Cap 4 5 y 6 Bazaraa

You might also like

Download as pdf
Download as pdf
You are on page 1of 159
FOUR: STARTING SOLUTION AND CONVERGENCE In the previous chapter, we developed the simplex method with the assumption that an initial basic feasible solution is at hand. In many cases, such a solution is not readily available, and some work may be needed to get the simplex method started. In this chapter we describe two procedures (the two-phase method and the big-M method), both involving artificial cariables to obtain an initial basic feasible solution to a slightly modified set of constraints. The simplex method is used to eliminate the artificial variables and to solve the original problem. We also discuss in more detail the difficulties associated with degeneracy. In particular we show that the simplex method converges in a finite number of steps, even in the presence of degeneracy, provided that a special rule for exiting from the basis is adopted. THE INITIAL BASIC FEASIBLE SOLUTION Recall that the simplex method starts with a basic feasible solution and moves 10 an improved basic feasible solution, until the optimal point is reached or else 197 138 STARTING SOLUTION AND CONVERGENCE unboundedness of the objective function is verified. However, in order to initialize the simplex method, a basis B with b= B~'b > 0 must be available, We shall show that the simplex method can always be initiated with a very simple basis, namely the identity. Easy Case Suppose that the constraints are of the form AX O where A is an m x 7 matrix and b is an m nonnegative vector. By adding the slack vector x,, the constraints are put in the following standard form: Ax + x, = b, x > 0, x, > 0. Note that the new m (m + n) constraint matrix (A, I) has rank m, and a basic feasible solution of this system is at hand, by letting x, = b be the basic vector, and x = 0 be the nonbasic vector. With this starting basic feasible solution, the simplex method can be applied. Some Bad Cases In many occasions, finding a starting basic feasible solution is not as straight- forward as the case described above. To illustrate, suppose that the constraints are of the form Ax 0 but the vector b is not nonnegative. In this case, after introducing the slack vector x,, we cannot let x= 0, because x, = b violates the nonnegativity requirement. ‘Another situation occurs when the constraints are of the form Ax > b,x > 0, where b > 0. After subtracting the slack vector x,, we get AX ~ x, = b, x > 0, and x, > 0. Again, there is no obvious way of picking a basis B from the matrix (A, =D with b= Bo'b > 0. In general, any linear programming problem can be transformed into a problem of the following form, Minimizes ex Subject to Ax x>0 where b > 0 (if 6, <0, the ith row can be multiplied by —1). This can be accomplished by introducing slack variables and by simple manipulation of the constraints and variables. If A contains an identity matrix, then an immediate basic feasible solution is at hand, by simply letting B = I, and since b > 0, then B™'b = b > 0, Otherwise, something else must be done, 4.1 THE INITIAL BASIC FEASIBLE SOLUTION 139 Example 4.1 2. Conder fllonng consis ne 30 Alter ing tess varios 35 nd g we ge A obvi rang tase fae ation gen ty ny =[2]=[4] and }-(] bo b. Consider the following constraints: my te $456 = 2x + 3x4 2x23 x 420 Note that x; is unrestricted. So the change of variable x, = xj" ~ xj is made. ‘Also the slack variables x, and x, are introduced. This leads to the following constraints in standard form: xt soy boat uty 2x 42x; $3x,42x, x5 XP, XT. te ty ty 4520 Note that the constraint:matrix does not contain an identity and no obvious feasible basis B can be extracted. 