354

IEEE TRANSACTIONS ON RELIABILITY. VOL. 44. NO. 2. 1995 JUNE

A Note on State-Space Decomposition Methods for Analyzing Stochastic Flow Networks

- Maximum

flow, minimum cut, network

Summary & Conclusions - Consider a flow network with i single source s and single sink t with demand d > 0. Assume that the nodes do not restrict flow transmissionand the arcs have fdte random discrete capacities. This paper has two objectives:

1. It corrects errors in well-known algorithms by Doulliez

mand is satisfied (or network reliability), b) the probability that an arc belongs to a minimum cut which limits the flow below d , and c) the probability that a cut limits the flow below d; Other, standard notation is given in Information for Readers 2. It discusses the applicability of these procedures. The D&J algorithms are frequently referenced or used by & Authors at the rear of each issue. researchers in the areas of power and communication systems and appear to be very effective for the computation of the network Assumptions reliability when the demand is close to the largest possible maximum flow value. Extensive testing is required before the D&J 1. The nodes do not limit flow transmission. algorithms are disposed i favor of alternative approaches. Such n 2. The capacity of a r c j is a discrete r.v. Bj taking values testing should compare the performance of existing methods i a n ,..., with respective probabilities variety of networks including grid networks and dense networks in the set {bj(1),bj(2) b j ( n j ) } pi( 1),pj(2),. . . , p j ( n j ) ,where 0 I bj( 1) < b j ( 2 ) < ... < of various sizes.

capacity state space ( v l , ...,vu): vector of capacity levels with vi E { 1,* ,nj} value of maximum s - t flow when the capacities are at levels v a (directed) s - t cut C j E bj(vi) : capacity of a (directed) s - t cut C demand placed at the sink t Pr { L(B) 1 d } : network reliability Pr{arc j is in a minimum cut and L( B) < d} Pr{cut C is minimum and L(B) < d} { a ; p } : subset of Q with rectangular structure (a1,. ..,a,):lower limiting element of R (PI,,..,Pa): upper limiting element of R.

1. INTRODUCTION A primary objective in the design of a transmission network for a single commodity is the establishment of flow between a source node and a sink node that exceeds a specified value d. If such a feasible flow does not exist, the detennination of the locations of the bottlenecks that limit the flow of commodity is an important problem in network design. The solutions to these problems are well known in the case of deterministic capacities. However, the capacities are often random variables in practice and the probability that a feasible flow exists is a meaxure of system performance, often called the network reliability. Another measure is the probability that a set of arcs (cut) limits the value of maximum flow below d. Notation

bj(nj) < 00. 3. The r.v.s B1,...,B, are statistically independent. 4. Every arc belongs to at least one s - t cut.

A point x in the capacity state space n is defined as an a-tuple ..., of values x = (bl(vl),b2(v2), b,(vu))where the index vj takes on values from 1 to nj. For notational convenience, the index vj is also used to designate the value bj(vj) itself so that the state point x is also denoted as v = ( vl,. ..,vu). The cut C is minimum if Z ( C , v ) = L ( v ) . Suppose that demand d is placed at the sink t. Then g (d) is the network reliability. A cut C is called critical (with respect to the demand d) if it is minimum cut while L( B) < d, Similarly, an a r c j is called critical if it belongs to a critical cut. This paper discusses the computation of The network reliability g ( d ) . The probability h 0) that arc j is critical. We refer to this probability as the criticality index of j . The probability q ( C) that a cut C is critical. We refer to this 4 probability as the criticality index of C.

G 37,

a
s, t Bj

B pj(l)

(rt,a,s,t): a network set of nodes {1, ..., set of arcs a}: A state v is called operating if flow of value d can be sup[source, sink] node plied from s to t when the capacities are (b,( vl), ...,b, ( vu)) and (random) capacity of arc j : Bj E {bj(1),bj(2),..., failed if no such feasible flow exists. If Q + is the set of bj(nj) 1 operating states, theng(d) = Pr{B E Q } A set R E n is . V I , . .. A ) called operating or failed if all states in R are classified as Pr{Bj=bj(l)}, 1 E {1,..., nj} operating or failed respectively. 0018-9529/95/$4.00 01995 IEEE

