Sie sind auf Seite 1von 7

Proc. AAAI-93 Spring Symp. on AI and NP-Hard Problems, Standord, CA, March 23-25, 1993, pp.160-6.

Truncated Branch-and-Bound: 
A Case Study on the Asymmetric TSP
Weixiong Zhang
Computer Science Department
University of California, Los Angeles
Los Angeles, CA 90024
zhang@cs.ucla.edu
ABSTRACT We further extend the idea of adapting BnB for ap-
proximate solutions to large combinatorial optimization
Branch-and-bound (BnB) is a general problem problems, and compare it to other approximation al-
solving paradigm, especially useful for nd- gorithms. In particular, we present an approximation
ing optimal solutions to most NP-hard com- method based on BnB, called truncated BnB, on the
binatorial problems. In fact, BnB can also asymmetric traveling salesman problem. Given n cities
be adapted as an ecient approximation al- f1; 2; :::; ng and a cost matrix (ci;j ) that de nes the cost
gorithm, providing a promising new alterna- between each pair of cities, the traveling salesman prob-
tive in practice. We show this by an ap- lem (TSP) is to nd a minimum-cost tour that visits
proximation scheme for the asymmetric trav- each city once and returns to the starting city. When
eling salesman problem (ATSP). Our method the cost from city i to city j is not necessarily equal to
is BnB truncated at di erent stopping points, that from city j to city i, the problem is the asymmetric
depending upon the total computation avail- TSP (ATSP). We choose the ATSP as our benchmark
able. Experiments indicate that this method problem because it is an important problem, and many
is superior to a local search algorithm that is NP-complete problems can be formulated as ATSPs,
the best existing approximation algorithm for such as vehicle routing, workshop scheduling, computer
the ATSP. wiring, etc [14]. Although we present the approxima-
tion method on the ATSP, our algorithm can be simply
applied to other problems as well.
1 Introduction and Overview We empirically compare our approximation method to
Many practical problems, such as planning and schedul- a local search algorithm of [9], the best existing ap-
ing, are computationally intractable, meaning that they proximation algorithm for the ATSP we found in the
cannot be solved exactly with limited computation time literature. Given a non-optimal complete tour, a local
and memory. In reality, we often do not need opti- search algorithm repeatedly improves the tour by local
mal solutions, but rather good ones that can be found perturbations [16, 9, 8]. In our experiments, we consid-
quickly. Therefore, approximation algorithms are of ered many ATSPs with di erent cost matrix structures.
considerable practical importance. The cost matrices we used include random matrices,
matrices with the triangle inequality, random matrices
Branch-and-bound (BnB) [15, 2, 13] is a general with ci;j from f0; 1; :::; i  j g, matrices converted from
problem-solving paradigm for nding optimal solutions no-wait workshop scheduling for 4 machines, which is
to NP-complete [6] combinatorial optimization prob- known to be NP-complete [20], and matrices from real
lems. Although it was originally developed for optimal problems [18]. Our empirical results indicate that the
solutions, it can, in fact, be adapted for nding approx- average tour quality from our method is better than
imate solutions as well. The idea, which is well-known that from the local search algorithm, with less or com-
in operations research, is to terminate the execution of parable computation time. Thus, the new approxima-
BnB whenever the limits of computational resources are tion algorithm provides a promising alternative in prac-
met. The previous research on this idea is very prelim- tice.
inary. It has only been applied to job-shop scheduling Section 2 discusses the approximation algorithm for the
problems with two machines and up to 50 jobs [1, 12], ATSP, and Section 3 presents the experimental results.
and simulated on a tree model [7]. Our discussion is in Section 4, and conclusions appear
 The research was supported by NSF Grant #IRI-
in Section 5. For interested readers, the detailed de-
9119825, a grant from Rockwell International, and a GTE scription and experimental results of the local search
fellowship.
algorithm of [9] can be found in [25]. 1 4
2 32