140 ‘STARTING SOLUTION AND CONVERGENCE ¢. Consider the following constraints: Xytayn 2s = = Ixy txt 34567 yh, 20 Since the right-hand side of the first constraint is negative, the first inequality is multiplied by —1. Introducing the slack variables x, and x, leads to the following system: Hx cxytigry = Sixty tiny +557 Xp Xy Tet? 0 Note again that this constraint matrix contains no identity Atificial Vari os After manipulating the constraints and introducing slack variables, suppose that the constraints are put in the format Ax = b, x > 0 where Ais an m % m matrix and b > Os an m vector. Further suppose that A has no identity submatrix (if A has an identity submatrix then we have an obvious starting basic feasible solution). In this ease we shall resort to artificial variables to get a starting basic feasible solution, and then use the simplex method itself to get rid of these artificial variables, To illustrate, suppose that we change the restrictions by adding an artificial vector x, leading to the system Ax +x, =, x >0, x, > 0. Note that by construction, we forced an identity matrix corresponding t0 the artificial vector. This gives an immediate basic feasible solution of the new system, namely x, = band x = 0, Even though we now have a starting basic feasible solution, and the simplex method can be applied, we have in effeet changed the problem. In order to get back to our original problem, we must force these artificial variables to zero, because Ax = b if and only if Ax + x, = b with x, = 0. In other words, artificial variables are only a tool to get the simplex method started, bbut we must guarantee that these variables will eventually drop to zero. ‘AC this stage, it is worthwhile to note the difference between slack and artificial variables. A slack variable is introduced to put the problem in equality form, and the slack variable can very well be positive, which means that the inequality holds as a strict inequality. Artificial variables, however, are not 4.1 THE INITIAL BASIC FEASIBLE SOLUTION 1m legitimate variables, and they may be introduced to facilitate the initiation of the simplex method. These artificial variables, however, must eventually drop 10 zero in order to attain feasibility in the original problem, Example 42 Consider the following constraints: xy +2x>4 = 3x, + 4x, > 5 2x +x <6 ym? 0 Introducing the slack variables x,, xy, and x, we get xy 42-5 <3ytdn = Ixy + xy eee) ne a 210) ‘This constraint matrix has no identity submatrix. We can introduce three artificial variables to obtain a starting basic feasible solution. Note, however, that x, appears in the last row only, and it has coefficient 1. So we only need to introduce two artificial variables x, and x,, which leads to the following system. Legitimate variables Artificial variables yt 2m ees -3xjtdx xy te Qe om Rs 7 Xie Xa Xa y Kay Xs, Ny 2770 Now we have an immediate starting basic feasible solution of the new system, namely x5 = 6, x= 4, and x, = 5. The rest of the variables are nonbasic and have value zero. ‘Needless to’ say, we eventually would like for the artificial variables x. and x, to drop to 2et0. 142 STARTING SOLUTION AND CONVERGENCE 4.2 THE TWO-PHASE METHOD ‘There are various methods that can be used to eliminate the artificial variables, One of these methods isto minimize the sum of the artifical variables, subject to the constraints Ax + x, = b, x > 0 and x, > 0. If the original problem has a feasible solution, then the optimal value of this problem is zero, where all the artificial variables drop to zero. More importantly, as the artificial variables drop to zer0, they leave the basis, and legitimate variables enter instead. Eventually all artificial variables leave the basis (this is not always the case, because we may have an artificial variable in the basis at level zero; this will be discussed later in greater detail). The basis then consists of legitimate variables. In other words, we get a basic feasible solution of the original system Ax = b, X > 0, and the simplex method can be started with the original objective ex. If, fon the other hand, after solving this problem we have a positive artificial variable, then the original problem has no feasible solutions (why?). This procedure is called the two-phase method. In the first phase we reduce artificial variables to value zero, or conclude that the original problem has no feasible solutions. In the former ease, the second phase minimizes the original objective function starting with the basic feasible solution obtained at the end of the phase I. The two-phase method is outlined below: Phase I Solve the following linear program starting with the basic feasible solution x = 0 and x, = b. Minimize 1x, Subject to Ax+x,=b x, x20 If at optimality x, # 0, then stop; the original problem has no feasible solutions. Otherwise let the basic and nonbasic legitimate variables be xq and xy. (We are assuming that all artificial variables left the basis. The case where some