformed to each other (or for that matter to any other so-called NP-complete decision problem). What is noteworthy here is that the transformations SET COVER + SET PARTITION or SET PACKING can be achieved via SPLP. More precisely, first T3 then T1 and T2 respectively: Proposition 3. Number of Number of var. constr. var. constr, T6: SETCOVER m =n — SET PARTITION: m+2mn_ n+mna TT: SETCOVER m nn = =SET PACKING = m+mn_— ntimna Note that a data instance of SET COVER for which no feasible solution exists will correspond to an unbounded objective function for any feasible solution to the instances of SET PARTITION and SET PACKING as defined via SPLP by Proposition 3. ‘The following example illustrates all seven transformations: Example 1. Let (m,n) = (3,4) and (:) { 00 | f={4], A=40 1 1 OF. 2 1110 For this particular instance, let y° and y? denote optimal solutions to SET COVER and SET PARTITION respectively. It is seen that y°=(1, 0, 1), f°=5, y= (1, 1, 0) and fp? =7, To save space we shall only consider the transformations TI and 2 as parts of T6 and T7 respectively. ‘The transformations are therefore presented in the order T3, T4, TS, T6 or T3-+ TI, T7 or T3 + T2. 46 J. Krarup, P.M. Prusan / The simple plant location problem 3: SET COVER = SPLP: Let (x°,y®) be the optimal solution to the instance of SPLP defined by m, 1, f, ¢, where ¢,,=0 if a), = 1; otherwise, ¢,j:= 002 0 wo 0 C=} 0 0 cf. O00) Thus, »°= =p. T4: SET PARTITION = SET COVER: L>¥,f,=9; L:=10; AeL = (20, 20, 30); f+ AeL = (23,24,32). Optimal solution y’ to SET COVER with data m, n, A, f+ Ael is »’=(1, 1, 0)=y". We observe that ~Int (f+ AeL)y’ = —40+47=7=f?. TS: SET PARTITION + SET PACKING: For L= 10, the optimal solution y” to SET PACKING with data m,n, A, —f-+AeL or m, n, A, (17, 16, 28) is y” = (I, 1, 0)=y?. We note that —Ln-+ (—f+ AeL)y” = 40+ (17, 16, 28)y” = -7= —p?. T6 (T3+TI): SET GOVER=SPLP SET PARTITION: The instance of SPLP defined by T3 is transformed to the instance of SET PARTITION shown below: 1,0, 1) =y% x9, (or x9) = 15 x% = x9, = x95 = 1 while all other x8 =0. 3,3,c,,2), + Sf =5 es _ 10 1 _ 1 1 0 a) 10 2 Tw 2 _ xy 2 to f% 2 1 cl Too _ Lo _ | 3 i ® 1 _ a _ 2 : ‘ ‘ : i : ' uy 3 ' 2 1 \ 3 Ls 3 to ao 16 tf 1 _ 2 ta dy ‘T6 (T3+TI): SET COVER=SPLP~SET PARTITION The circled entries correspond to the starred variables equal to 1 in the optimal solution and clearly demonstrate that the solution found does in fact constitute a partition. The cost vector associated with the J. Krarup, P.M. Pruzan / The simple plant location problem 47 m+ 2nmn = 27 variables is shown in the extreme right-hand column. For the solution shown, note that Dp — LHI +S, —4t9=5= fh". 17 (T3 + T2): SET COVER = SPLP = SET PACKING: The same instance of SPLP is finally transformed to an instance of SET PACKING with m+ n= 15 variables. For L>max,(min,( f+ ¢,,}}=3 or, for example, for L:= 5, we obtain the instance of SET PACKING given below: Kye 17 (13472): SET COVER+SPLP-»SET PACKING For the solution shown, 3,3,(L—¢,)xjj+ 3,f,5)— nb —3,f X444-5xX4-9=-S=-f. O Extensive use of transformations T1-T3 have been made in various contexts. Examples include studies of the SPLP-polytope discussed in Guignard and Spielberg (1977) and Guignard-Spielberg (1980), the polynomial heuristics devised in Hochbaum (1979), and the investigations of facets of the SPLP-polytope reported on in Cornuéjols and Thizy (1980) and discussed in Section 12. In addition to the transformations shown, it will be established in Section 11 that p-MEDIAN and p-CENTER are transformable to SPLP such that the number of strategic y-variables is maintained. 3. The computational complexity of SPLP and related problems According to Guignard and Spielberg (1977): “The SPLP is one of the simplest mixed integer problems which exhibits all the typical combinatorial difficulties of mixed (0-1) programming and at the same time has a structure that invites the application of various specialized techniques”. This statement indicates that SPLP is a hard nut to crack, of, to use a more precise characterization, that it is highly unlikely that an exact polynomial time bounded algorithm can ever be devised for its solution. In the following, we will provide a series of characterizations of the four related integer programming problems in terms of computational complexity, to demonstrate that they all indeed belong to the class of combinatorial optimization problems termed NP-hard. Since excellent textbooks and expository articles on the theory of NP-completeness are available, we shall resort to intuitively appealing on definitions of the classes @ and N9, sufficient for conveying our main message: SPLP itself is termed NP-hard and the so-called decision problem related to SPLP is NP-complete. Pertinent references include Garey and Johnson (1979) and the survey-type paper by Lenstra and Rinnooy Kan (1979). First a word about the distinction between polynomial and exponential time bounded algorithms. For our purposes, we can consider algorithms as step-by-step procedures for solving problems in a finite 48 J. Krarup, P.M. Pruzan / The simple plant location problem computing time for all data instances. For a given problem type (eg. SPLP) and a set of data instances of a given size corresponding to a given input length (e.g, SPLP (m1), the complexity function