ALEXOPOULOS: STATE-SPACE DECOMPOSITION-METHODS FOR COMPUTING PERFORMANCE MEASUREMENTS

355

Since the evaluations of g ( d )and h 0) are NP-hard problems (Ball [5], Alexopoulos & Fishman [3]), no polynomial algorithm is known to exist for computing them. Exact methods for computing performance measures for flow networks include Doulliez & Jamoulle [6], Evans [7], Kulkarni 8z Adlakha [9], Lee [lo], Rueger [1 11, Shogan [121 while Monte Carlo methods are described in Alexopoulos [2], Alexopoulos & Fishman [3, 41, Fishman & Shaw [8]. This paper has the following two objectives: 1. It identifies and corrects errors in the algorithms by Doulliez & Jamoulle [6] for computing the probabilities g ( d ), hO), do; 2. It discusses the applicability of these procedures. 4 The methodology of Doulliez & Jamoulle [6] is based on iteration and, in short, evaluates g (d) as follows: At each iteration a subset of Q is partitioned into non-overlapping operating, failed, and undetermined sets. An undetermined set is one whose states cannot be classified at the current iteration as operating or failed. The existing undetermined sets are used as input to subsequent iterations. The method continues in the same fashion until all the undetermined sets have been considered. These sets are a-dimensional discrete rectangles in the sense that each such set, say R, has limiting elements a s a[R] and /3 = P[R]such that every integer vector vj with aj 5 vj IPi for a l l j belongs to R. This rectangle can then be denoted by R = {a$}.Since the capacities are statistically independent, then, Pr{R}

ajwe set vJ: = aj. Then for each a r c j = (k, 1 ) with ! > aj v we identify a minimum s -t cut, say C, containing j ( C = C for j E C) define vJ: to be the smallest index m E {cyj,aj+ and 1, ...,v?} such that b j ( m ) CiEcly) bi(Pi) 2 d. Clearly, every state v E R with vj < is failed because the cut C has capacity Z ( C,v) < d. For j C, the cut C is a minimum cut in the network resulting from G by adding two arcs (s,k) and (At) with infinite capacities and contains arcj because: a) the arcs in G have finite capacities and b) j belongs to at least one s - t cut. Now for j E a define:

vi

Fjl= {v: v E Randvj < vi}

v = {v: v E R, vj < $and

vk 1 v i for all k E

a}.

Clearly, F, = 0 if vJ: = aj and ujr = 0 if vJ: = v Note that ; the sets are failed and that the states in Uj= ujcan not be classified as operating or failed with the existing flow. Since vj overlap, we partition this union into the rectangles:

= {(v? V~-l,Vj,Vj,l ,...,vi); ,...,

* .

..,Pj - 1,

1 Pj+ 1

9.

.., P a ) I ,

Similarly, we partition UT= Fj into the rectangles:

Pr{B E R} =

n
a
j=1

6 pj(Z).
l=a,

= { (vl,...,vj-l,ofi,~j+l,... t a u ) ;

Section 2 describes the modified algorithm for evaluating the network reliability g ( d ). Section 3 contains the corrected algorithm for computing h 0) and discusses the evaluation of dC).

(PI

9..

.J j - I,vj.-

1 p P j + 17..

.P P u ) } .

2. COMPUTING THE NETWORK RELIABILITY

Consider an undetemined rectangle R C Q with lower and upper limiting points a = (a1 a,) and P = (PI,...,Pa) ,..., respectively. This set is partitioned by determining indices v : and vJ: for each arc j such that the states v E R for which vy Ivj IPi, j E a are operating and the states v E R for which vj < vJ: for some j are failed. These indices are obtained as follows: Create a fictitious demand node T, add the arc e = (t, T ) with capacity d and determine a maximum s - T flow = ~ .. . ) with capacities P for the arcs in a. If the , value of this flow is less than d, then none of the states in R can satisfy the demand at node t making R a failed rectangle. Otherwise, for each arc j E (2 define v = min{v: aj II I : Pj and bj ( 1 ) 2 $} . Obviously, all states v with vy Ivj IPi, j E (2 form an operating rectangle W . We now show how the indices vJ: can be obtained: Using the existing flows we compute a maximum s - t flow with value L (6) and a minimum cut C. For each arc j with v = :

Clearly, the rectangles W, q, 4 partition R. Now, each nonempty undetermined rectangle Uj remains to be decomposed similarly to the original set R. Procedure RELIABILITY summarizes the decomposition method for calculating g ( d ). The set r contains all the remaining undetermined rectangles. At termination the bounds gL( d ) and gu( d ) are equal to g ( d ). Note that only the boundary points of each set must be stored. Efficient algorithms for maximum flow evaluations are described in Ahuja, Magnanti, Orlin [l].
2.1 Procedure RELIABILITY

E&

e,

1. Start with the rectangle R = Q . Set I = 0, g L ( d ) = 0 , and g u ( d ) = 1. 2. Let a and @ denote the limiting points of R. If L ( @ ) < d, set I = I\{R} and go to step 5. Otherwise, decompose R into an operating rectangle W, failed rectangles 5, and undetermined rectangles 3. Set g L ( d ) = g L ( d ) Pr{W}. 4. F o r j E a: If 4 z 0 , set g u ( d ) = g u ( d ) - Pr{F/}. If # 0 , set r = r U ( q } . 5 . If r # 0 ,choose a set from r, call it R, set I = I\ {R}, and go to step 2. 6. End with g ( d ) = g L ( d ) = g u ( d ) .

q.