artificial variables remain in the basis at zero level will be discussed later.) Go to phase II Phase I! Solve the following linear program starting with the basic feasible solution X,=Btband x, = 0. Minimize egxy + ©yXy Subject to Xp + B'Nxy= Bo'b Xp. Xy> 0 4.2 THE TWO-PHASE METHOD 143 The optimal solution of the original problem is given by the optimal solution of, the foregoing problem. Example 4.3 Minimize x,-2x, Subject to x, + 24>2 -xyt xP 3 xy 20 ‘The feasible region and the path taken by phase I and phase Il to reach the optimal point are shown in Figure 4.1. After introducing the slack variables x, q-%s the following problem is obtained. Minimize —x,-2x, Subject to 4+ x = -xty 7m <1 % $ayo3 By Xa Xa Xu ¥520 An initial identity is not available. So introduce the artificial variables x, and x, (note that the last constraint does not need an artificial variable). Phase I starts bby minimizing xp = x, + x7 Figure 4.1. Example of the two-phase method. 14 STARTING SOLUTION AND CONVERGENCE PHASE! ArmiFiciaLs ko eR i700 0 0 0 -1 -1] 0 oj; 1 1 -) 0 0 1 0] 2 o}-1 1 0 -1t 0 69 1.1 oo}; 0 1 0 1 0 013 ‘Add rows 1 and 2 to row 0 so that 2, — cg = 2, — ¢; = 0 will be displayed. Xo Mee »[1]¢_2-! -1 0 0 ~~ 07 3 x fo; i a -t 0 0 1 oj] 2 x] o}-1 @ 0 -1 0 0 «1/4 x{o}| o 1 0 o 1 0 0] 3 eee ese x fl 20 -1 To 0 -2) 7 x|o! @ o -1 1 0 1 -1) 9 ee eo x3 Lo| 1 0 0 1 1 0 -1})2 ets nus x [1] oo 0 oo 0 ufo| 1 0 -} 4 0 + my) oo) 0 1 -$ -} 0 3 yi oo o $ 2} 1 i This is the end of phase I. We have a starting basic feasible solution, (xx) = (2, 2). Now we are ready to start phase II, where the original objective is minimized starting from the extreme point (1, 3) (see Figure 4.1) ‘The artificial variables are disregarded from any further consideration. 4.2. THE TWO-PHASE METHOD 145 PHASE I! Multiply rows I and 2 by | and ~2 respectively and add to row 0 producing 4-4. 2 T =r x | 0 1 x] 0 o 8 Oa Since z)— ¢ <0 for all nonbasic variables, the optimal point (0, 3) with objective ~6 is reached. Note that phase I moved from the infeasible point (0, 0), tothe infeasible point (0, 1), and finally to the feasible point (|, 3). From this extreme point, phase IT moved to the feasible point (0, 2), and finally to the optimal point (0,3). This is illustrated in Figure 4.1, The purpose of phase Tis to Bet us {0 an extreme point of the feasible region, while phase II takes us from this feasible point to the optimal point. 146 STARTING SOLUTION AND CONVERGENCE Analysis of the Two-Phase Method At the end of phase I either x, % 0 or else x, = 0. These two cases are discussed in detail below. Cose A: x, #0 Its, 0, then the orginal problem has n feasible solutions, because if there is anx > O with Ax =b, ten (3) is a feasible solution ofthe phase T problem and 0(x) + 1(0) = 0 < 1x, violating optimality of x,. Example 4.4 (Empty Feasible Region) Minimize Bx, t4x, Subjectto xt 5 4 2xy+3x, 218 ree) The constraints admit no feasible points, as shown in Figure 42. This will be detected by phase I. Introducing the slack variables x, and xy we get the following constraints in standard form: mt mts =4 Qxt3x, x= 18 My Re Xe MPO Since no convenient basis exists, introduce the artificial variable x, into the second constraint. Phase I is used to get rid of the artificial Figure 4.2. Empty feasible region. 4.2 THE TWO-PHASE METHOD a7 PHASE | Xo a x x Xs RES oO 0-1 0 1 0 0 4 0 1 1 18 is displayed, X eyes 4 % Xs Rus Xo 1 2 3 0-1 0 18 ‘The optimality criterion of the simplex method, namely z, ~ c, < 0, holds for all variables; but the artificial x, > 0. We conclude that the original problem has no feasible solutions Cote B: x, = 0 This case is further divided into two subcases. In subcase BI all the artificial variables are out of the basis at the end of phase I. The subcase B2 corresponds to the presence of at least one artifical in the basis at zero level. These cases are discussed below. Subcase BI (All Arifcials Are Out of the Basis) Since at the end of phase I we have a basic feasible solution, and since x, is out of the basis, then the basis consists entirely of legitimate variables. If the legitimate vector x is decomposed accordingly into x, and xy, then at the end of phase I we have the following tableau. Xo xy x Rus ~ Ce Let en oe Te | ‘Now phase Il can be started, with the original objective, after discarding the columns