for an algorithm expresses the largest amount of time required for solving the problem for an arbitrary data instance of that Siew, Am algorithm is called polynomial (rime bounded) or good ot fast if for all data instances its time complexity function is bounded by some polynomial function of the input length; otherwise the algorithm is called exponential. Since this definition involves ‘all data instances’, we can alternatively say that an algorithm is either polynomial or exponential in the worst (most time consuming) case. ‘a contrast to optimization problems dealt with so far, a decision problem 7 has only two possible solutions, either the answer ‘yes’ or the answer ‘no’. As an example, a decision problem 7 (SPLP) related to SPLP is: For a given instance m, n, C,fand a given threshold value k, does SPLP have a solution of value at most k? ‘The formal theory of NP-completeness is based on the concept of deterministic and nondeterministic ‘Turing machines, designed to provide yes/no answers to decision problems posed in terms of language recognition. For a given alphabet and language, an input consisting of 2 finite string of symbols from the “lphabet is accepted by the machine (je. solves the decision problem) if and only if it belongs to the language. - Ce ————— CE of all its feasible instances. A decision problem as to the feasibility of a data instance for @ given problem type is said to belong to the class 9 if feasibility or infeasibility of any data instance can be determined by some algorithm in polynomial time on a digital computer, idealized to accept input of arbitrary length. Thus, 7(SPLP) € 9 if we for given m, n, C, f, k in polynomial time can either confirm or refute its membership of the set of all feasible instances, Actually, it is not known whether =(SPLP) in general belongs (0 or not. Tt does however belong to the (wider?) class IF which can be characterized without reference to the rather exotic concept of a nondeterministic Turing machine as follows: For a given instance ¥ of a decision problem (eg. 2(SPLP) defined by m,n, Cy f, k) its feasibility implies the existence of an appropriate Prnicture 8 associated with y (eg. a binary vector of length m-t mn whose elements correspond to the assignment of values 0 or 1 to all variables y, x in (8)~(11)). If the ength of 6 (the encoding of the desired structure) is bounded by some polynomial in the length of y and if we for given (y, 8) can affirm the feasibility of 8 (eg. (8)-(11) are satisfied and the value of the objective function (7) does not exceed k) in polynomial time on a digital computer, this decision problem is then said to belong to the class U9. This nay also be informally expressed within a framework of nondeterministic computation by stating that NP is the class of all decision problems that can be solved by polynomial time, nondeterministic algorithms. Evidently, 1(SPLP) € 9.9 as asserted since both conditions for its membership are satisfied; the Single ‘Assignment Property ensures that 5 is of length m+ mn and to verify its feasibility requires computations of the order m-+ mn. If for any data instance of a decision problem =’ we can construct in polynomial time a data instance of decision problem 7 such that the instance of 7 is feasible if and only if the instance of 7 is feasible, then aris said to be polynomially transformable to = (notation: =’ cc). Thus, 2! 6 = implies that 7’ can be Viewed as a special case of = and, consequently, 7 is at least as difficult to solve as a’ If a! xa for all oe —————“—i—iE= se which is then called NP-hard, Finally is called NP-complete if a is NP-hard and =< QP. This definition of NP-completeness does not provide an efficient basis for proving that a decision problem a is NP-complete. Howover, it can be shown that if some 2’ cz, 7 belongs to OP and a’ is NP-complete, then @ is also NP-complete. ‘The Main ‘Theorem in Karp (1972) asserts the NP-completeness of 21 decision problems including a(SET COVER). If we for a given instance, m, n, A, f, k of a(SET COVER) define an instance of SPLP as Shown earlier (Proposition 2, transformation T3) then the answer to »(SET COVER) will be positive if and only if the answer to the question: “Does this instance of SPLP have a solution of value at most k?” is positive, Since a(SPLP) is in 99 and since the transformation «(SET COVER) = x(SPLP) is polynomial (in this case linear) in mm then (SPLP) is NP-complete. ‘Our primary concern in this paper is certain optimization problems and their computational tractability J. Krarup, P.M. Pruzan / The simple plant location problem 4 whereas the temporary interest in decision problems is due merely to the fact that key concepts like NP-completeness and NP-hardness are only defined with respect to these. What remains in order to characterize the computational complexity of SPLP itself is therefore to elaborate upon the connection between a given optimization problem and some related decision problem. For a given minimization problem OPT, the most direct way of deriving a related decision problem 7(OPT) is to introduce a threshold k and ask the question: “Does OPT have a feasible solution of value at most k?”, If m(OPT) - which for given OPT and given k is uniquely defined ~ possesses a certain property (e.g. NP-completeness) and if this