356

IEEE TRANSACTIONS ON RELIABILITY. VOL. 44. NO. 2. 1995 JUNE

The application of this procedure to the network in Doulliez

experiment.
2.2 Remarks on the D&J Algorithm

Doulliez & Jamoulle [6: top of p 551 proposed the following procedure for deriving the indices v Given the feasible , : flow fo,for each arc j = (k,Z) E Q. let Lj denote the value of maximum flow that can be transmitted from node k to node Z in addition to the existing flow without using this arc. If Lj 2 f any capacity aj 5 I I for arc j satisfies the demand d when all other capacities are fixed at &, k # j and then vi = cyj. Otherwise, arc j is in a minimum cut that blocks the value of maximum s - t flow below d and every state v E R with bj(vj) < fjo - L~is failed. Unfortunately, the last statement is not valid in general as the following example demonstrates. We should also mention that the statement remains false when Lj is the value of a maximum s-Z flow or a maximum k - t flow.

Y,

measures in several models (such as maximum flow, shortest path, and minimum spanning tree models), this claim appears to be unjustified. If procedure RELIABILITY is executed entirely (until r becomes empty), then I recommend that r be maintained by means of a stack (or LIFO list). Using this strategy during the evaluation of unconditional measures, we never had to maintain more than 15 - 20 sets at any time for networks with up to 20 nodes. If high capacity levels I are associated with large probabilities pi( I ) and the objective is a quick derivation of tight bounds, then I recommend that these collections be implemented as heaps with nodes corresponding to rectangles and node weights equal to the negatives of the rectangle probabilities. Obviously, each procedure removes the set corresponding to the root of its respective heap. In the special case where each component has only two states, an operating state and a failed state, an undetermined set can be stored by recording only the set of operating components and the set of failed components.

3. COMPUTING CRITICALITY INDICES

Our method for computing the criticality indices is based on the partition of the failed rectangles produced by procedure RELIABILITY. Indeed, consider a failed rectangle F = {a!;@} generated in step 2 and, with capacities P, find a maximum s - t flow fand a minimum cut C. Then we use this flow to identify indices v?, j E Q. such that the states in the classz$ed rectangle:

Figure 1. Network for Remark 1

The network in figure 1 has s = 1, t = 4, demand d = 6, and arcs numbered as 1, 2, ..., 5. Suppose that the capacity of each arc takes values from the set (1, 2, 3, 4, 5). This assumption implies bj(I) = 1 for every j E Q. and I = 1,...,5. Consider the rectangle R with limiting points a! = (1, 1, 1, 1, 1) and p = (5, 3, 2, 3, 4). The s - t flowf' = (3, 3, 1, 2, 4) has value d and, clearly, vo = (3, 3, 1, 2, 4). Consider the arc 2 = (1, 3). The maximum flow that can still go from node 1 to node 3 without using arc 2 has value & = 1 < f l ( 1 unit through arcs 1 and 3) and thenfi - & = 2. Now, reduce the capacity of arc 2 tof2 - & - 1 = 1 2 a2 and leave the remaining capacities at &, k # 2. The flow (5, 1, 2, 3, 3) has value d = 6 and the capacity state (5, 1, 2, 3, 4) E R is operating. The same erroneous statement reappears in Shogan [12: section V.C, step b]. To the best of my knowledge, this error has not been pointed out previously. The D&J method has also been criticized for having to maintain large lists of undetermined sets; see for example Lee [101 and Rueger [111. Based on computational experience with state space partitioning methods for computing a variety of

have an identical set S of critical arcs. This set consists of the arcs in C and the arcs j C with fj = bj(fij)for which a minimum s - t cut containingj (found by the method in section 2) has capacity equal to L ( ) Then Pr {D} part of h 0') only 0. is for j E S. If s = C, we set vjo = cyj, j E S; otherwise, we set vjo = Pj, j. E S. Now for each arc j Q S, we identify the index k = mm{l: cyj II I and bj(Z) 2 fj} and set vo = k if j does not enter a minimum cut at level k or vj = min{k+l, Pj} otherwise. The partition of F ends with the determination of the unclussijed rectangles:

Procedure CRlTlCAL summarizes the preceding discussion. At termination, the bounds h L ( j )and h&) are equal to h 0'). The set A contains the unclassified rectangles.
3.1 Procedure CRITICAL

tr ..., 1. S a t with the collection A = {Fl, FL} of failed rectangles determined by procedure RELIABILITY. Set hLU) = 0 andhu(j) = 1 - g ( d ) for a l l j E (2.