corresponding to x,. (These columns may be kept for the purpose of vas STARTING SOLUTION AND CONVERGENCE bookkeeping since they would present B~' at each iteration. Note, however, that an artificial variable should never be permitted to enter the basis again.) The 2, — 6's for the nonbasic variables are given by the vector ¢B-!N ~ cy, which can be easily calculated from the matrix B~'N stored in the final tableau of phase I. The following initial tableau of phase II is constructed. Starting with this tableau, the sisaplex method is used to find the optimal solution. ‘Subcose 82 (Same Arifcials Are in the Basis at Zero Level) In this case we may proceed directly to phase II, or else eliminate the artificial basic variables, and then proceed to phase II. These two actions are discussed in further detail below. PROCEED DIRECTLY TO PHASE 1! First eliminate the columns corresponding to the nonbasic artificial variables of phase I. The starting tableau of phase II consists of some legitimate variables and some artificial variables at zero level. The cost row consisting of z; ~ 67s is constructed for the original objective function so that all legitimate variables that are basic have z, ~ c= 0. The cost coefficients of the artificial variables are given value zero (justify!). While solving the phase II problem by the simplex method, we must be careful that artificial variables never reach a positive level (since this would destroy feasibility). To illustrate, consider the following, tableau, where for simplicity we assume that the basis consists of the legitimate variables x), x3,...,% and the artificial variables xq4.¢41) ++ Xpym (the artifi= cial variables x,,.,,---+ yx left the basis during phase 1), 2M te Meee See Brom RUS 4.2. THE TWO-PHASE METHOD 149 Suppose that :; ~ 6 > 0, and so x; is eligible to enter the basis. If yy > 0 for + 1,...,m, then the artificial variable x,,, will remain ze70 as x enters the basis and the usual minimum ratio test is performed. If, on the other hand, fat least one component »,, <0 (r= k+ l,..., or m) then the artificial variable x,4, becomes positive as x, is increased, This action must be prohibited since it would destroy feasibility. This can be done by pivoting at y, rather than using the usual minimum ratio test’ Even though y,; < 0, pivoting at y,, would maintain feasibility since the right-hand side at the corresponding row is Zero. In this case x, enters the basis and the artificial x,,, leaves the basis, and the objective Value remains constant. With this slight modification the simplex method is used to solve phase II FIRST ELIMINATE THE BASIC ARTIFICIAL VARIABLES AT THE END OF PHASE I Rather than adopting the preceding rule, which guarantees that artificial vari ables will always be zero during phase II, we can eliminate the artificial variables altogether before proceeding to phase II. The following is a typical tableau (possibly after rearranging) at the end of phase I. The objective row and column do not play any role in the subsequent analysis and are hence omitted. off ] oo oO] & a ' 8, n foo. of & 5 || 1 0 Gens Oils ves POO 3 7 7 8 ® ' ° sun Lo 0 ° fo We now attempt to drive the artificial variables Xe - +. Neem OUt of the basis by placing m — & of the nonbasic legitimate variables x,y.» =» %, into the basis. For instance, x,,,,,, can be driven out of the basis by pivoting at any nonzero element of the first row of R,. The corresponding nonbasic legitimate variable enters and x,,,, leaves the basis, and the tableau is updated. This process is continued, if all artificial variables drop from the basis, then the basis ‘Actually, if any one of the y's is potive for r= K+ Ly... my then We may wse the usual ‘minimum rato test a 4 will entr the Basis at zero level, 150 STARTING SOLUTION AND CONVERGENCE would consist entirely of legitimate columns, and we proceed to phase II as described in subcase BI. If, on the other hand, we reach a tableau where the matrix Ry = 0, none of the artificial variables ean leave the basis by introducing Means Meaae sees OF ye Denoting (x), 55-665 %) amd (ipaq %_) BY fAn| Aw and x respectively, and decomposing A and b accordingly into | F°-2- [o and. b |, it is clear that the system [be [4Au| An] is transformed into kn-k k [EPR 2] fo, wv alotellal-[2 through a sequence of elementary matrix operations. This