property is retained upon certain simplifications of (OPT) then alternative decision problems 7,(OPT),... possessing that property can be formulated. Consider e.g. OPT = SET PARTITION. For a given data instance m,n, A, f, k, we have m(SET PARTITION): “Does there exist a y which is feasible for SET PARTITION and satisfies jy2°, the approximation algorithm is a heuristic. A heuristic is thus an algorithm producing feasible solutions for all data instances without guaranteeing optimality. It is practical and commonly accepted to characterize algorithms by a measure related to their time complexity function. For a specific problem type OPT and a specific algorithm A, the time complexity function f() will, as mentioned in Section 3, express the largest amount of time required for solving any instance of length n. To characterize the order of f(x), we say that f(12) is O(g()) whenever some constant a exists such that | f(22)| 0 All slack variables for the tighter constraints y, — x,,> 0 are upwards bounded by +1 while a slack variable corresponding to one of the aggregated inequalities in the more loosely constrained formulation can attain any nonnegative value less than n,, the number of clients which can be supplied from the ith facility. For 1,=n, each aggregated constraint is simply the sum of » disaggregated constraints. A feasible solution to SRS will therefore be feasible for WRS as well while the converse is generally not true. Thus, the adjectives J. Kearap, P.M Prazan / The simple plant location problem 3 ‘strong’ and ‘weak’ have the intuitive meaning that SRS with disaggregated constraints is stronger (closer to the underlying SPLP) than the weaker (less tight) form WRS. On the other hand, SRS has (m+ 1) constraints and m(2n-+ 1) variables including slack while these figures reduce to m+n and m(n-+2) respectively for WRS. Disregarding other features such as how to select the next node in the search tree from which to branch, a successful branch and bound algorithm is a lucky compromise between the quality of the bounds and the computational effort involved in their evaluation, For designers of branch and bound algorithms for SPLP, the schism at this stage is essentially reduced to a choice between WRS and SRS (or their duals): WRS with significantly fewer constraints will in general require less computational effort in each bound evaluation while SRS in general will produce tighter bounds. Since either of the two linear programs will have to be solved repeatedly, the substantial difference in the number of constraints led most early researchers to exploit WRS. The remainder of the paper however bears considerable evidence of the prudence of choice of SRS. 6.1. Optimum integral solutions to SRS It is nof true that an optimal solution to SRS automatically will be integral and thus solve SPLP itself. Balinski (1965) noted however that “it might be conjectured” but quoted the following counterexample suggested by Gomory: Example 2. An SPLP (3,3) is defined by lb Et Minimum cost for establishing any subset of at most two facilities is 5 while a minimum cost of 6 is incurred for a establishing all three facilities. An optimal fractional and unique solution (x',y"’) to SRS is 4, all i, xj, =4, ij and x,,=0, all i, for which the objective function of SRS attains the value 3. 0 Is it possible to construct an instance of SPLP(m, 1) such that min{m, ) is less than 3 and such that the corresponding SRS has an optimal fractional solution better than any discrete? The following proposition settles the question in the negative Proposition 5. For any SPLP with min(m,n} <3 there exists an optimal solution in integers to the strong LP-relaxation SRS. Proof. The proposition is trivially true for min(m, n} <2. m>2, n=2: For the constraints of SRS written in the form Ax — 6, it suffices to prove that A is totally unimodular ie. that every square, nonsingular submatrix of A has a determinant equal to +1. Upon addition of slack variables s,, to the disaggregated constraints, the coefficient matrix of SRS is yy tt Im | Xu XR Xm2| Su Sm2 Ry Re 1 1 : il 1 . 1 * 1 a al * 1 eal -1 . 1 il =i * 54 J. Krarup, P.M, Pruzan / The simple plant location problem Every column contains at most two nonzero entries and every entry is 0, +1, —1. The rows have been partitioned into two sets R, and R, (members of these sets alternate in the actual case as shown) such that the two nonzero entries in a column with the same sign (here the y-columns) do not appear in the same set R, of rows. Finally, the two nonzero entries in a column with different signs (here the x-columns) belong to the same set R, of rows. These are the sufficient conditions due to Heller and Tompkins (1956) for establishing the total unimodularity of the matrix. m=2, n>2: In this case the matrix reads: Yu Fa} Xie a Xan | St By an argument similar to the one used in the previous case, B" is totally unimodular. It is then easily verified (for details, see e.g. Garfinkel and Nemhauser (1972)) that the same applies for AT and hence for A itself. Besides det SRS, Propos exists. Nevertheless, counterexamples or not, Balinski'’s conjecture is almost correct. Though vaguely for- mulated and only supported by empirical evidence, we state it formally to emphasize its indebatable significance: ing a class of SPLP’s