ALEXOPOULOS: STATE-SPACE DECOMPOSITION-METHODS FOR COMPUTING PERFORMANCE MEASUREMENTS

357

2. If A = 0 ,go to step 5 . Otherwise, remove a rectangle R = {a$} from A and set A = A\{R} . Decompose R into a classified rectangle d and unclassified rectangles Qj. Let S be the set of critical arcs for the state P. 3. F o r j E S: Set hL(j) = h L ( j ) Pr{D}. For j L S: Set h&) = hu(j) - Pr{D}. 4. F o r j E a: If Qj # 0 , set A = A U { Q j } . Go to step 2. 5 . End.

REFERENCES
[ l ] R.K. Ahuja, T.L. Magnanti, J.B. Orlin, Network Flows: Theory, Algorithms, and Applications, 1993; Prentice Hall. [2] C. Alexopoulos, Computing criticality indices of arcs and the mean maximum flow value in networks with discrete random capacities, Technical Report J-93-01, 1993; School of Indl & Systems Engg, Georgia Institute of Technology. [3] C. Alexopoulos, G.S. Fishman, Characterizing stochastic flow networks vol21, 1991, pp 775-798. using the Monte Carlo method,.Networh, [4] C. Alexopoulos, G.S. Fishman, Sensitivity analysis in stochastic flow networks using the Monte Carlo method, Networks, vol23, 1993, pp 605-62 1. [5] M.O. Ball, Computational complexity of network reliability analysis: An overview, IEEE Truns. Reliability, vol R-35, 1986 Aug, pp 230-239. [6] P. Doulliez, E. Jamoulle, Transportation networks with random arc capacities, R.A.I.R.O., vol 3, 1972, pp 45-60. [7] J.R. Evans, Maximum flow in probabilistic graphs: The discrete case, Networks, vol 6, 1976, pp 161-183. [SI G.S. Fishman, T.Y. Shaw, Evaluating reliability of stochastic flow networks, Probability in Engineering & Informutionul Sciences, vol 3, 1989, pp 493-509. [9] V.G. Kulkami, V.G. Adlakha, Maximum flow in planar networks with exponential arc-capacities , Communications in Statistics Stochastic Models, vol 1, 1985, pp 263-289. [lo] S.H. Lee, Reliability of a flow network, IEEE Trans. Reliability, vol R-29, 1980 Apr, pp 24-26. [ l l ] W.J. Rueger, Reliability analysis of networks with capacity constraints and failures at branches and nodes, IEEE Trans. Reliability, vol R-35, 1986 Dec. DD 523-528. [12] A.W. Shogan, Modular decomposition and reliability computation in stochastic transportation networks having cutnodes, Networks, vol 12, 1982, pp 255-275.

Observe that procedure CRITICAL can be time consuming. In fact, it starts with A containing all failed rectangles determined by procedure RELIABILITY. 3.2 Comments on the D&J Algorithm Doulliez & Jamoulle [6: bottom of p 551 considered the evaluation of h U),j E a and proposed the following decomposition of a failed set F: Find a maximum flowf and a m n i i cut C with capacities at levels 0 and set vj = aj f o r j E C and vy = min(1: aj I1 I Pjandbj(l) > A } f o r j L C. Thenthey claimed that only arcs in C are in minimum cuts for each state in the resulting set d. This statement is wrong because: a) an arc may belong to an alternative minimum cut for the state 0, b) j may enter a minimum cut when its capacity is or at level vj < Pj. This error appears to affect the computation of the criticality index of a cut Cas Doulliez & Jamoulle [6: top of p 561 state that this probability can be computed by summing the probabilities of all sets d with C as the minimal cut. Unfortunately, the approach in the previous paragraph may fail to identify C as a minimum cut. Procedure CRITICAL does not seem to be applicable to the computation of the criticality indices of al cuts l because the identification of all cuts with arcs in the set S is a hard problem. It appears that a separate decomposition procedure must be executed for computing the criticality index of each cut. Another alternative is the simultaneous estimation of the criticality indices of several cuts by using the method in Alexopoulos & Fishman [4]. ACKNOWLEDGMENT This research was supported in part by the Air Force Office of Scientific Research under contract F49620-93-1-0043.

__

AUTHOR
Dr. Christos Alexopoulos; School of Industrial & Systems Engg; Georgia Institute of Technology; Atlanta, Georgia 30332-0205 USA. Internet (e-mail): christosQisye.gatech.edu Christos Alexopoulos received his BSc (1982) in Mathematics from the Aristotle University in Thessaloniki, and his PhD (1988) in Operations Research from the University of North Carolina at Chapel Hill. Since 1988, he has been with the School of Industrial & Systems Engineering at the Georgia Institute of Technology, where he holds the rank of Associate Professor. His research interests are in applied probability, statistics, and optimization of stochastic systems. His recent work involves problems related to the optimal design of telecommunication & transportation systems. Since 1993 he has been an Associate Editor for the Simulation Department of the IIE Transacrions. Manuscript received 1993 November 26. IEEE Log Number 92-15602