shows that rank (A, b) = k 0. Minimize ex Subject to AX=b x0 If no convenient basis is known, we introduce the artificial vector x,, which leads to the following system: Axtx=b Xx, 0 ‘The starting basic feasible solution is given by x, = b and x = 0. In order to reflect the undesirability of a nonzero artifical vector, the objective function is ‘modified such that a large penalty is paid for any such solution. More speci cally consider the following problem. Minimize ex+M Ix, Subject to Ax+x,= b x20 where M is a very large positive number. The term M Ix, can be interpreted as a penalty to be paid by any solution with x, %0. Even though the starting solution x = 0, x, = b is feasible to the new constraints, it has a very unattrac- tive objective value, namely M tb. Therefore the simplex method itself will try to Bet the artificial variables out of the basis, and then continue to find the optimal solution of the original problem, 4.3 THE BIG-M METHOD 155 ‘The big-M method is illustrated by the following numerical example. Valida- tion of the method and possible complications will be discussed later. Example 4.6 Minimize x,-2x, Subject to x4 x2 xt Pl <3 bor ae0) This example was solved earlier by the two-phase method (Example 4.3). The slack variables xy, xy, x, are introduced and the artificial variables x, and x, are incorporated in the first two constraints. The modified objective function is 2 x; ~ 2x, + Mx, + Mx,, where M is a large positive number. This leads to the following sequence of tableaux. Pom 8 8 mw em kas [ip =t 20 0 0 -M -M | 0 fo) a 1-1 00 1 0 2 o| -1 1 0-10 0 7 1 o| o i 0 o1 06 Oo 3 Multiply rows 1 and 2 by M and add to row 0. 2 omy me me z 7|—1 242M -M -M 0 0 ee x | 0 1 1 -!r 0 0 Tt O 2 slo) -1 @ 0-1 0 0 1 1 x[o| 0 1 0 o 1 0 o 3 z 1] i+iM 0 «fol @ 0 = alo! -T 1 slo) 1 oo 156 STARTING SOLUTION AND CONVERGENCE Rus Since z, ~ ¢, < 0 for each nonbasic variable, the last tableau gives the optimal solution. The sequence of points generated in the (x,x,) space is illustrated in Figure 43. Figure 4.3. Example of the big-M method. 4.3 THE BIG-M METHOD 187 ‘Anelysis of the Big-M Method ‘We now discuss in more detail the possible cases that may arise while solving the big-M problem. The original problem P and the big-M problem P(M) are stated below, where the vector b is nonnegative. Problem P: Minimize ex Subject to Ax =b x20 Problem P(M): Minimize x +M1x, Subject to Ax +x, =b x x20 Since problem P(M) has a feasible solution (say x = 0 and x, = b), then while solving it by the simplex method one of the following two cases may arise J. We shall arrive at an optimal solution of P(M). 2. Conclude that P(M) has an unbounded optimal solution, that is, z -» — 20. OF course, we are interested in conclusions about problem P and not P(M). ‘The following analysis will help us to draw such conclusions. Cose A: Finite Optimal Solution of PM). Under this case, we have two possibilities. First, the optimal solution to P(M) has all artificials at value zero, and second, not all artificials are zero. These cases are discussed below. Subcose Al: (x*, 0) is an optimal solution of P(M). In this case x* is an optimal solution to problem P. This can be easily verified as follows. Let x be any feasible solution to problem P, and note that (x, 0) is a feasible solution to problem P(M). Since (x*,0) is an optimal solution of, problem P(M), then ex* + 0 < ex +0, that is, ex" < ex. Since x* is a feasible solution of problem P, then x* is indeed an optimal solution of P. This case was illustrated by Example 4.6 above. Subcase A2: (x*, x2) is on optimal solution of P(M) and x3 % 0. If M is a very large positive number, then we conclude that there exists no feasible solution of P. To illustrate this case, suppose on the contrary that x was 158 ‘STARTING SOLUTION AND CONVERGENCE a feasible solution of P. Then (x, 0) would be a feasible solution of P(M) and by optimality of (x*, x7) we have ex" + MIX < ex +0 ex. Since M is very large, xf > 0 and x$ % 0, and since x} corresponds to one of the finite number of basic feasible solutions, the preceding inequality is impossi- ble (why?). Therefore x could not have been a feasible solution of P, A formal proof is left as Exercise 4.13, Example 4,7 (No Feasible Solutions) Minimize — x,-3x;+ x3 Subjectto xt x pt2y <4 0 is very large, then z, — ¢ < 0 for