which is optimally solvable by solving the corresponding instances of ion 5 asserts that no smaller (in terms of mn) counterexample than Gomory’s SPLP (3,3) Proposition 6. A very large proportion of all instances of SRS will have have an integer-valued optimum solution. Since all f; are assumed nonnegative, any optimal solution (x’,y’) to SRS will satisfy y/ = max,x/,, all i. If any client's entire demand is supplied by the ith facility, then some x/, will equal 1 and so will y/ Since an SRS related to an instance of SPLP generated at random or with ‘real-world’ data most often posssesses the Single Assignment Property, split assignments (where clients are served by two or more facilities) will seldomly occur and an optimum solution in integers to SRS results. As will be elaborated upon in Sections 7 and 8, several designers of branch and bound codes for SPLP based on SRS or its dual have noted this fact and realized that in the vast majority of cases no branching at all is required upon the initial solution of SRS. J. Krarup, P.M. Prusan / The simple plant location problem 35 Table 1 ‘Optimal solutions to SRS and SPLP coincide in 96% of the instances tested (Morris (1978)) SRS and Soi(==)p SRS m Sense see ts) ease n 5 6 6 s 6 8 Sy =2 =2 >A Number of integer solutions 100-9310, 9590100 The extent to which Proposition 6 is true has been empirically investigated by Morris (1978). He augments SRS with the additional constraint 3,y,(=, >)p, solves three series each of 100 data instances for the augmented problem and records the number of integer solutions found. Another set of three series each of 100 instances deals with the pure SRS. The results found are summarized in Table 1. A few runs terminated with some y,=2 indicating the need for the upper bound constraints y, <1. It appears from Table1 that a well-performing approximate algorithm for SPLP is merely some LP-technique for solving SRS. Yet, for large values of m or n, the number mn of disaggregated constraints may well become prohibitively large. However, as noted by ReVelle and Swain (1970), there is at least one way to reduce the number of constraints in such a case: ignore them and re-introduce them only when needed. ReVelle and Swain consider what corresponds to SRS for the p-median problem in a network a (1,...:m), no fixed costs but a prescribed number p of plants to be established, ie. 3,y,=p)- Without giving exact figures paralleling those of Table 1, they characterize fractional solutions as ‘unlikely events’ and observe no such solutions in their computational experiments. On the other hand, to demonstrate that fractional solutions better than any integral solutions can occur, they present a series of ‘highly contrived” data instances. ‘The idea of ignoring the mn disaggregated constraints y,— x,,>0 in SRS has been further pursued by Morris (1978). Initially he solves min{3,Z,c,,x, +3, 43)|3,x1,= 1, all j, x, yO). Those y,—x,,0 violated by the solution found are added to the problem and the next solution is provided via dual simplex iterations, The process continues until all y; — x;,> 0 are satisfied. Limited computational experience based ‘on a series of rather small problems has indicated that less than 25% of the min constraints will need to be added and that the expected proportion of added constraints decreases for increasing values of inn. “Yet”, as Morris once remarked, “25 of a large number is still a large number”. A variant of the ReVelle-Swain-Morris “ignore the constraints approach” to circumvent the dimen- sionality problem for large mn is the variable upper bound (VUB) technique proposed by Schrage (1975). It is well known that the simplex method of LP can be modified to accommodate simple upper bounds (SUB's) (eg. x;} anda constant on the right-hand side. However, a VUB-constraint arises if y,—x,,20 is replaced by x,y, whiere y, conceptually no longer is an ordinary variable but a variable upper bound on x,,. Schrage shows how such constraints can be implicitly treated and implemented in a revised simplex algorithm with product form inverse. He furthermore exhibits three classes of problems (including strong LP-relaxations of P-MEDIAN and SPLP) where the proportion of VUB-constraints is particularly large. Computational experience comprises strong LP-relaxations of three instances of SPLP where the values of the c,,’s are intercity distances for published travelling salesman problems. The results found are shown in Table 2. The 56 P.M, Prusan / The simple plant location problem Table2 “Three instances of SRS solved with implicit representation of variable upper bounds (Schrage (1975). am 7m Si7m Number of constraints Pivots required Seconds IDM 370/168 Total Explicit 33 33 184 1122 33 33 LAs 3 2 295 1122 2 4 139 3 3 2000 2 3 07 236 VUB-algorithm terminated with integral solutions in all three cases. Schrage has also solved a number of SRS’s corresponding to the 16 instances of SPLP (25,50) considered by Kuehn and Hamburger (1963). Integer solutions were found in all cases within 2 to 3 seconds of computer time. Proposition 6 is also supported by the computational experience with the SRS-dual-based exact algorithms reported on in Section 8. Let us close this part of the SRS-discourse by discussing properties of another measure of the tightness of SRS with respect to theunderlying SPLP. Let 29,5 and 2851p denote the minimum values of the objective funetions for SRS and SPLP respectively. Due to the assumption of non-negative data, 0< 2845 = 291» for all instances. For a given feasible instance, the absolute duality gap G, is defined by G,=25.p— 28,5 and the relative duality gap is G,= G,/22y.p for 28,» >0 and G,=0 for 29 p(= 2845) =0. Clearly, 00, all i, j. Hence, the dual formulation reads: J. Krarup, P.M. Pruzan / The sinsple plant location problem 37 DSRS: The dual of SRS MAX Zpsps = {Se yw, Se, alli,j; Dw, Sf, all #5 w,,>0, atti}. i 7 Since for all j the maximum value of v, can be no larger than the minimum value of any ¢,, + w,, and since no other constraints are imposed on the v’s there exists an optimal solution to DSRS satisfying v= min{c,+m,) and mar Ze} = min, {e+}. (12) j For the feasible solution {v my") for ¢;=0, y= 0 for ¢,,=1 (12) yields oj:= 1/(751), all j, and zpses= 1 /(0 M/S") = m/t= s, Since zpsas= spss a pair of optimal sohitions for SRS and DSRS have been identified due to the Strong Duality Theorem of LP; hence o= zag = 5. Finally, for this particular family of data instances, we obtain for the relative duality gap G, Bupa za5_ st—-ttl—s st—t+1 0 ZSPLP for to and fixed s,s>1. O 6.2. WRS is solvable by inspection Besides having fewer constraints than SRS, the weak LP-relaxation WRS possesses another favourable property: it is optimally solvable by inspection. This useful result was established by Efroymson and Ray (1966). For a given instance of SPLP(m,n), let a partial solution be an assignment of binary values to an index set WCJ={l,...,m)} of the strategic y-variables. IV can be partitioned into two subsets SO and S1 corresponding to the different states of the facilities involved: SO= {i]i€ IW and y,=0)}, S1=(iJiEW and y,=1) while the remaining variables, y,, iGJ\IV constitute an index set SF of free variables SF= {ii /\WV). To simplify the notation, let R denote the set $1 U SF. For the given SPLP(m, n), any partition of F into 50, S1, SF defines a subproblem for which the corresponding WRS reads min J Leyryt & ft 3 fe ‘er 7 iesr Gest Dx allj, ,%-Dx,=0, i€SF, ieR i W20,iESF, x,,=0, alli, j. SF is assumed nonempty since the partial solution otherwise is complete. Since f,>0, no free y, need take on a larger value than necessary; thus in an optimal solution (x', y’) to WRS, all aggregated constraints will hold as equalities n,y/ =3,xj, or yf =(1/n,)3)x/p 1€ SF. Substituting this value of y/, i SF, into the objective function and discarding the constant term 3,es1 fj, WRS reduces to min DY Lapryt D Vay th/n)xyr est j iesr 7 24, iER , all, x20, alli, j 38 J. Krarup, P.M, Pruzan / The simple plant location problem or min ZT Lexy) y= loalljs =o} ier j ier where cf, =, for /€ SI and cj, =/,/n, for i SF. For all , the th client cannot be served at a cost less than ci.) ;= min;cqcf,s hence, the optimal solution is xj,:=1 for some i=i(j), i. ties are resolved arbitrarily, and 0 otherwise, all j. Since the number of computations involved in ‘solving WRS by inspection’ is proportional to | Ror (i — [S0))n, we have thus proved Proposition 8. For a given partial solution to SPLP with |S0| and |$1| variables fixed at 0 and | respectively, the WRS corresponding to the sub-problem so defined is optimally solvable by inspection (in O((m —|SO))n) time). Actually, as could be expected, the dual of WRS is also solvable by inspection. Though stated slightly differently, (we believe) this observation was first made by Spielberg (1969a), Let v=(v,,....0,) and w=(Wy,..4,¥%,) be dual variables associated with the 1 equations ¥jx,,=1 and the m inequalities n,J,~ 3)X;j2 0 respectively. Then DWRS, the dual of WRS, reads max{ Soo ~ mje all i,j; wn, =f, all i; » free, wea}. 7 To maximize B,v,, we can do no better than to maximize each component v, separately, that is, to choose vj =min(w,+¢,} which in turn implies that the optimal choice of w, is to take w,=f/n. Obviously, a similar argument applies for the dual of a subproblem of WRS defined by any partition ($0, S1, SF) of J. Proposition 9. The optimal objective function value for the dual DWRS of a given WRS (and of WRS itself) equals 3, min,{c,,+f,/n,) which can be evaluated in (ann) time. 7. Exact algorithms based on WRS or SRS. Disregarding the complete enumeration approach in Stollsteimer (1963), credit for the first attempt to solve SPLP to optimality should go to Balinski and Wolfe (1963) who devised an algorithm based on Benders decomposition. However, as reported on in Balinski (1965), the computational experience was discouraging. It seems therefore reasonable to attribute the first efficient (not to be confused with polynomial) algorithm for the exact solution of SPLP to Efroymson and Ray (1966). The Efroymson—Ray algorithm, which was independently developed by Zimmerman (1967), combines the then relatively novel branch and bound technique with WRS for providing lower bounds. Branchings are performed on the strategic y-variables with two branches emanating from each node corresponding to a selected, free y,, fixed at either 0 or 1, Since WRS is solvable by inspection due to Proposition 8, several hundred bounds per minute could be evaluated on'the computer used (IBM-7094). On the other hand, the number of positive x,, found at