all nonbasic variables; that the simplex optimality criteria hold. Since the artificials x, and x, still remain in the basis at a positive level, then we conclude that the original problem has no feasible solutions. Cose 8: P(M) has an unbounded optimal solution, that is, z > — co. Suppose that during the solution of the big-M problem, the updated column yp i <0, where the index k is that of the most positive z, ~ c,- Then problem P(M) has an unbounded optimal solution. In the meantime, if all artificials are ‘equal to zero, then the original problem has an unbounded optimal solution. Otherwise, if at least one artificial variable is positive, then the original problem is infeasible. These two subcases are discussed in more detail below. Subease BI: — 6 = Maximum (z,— 6) > 0, < 0, ond all arifcials are equal to 2070. In this case we have a feasible solution of the original problem. Furthermore, since problem P(.M) is unbounded, then there is @ direction, (d), @,) > (0, 0) of the set ((X, X,) ! AX-+, = B,x > O,x, > 0) such that ed, + Md, < 0 (recall that this is the necessary and sufficient condition for unboundedness of P (M)). Since M is very large and d, > 0, then ed + Md, < 0 implies that dy = 0 and hence ed; < 0. Thus we have found a direction d, of the set (x: Ax = b, x > 0) such that ed, <0 (why?). This implies that problem P has an unbounded optimal solution, Example 48 Minimize —x,-x, Subjectto yy yd ax taytlxy- x, My ty My 420 Introduce the artificial variables x, and x, with cost coefficient M. This leads to the following sequence of tableaux. 7 1 0 -M -M | O 0 1 T 0 1 ord (On |[eett eal 0 ele 160 STARTING SOLUTION AND CONVERGENCE Multiply rows | and 2 by M and add to row 0. The most positive z ~ ¢ corresponds to x, and y, < 0. Therefore the bigeM problem is unbounded. Furthermore, the artificial variables x, and x, are equal to zero, and hence the original program has an unbounded optimal solution along the ray ((3, 0, 2, 0) + A(1, 1, 0,0): A > 0). 2, — G7 Maximum (2,~¢) > 0, ¥4 <0, ond not ll orifcials ore In this case we show that there could be no feasible solution of the original problem. To illustrate, suppose that the basis consists of the first m columns, where columns I through p are formed by real variables and columns p + 1 through m are formed by artificial variables. The corresponding tableau is depicted below. 4.3. THE BIG-M METHOD 161 Noting that ¢ = M for i= p+ 1,...4m then for j= m+ 1... we get -o-Seu+m( § n)-5 1) First note that 37, 4,94 < O for all = m+ 1... To show this, frst recall, that y, <0 and hence 37, <0 holds trivially. By contradiction, suppose that 3%.,,,¥, >0 for some nonbasic variable x. From Equation (4.1) and since Mis very large, then z, ~ ¢, is a large positive number, violating the definition of % ~ e (recall that 2 ~ c = Maximum (2; q) and y, <0). Therefore E7-p4.%) <0 for all j= m+ 1,...,n. Summing the last m— p ‘equations of the tbove tableau, we get Sane 3 a( 3 vj- 88 2) By contradiction, suppose that problem P has a feasible sohtion. Then = 0 £ al srial veils and fence =O forf=p-tlreesm Also 230 eer tar ta eae ears ee eee hand side of Equation (4.2) is < 0. But the right-hand side is positive since not all artificials are equal to zero. This contradiction shows that there could be no feasible solution of the original problem. ‘Note that we cannot conclude that the original problem is inconsistent if some z) = 6 > 0, y, <0 and not all artificials are zero. It is imperative that we use the most positive s, — ¢,. See Exercise 4.21. Example 4.9 Minimize x, Subject to xy-x,> 1 -xytm> 1 xy, 420 4 fo Figure 4.4. Empty Feasible 162 ‘STARTING SOLUTION AND CONVERGENCE This problem has no feasible solutions, as shown in Figure 4.4. We shall show that the optimal solution of the big-M problem is unbounded, Introduce the slack variables x, and x, and the artificial variables x, and xg ‘The last tableau indicates unboundedness since 2; ~ e > 0 and y, < 0. Note, however, that x, is positive, so we conclude that the original problem has no ae cas we Optima! is | Optimal is htcunded_| Setewe AZ Subcaw 8Y Subese 62 =. one | secu |f orgem || Peteted || Stone” Pinboendes | iconsuent Vee ee Figure 4.5. Analysis of the big M method. 