each node is in general small compared to 1,;; accordingly, the amount of the fixed cost absorbed by the ith facility is rather small, which in turn influences the tightness of the bounds obtained. Three simplifications have been incorporated to overcome this difficulty; simple tests are performed to check whether it pays to open or close certain facilities or to reduce some 1,’s. The report on the computational efficiency is somewhat laconic. Efroymson and Ray state that “a number of SPLP’s (m= 50, 2 = 200) have been solved in an average of 10 minutes on an IBM-7094”, However, compared with earlier works, the results obtained are indeed significant for their time. A three line footnote in Efroymson and Ray’s paper, largely overlooked by subsequent researchers but later brought into focus by Rosing et al. (1979), reports on a series of experiments with SRS resulting in J. Krarup, P.M. Pruzan / The simple plant location problem 59 ‘optimum integer solutions. Apparently stimulated by this observation, Ray joined the very few who in the following decade concentrated on SRS rather than on WRS; his work from 1969 with Davis (Davis and Ray (1969) demonstrates the significance of SRS-like LP-relaxations for an even wider class of location problems. The adjective ‘simple’ in SPLP, now widely accepted as synonymous with ‘uncapacitated’, is originally coined by Spielberg (1969a). Some features of the algorithms presented resemble those of Efroymson and Ray but their root is a general direct search mixed integer algorithm due to Lemke and Spielberg (1967). For SPLP, the mixed integer algorithm operates on WRS and its dual which are solved for a sequence of fixed y-vectors, generated by a search algorithm due to Balas (1965). This is a single branch scheme in contrast to, e.g. Efroymson and Ray. It starts with all y,=0 and proceeds by assigning 1’s to selected y, on ‘forward’ steps of the search, If the optimal solution has several open facilities, it can be expected that a good feasible solution will be found far away from the root of the search tree, It is then reasonable to relocate the search origin, say, to the best feasible solution known and to restart the search from that pi The ‘add or drop’ question which was briefly discussed in connection with heuristics plays an essential role in Spielberg's paper. Two basic approaches with natural search origins — either all facilities open (drop) or all facilities closed (add) ~ are investigated. If the fixed costs are relatively small or if the cost matrix C is sparse (i.¢; relatively few clients can be reached from each facility) it is reasonable to expect a large number of open facilities in an optimal solution and, consequently, one may expect a search origin with all facilities open (drop) to perform particularly well for such cases. Conversely, with large fixed costs and a dense C-matrix, the natural search origin will be to close all facilities (add). Spielberg sketches an ‘adaptive search algorithm’ where the search origin is none of the two extremes (all open, all closed) but preferably selected by the program itself, based on the actual data. This idea is pursued in another 1969-paper on generalized search origin, Spielberg (1969b), and further elaborated upon in Guignard and Spielberg (1969). The algorithms exploit the disaggregated form in terms of dual variable analysis leading to strong Benders inequalities and gain functions. Both local and global gain functions are devised for curtailing the search. The local gains measure the local improvement of the objective function when a free facility is either opened or closed while the global gains at a certain node in the search tree measure the maximal improvement (over all successor nodes) incurred by a forward step. These conceptu- ally simple devices together with different tests appear to be quite effective in looking ahead into the tree or to ascertaining optimality. An in-depth treatment of the significance of SRS or its dual as basis for the gain functions developed as well as the desirability of adapting SPLP algorithms to the actual data instances can be found in Guignard and Spielberg (1969) and Spielberg (1970). The mixed results obtained make it difficult to draw valid conclusions as to the data dependency of the algorithm's efficiency. Comparative results for programs with rigid origin exhibit in general a substai decrease in the number of iterations required by the modified algorithms; however, the computing time per iteration increases considerably for these. Although approaches based on generalized search origin seem promising for other problems of special structure and for integer programming in general, Spielberg (1969b) summarizes his opinion of the SPLP-situation with the following characterization: “An optimist might add that there is still ample room for improvements in strategy, corresponding to the great difficulties that still remain”, At least two noteworthy exact WRS-based methods for SPLP appeared in 1972 (Khumawala (1972) and Hansen (1972)). Since the latter refers to the first, we start by giving an account of the results achieved by Khumawala whose approach can be regarded to be a significant improvement of the Efroymson-Ray algorithm with respect to storage requirements and computing time. Like his forerunner, Khumawala operates on WRS but devises an efficient updating procedure, based on the three simplifications proposed by Efroymson and Ray. The purpose is to tighten the aggregated constraints 1,y,— 3,x;,# 0 by seeking n, reduced and to determine bounds at each node of the search tree according to which certain facilities can be fixed either open or closed when the next branching is performed. Finally, an efficient storage scheme for bookkeeping information is developed. Eight different branching decision rules are investigated and compared for 16 instances of SPLP (24,50). The so-called ‘Largest Omega’ rule, based on the maximum net gain achieved by opening a free facility 60 J. Krarup, P.M. Pruzan / The simple plant location problem appears to be superior in efficiency to the other branching rules examined. The average computing time with ‘Largest Omega’ for the 16 test problems is approximately 3.8 seconds on a CDC 6500. It is interesting to note, that Khumawala’s results led to the following observation by one of his referees: “*... the problems for which there are about eight warehouses [facilities] open in the optimum solution seem to be most difficult; whereas, those with appreciably less than or greater than cight are relatively easier”. Similar phenomena have been noticed by other researchers; the point is not that eight is a crucial number but that the optimum for these test data is flat in the sense that near-optimal solutions exist for a wide range of open facilities. Hansen (1972) deals with two implicit enumeration algorithms exploiting the concepts of additive penalties (akin to the gain functions introduced by Spielberg). Recalling the definition of the three disjoint index sets (SO, S1, SF) at each node of the search tree, a penalty can be viewed upon as an increment of some lower bound z on the objective function z when a free variable y,, i € SF, is fixed at either 0 or 1. If the values of a subset of free variables are fixed simultaneously, some at 0 and some at 1, and if the addition of the corresponding penalties to a known lower bound can be shown to constitute a new valid bound, then these penalties are said to be additive. Exact, dual-based algorithms Up to this point, the different methods covered have been presented in a chronological order. In this section, however, we shall simultaneously deal with two algorithms due to Bilde and Krarup (1967) and Erlenkotter (1978) which conceptually are closely related although their development spans about a full decade. Both methods, of which the latter is the most powerful known to date for the exact solution of SPLP, can be characterized as dual-based or parametrized Lagrangian relaxation, and both methods operate ‘on SRS, the strong LP-relaxation of SPLP, and its dual DSRS. DSRS was introduced in Section 6 as =(Zole max Zpsrs = wy Se; all i,j; Dy =f, all i; w,; 0, all i} J and we observed that there exists an optimal solution to DSRS satisfying max{ Zo} =S min (6,4) i a =ZBsrs (13) which leads to the following equivalent formulation in Bilde and Krarup (1967) involving only the explicit variables 1, : Lower Bound Maximization Problem MAX Zpsps {z min {ey} | D0 Sh, all i; w,=0, all isl} (14) i d Since B55 = 2fns by the Strong Duality theorem of LP, ‘lower bound! above refers to the inequality ZDsrs = ZspiP* ‘An alternative way of reasoning provides another equivalent form of DSRS, now involving only the v, as, explicit variables. For any set of given »,’s, we can without affecting the objective function assign the lowest possible value to all's max{0, 6)~¢,,)- (1s) This observation was made earlier by Spielberg (1969) but is here further exploited by Erlenkotter (1978) ‘on his way to the so-called condensed dual, obtained by substituting (15) into (13): Condensed Dual {Zaz max{0, 9, j wy mMaX Zpsp )0 so as to increase min, {¢, + 4.) = v,. To obtain the desired effect we must simultaneously increase (at least) all 1,,’s corresponding to the admissible cells, Unless A, = J, at least one new admissible cell will appear in column & provided that none of the constraints 3,1, b, Bx >d, x =0, some x, integer} (21) the Lagrangian relaxation of (21) relative to the constraint set Ax>b and a conformable nonnegative J. Krarup, P.M, Pruzan / The simple plant location problem 6 vector A is defined by min z, = (cx +M(b— Ax)| Bx, x*0, some x, integer}. (22) If x’ is feasible for (21), x’ is also feasible for (22). Besides, Ax’ > b whereby z= cx’ = cx’ -+Mb— AX’) = 21 for all A>0 in accordence with the definition of a relaxation in Section 4. The essence of Langrangian relaxation is to identify a set of ‘complicating constraints’ (Ax > b), weight these by multipliers and insert them in the objective function in order to obtain a new problem which, hopefully, is simpler to solve than the original. One of the earliest and perhaps most convincing examples of this technique is the work of Held and Karp (1971) on the symmetric traveling salesman problem where the NP-hard problem of finding a Hamiltonian circuit of minimum weight is relaxed to the simple problem of finding a minimum-weight 1-tree, which is solvable in polynomial time by the greedy algorithm. Lagrangian relaxation can thus be viewed as an alternative to LP — and frequently a very attractive alternative ~ as a means of