4.4 THE SINGLE ARTIFICIAL VARIABLE TECHNIQUE 163 feasible solutions, This is also clear by examining the second row, which reads xs tx e dt ay t ty Since x, x4 > 0, then x; + x6 will always be positive, original system is inconsistent, The possible cases (Al, A2, BI, and B2) that may arise during the course of solving problem P(M) are summarized in Figure 4.5. indicating that the Comparison of the Two-Phato and Big-M Methods The big-M method has two important disadvantages in comparison to the two-phase method. First, in order to conduct the big-M method we must select a value for Mf. Without solving the linear program it is difficult to determine just how large M should be in order to ensure that the artificial variables are driven cout of the basis. In Exercise 4.18 we ask the reader to consider this question, The second major difficulty with the big-M method is that a large value of M will completely dominate the other cost coefficients and may result in serious round-off etror problems in a computer. 4.4 THE SINGLE ARTIFICIAL VARIABLE TECHNIQUE ‘Thus far we have described two methods to initiate the simplex algorithm by the use of artificial variables. In this section we discuss a procedure that requires only a single artificial variable to get started. Consider the following problem, Minimize ex Subject to Ax xo Suppose that we can partition the constraint matrix A into A = [B, N}, where B is a basis matrix, not necessarily feasible. This can certainly be done if the constraints were originally inequalities (that is, Ax > b or Ax < b) by utilizing the slacks as basic variabies. Index the basic variables from I to m. Multiplying the constraints by B~' (which would be =I for slack variables), we get Ixy + BUINay = 5 where b = B-'b and b + 0 (ib > 0, we have a starting basie feasible solution). 164 STARTING SOLUTION AND CONVERGENCE To this system let us add a single artificial variable, x,, with a coefficient of — 1 in each constraint: Inp + BOINay — 1x, = Now, introduce x, into the basis by selecting the pivot row r as follows: 6 Misinum(4) Note that 5, < 0, On pivoting in row r (that is inserting x, and removing s), we get the new righthand-side values y= ~ (> 0) — B(> 0) ‘Thus by entering x, and exiting x, we have constructed a basic feasible solution to the enlarged system (the one including the artificial variable). Starting with this solution, the simplex method can be used, employing either the two-phase or big-M method, Example 4.10 Minimize 2%, +3%2 Subjectto + 3 = 2yt x22 er) Subtracting the slack variables x, and x, and multiplying by —1, we get the ‘basic system -ucuty 273 Wyre xe? Appending a single aril variable x, with activity vector (1) to the il tableau ofthe phase I problem, we get the following tableau, 4.5. DEGENERACY AND CYCLING 165 mf oo 0 Ts x | 0 = 0 3 Sy | nce |e oe ee oe oie ed hes RBS eel 1 t r 0 0 3 xf o | t 7 7 ° 1 3 o | 3 ° 1 1 o 1 x a ‘The tableau above is ready for application of the two-phase method. Subsequent tableaux are not shown. An analysis of the two-phase method and the big-M method for the single artificial variable technique discussed in this section can be made similar to the analysis of Sections 42 and 4.3. The details are left to the reader in Exercise 428. 4.5. DEGENERACY AND CYCLING In Chapter 3 we developed the simplex method with the assumption that an initial basic feasible solution is known, We then described in this chapter how to obtain such a starting basic feasible solution. In Chapter 3 we also showed that the simplex method converges in a finite number of steps provided that the basic feasible solutions visited were nondegenerate. In the remainder of this chapter ‘we examine the problems of degeneracy and cycling more closely and give a rule that prevents cycling, and hence guarantees finite convergence of the simplex algorithm, We have seen that the simplex algorithm moves among basic feasible solu- tions umtil optimality is reached, or else unboundedness is verified. Since the number of basic feasible solutions is finite, the simplex method would converge in a finite number of steps, provided that none of the bases are repeated. Now suppose that we have a basic feasible solution with basis B. Further suppose that we have a nonbasic variable x, with , — ¢, > 0 (for a minimization problem). ‘Therefore x, enters the basis and xp, leaves the basis, where the index ris determined as follows: Minimum { Ya tsiem

You might also like