providing bounds in a branch and bound algorithm, ‘Two different Lagrangian relaxations of SPLP almost suggest themselves. Within the context of the dual-based algorithms discussed above, the first is obtained by taking y, — x,,> 0 as the set of ‘complicating constraints’ with the w,,’s as the corresponding set of nonnegative multipliers: Lagrangian relaxation of SPLP VUlay tm) +B (4- Boy) Bx minz,(w (23) allj, — x>0,y,E (0,1), alli, 7. The second Lagrangian relaxation is obtained by taking 3jx,,=1, all j, as the set of ‘complicating, constraints’. The discussion of this approach will be deferred to Section 10. If the values of all multipliers are fixed, the y-variables can be determined by inspection. Then (23) may be written as mins) = {8 Zleyt w)xy+ dD msin{ S000} | Dxj=1, x20, all aj. ay i i t If for some i min{ f,— Z,1v,j, 0) is negative, then its value can be increased to 0 by decreasing E,ys,, until this sum equals j,. Such a change in the w,, will increase the minimum value of the objective function. Therefor, in order to obtain as good bounds on 22, as possible via maximizing z,(.), only nonnegative ny satisfying Z,v,,0, all i, }. For such w, 2,(w) is solvable by inspection; the minimum value 29(.) of zy(av) is simply the sum of the column minima of the (C+ 1’)-matrix 2E(v) = min {c,,+,,). (24) (24) is recognized as the expression used previously for removing the v-variables from DSRS on our way to the simplified formulation (14). (24) is, in other words, the lower bound maximization problem for fixed multipliers w,,, but the quality of the bound thus obtained is, of course, strongly dependent on how the multipliers are selected. The sharpest bound is found by maximizing 2?(w) over the set of feasible multipliers maxz9( [Zmjoley +912, hav, 0,alli,j} (25) which is seen to coincide with the bound obtained via the LP-relaxation (14) or its equivalent form (16). 64 J. Krarup, P.M, Prusan / The simple plant location problem (25) is also known as the formal Lagrangian dual of (21) with respect to the constraints Ax> b since its solution respresents the best possible choice of multipliers. We note that this result whereby both Lagrangian relaxation and DSRS lead to the lower bound maximization problem for determining bounds on the optimal solution of an SPLP does not hold in general for an arbitrary combinatorial optimization problem. In fact, for a given integer programming formulation it can be shown that the Lagrangian relaxation will provide at least as sharp bounds as the corresponding LP-relaxation and that the two relaxations will provide equally sharp bounds if the bound provided by the Lagrangian relaxation is not increased by relaxing the integrality restrictions in which case the Lagrangian relaxation is said to possess the Integrality: Property, see Theorem 2, in Geoffrion (1974). Thus, to summarize: In terms of Lagrangian relaxation, the dual-based approach can be viewed upon as a parametrized relaxation where the dual ascent (for DUALOC supplemented by dual adjustment) is an approximate algorithm for setting the parameters (multipliers) in an attempt to obtain the best possible bounds. ‘According to recent surveys by Shapiro (1979) and Fisher (1981), Lagrangian relaxation has been successfully applied to several hard combinatorial optimization problems over the past decade. In most cases, it is quite obvious how the set of complicating constraints should be selected. SPLP however is an exception in this respect as will be demonstrated in Section 10 where the second ‘natural’ Lagrangian relaxation of SPLP (where the equations S,x,,= 1, all j, are viewed as ‘complicating’) will be investigated. 9. A class of SPLP’s solvable in polynomial time As introduced in Section 1, a client j can be served by the ith facility only if ¢,,j. ‘There is no need to give the symbolic formulation of ELS since it evidently becomes identical to that of SPLP, It appears furthermore that all elements in the lower triangular part of C are infinite, preventing units from being sold before they are manufactured, All finite elements (at least one in each column) are consecutive within each row. Accordingly, the C-matrix possesses Property A and it is fairly simple to verify that C also possesses the more restrictive Property B. ELS is traditionally solved by dynamic programming and a DP-formulation taking advantage of the fact that all x,,’s will be integers in an optimal solution will not require a greater number of computations than the linear algorithm, The SPLP-formulation (or, to be more specific, the properties of the corresponding SRS), however, enables the user to solve even very large problems by hand. Finally, the link to LP implies that the entire arsenal of post-optimal analyses of LP is immediately at hand for any instance of SPLP possessing Property B though care must be taken due to the integral nature of SPLP. The following decision problem is known as CONSECUTIVE ONES SUBMATRIX (COS); see Garey and Johnson (1979). Given an m Xn matrix A of zeros and ones and a positive integer k: does there exist an mX k submatrix B of A that has the ‘consecutive ones’ property, ie. such that the columns of B can be permuted so that in each row all the 1’s occur consecutively? As shown in Booth (1975), COS is NP-complete for k

    You might also like