2 Truncated BnB on the ATSP E={(2,3)}


5 6
E={(3,2)}
I={} I={(2,3)}
2.1 Introduction: Branch-and-Bound 1 4
A B
2 3 1 4
Since our new approximation method is based on 2 54 3
the branch-and-bound (BnB) subtour elimination al-
6 3
6 5
gorithm [2] for optimally solving the ATSP, we rst E={(3,2),(6,2)} E={(3,2),(3,6)}
describe this algorithm. I={(2,3)} I={(2,3),(6,2)}

The BnB subtour elimination algorithm uses the as- C D


signment problem as a lower-bound function. Given n 2 3 6 6 2 3
cities and a cost matrix (ci;j ), the assignment problem 4 1 5
5
4 1 5
3
(AP) [17] is to assign to each city i another city j , with
ci;j as the cost of this assignment, such that the to-
tal cost of all assignments is minimized. The AP is a Figure 1: BnB subtour elimination on the ATSP.
generalization of the ATSP without the restriction of
a single tour, allowing collections of disjoint subtours.
Speci cally, the solution of the AP is a permutation of The process of subproblem selection and decomposi-
the n cities which is either a complete tour, or a collec- tion repeats until all active subproblems have AP costs
tion of subtours, and can be computed in O(n3) time greater than or equal to the cost of the best tour ob-
[17]. If the AP solution happens to be a complete tour, tained so far, or no active subproblems remain. More
it is the solution to the ATSP as well. speci cally, with an active list A, containing all active
subproblems, a variable BT , representing the best tour
Figure 1 is a simple example of BnB subtour elimina- obtained so far, and its cost U , the process of BnB
tion. The edge costs and the calculation of the AP costs subtour elimination is as follows.
are not shown. The AP solution to the original ATSP
contains two subtours that are in the oval at the root of 1. (Initialization) Set the active list A = fX0 g, the
the tree. Assume that the cost of the AP solution is 2, best tour BT = ;, and its cost U = 1, where X0
as shown in the black box. The subtour 2 3 2 is cho- is the original problem to be solved.
sen to be eliminated. The problem is then decomposed
into two subproblems, subproblem A with cost 4, and 2. (Subproblem selection) If A = ;, stop. BT is an
subproblem B with cost 3. The lowest-cost subproblem, optimal tour with cost U .
B , is selected for decomposition next. By breaking the Otherwise, select an X 2 A, and remove it from
subtour 2 3 6 2, subproblem B produces subprob- A.
lems C and D. D is the next lowest-cost subproblem 3. (Lower-bounding) Let AP be the assignment prob-
among all undecomposed ones, A, C and D. The AP lem of X , and ApT be its solution, whose cost is
solution to D is a complete tour. The algorithm termi- C.
nates with the complete tour in node D, and its cost 3, If C  U , go to 2. (This subproblem is pruned.)
since subproblems A and C have costs greater than 3, If C < U and ApT is a single tour, BT = ApT ,
and hence cannot generate a better solution than the U = C , and go to 2. (This is a better tour.)
one in node D (explained below). Otherwise, go to 4.
BnB subtour elimination rst solves the AP for all n 4. (Decomposition) Expand X , generating subprob-
cities. If the solution is not a single tour, then the prob- lems X1 ; X2; :::; Xt, with the costs of their AP so-
lem is decomposed into subproblems by excluding dif- lutions C1; C2; :::; Ct, respectively.
ferent arcs in a subtour, which eliminates the subtour. A = A [ fXi jCi < U; i = 1; 2; :::;tg. Go to 2.
A subproblem has some excluded arcs that are forbid-
den from its solution, and some included arcs that must The original problem has empty excluded and included
be present in its solution, and is used to decompose the sets. A newly generated subproblem rst inherits its ex-
problem space into mutually exclusive components. A cluded and included sets from its parent, and adds one
subproblem is then selected from all subproblems that new excluded arc to its excluded set, and new included
have been generated but not decomposed, which are arcs, if any, to its included set. Since a subproblem
called active subproblems. If the AP solution to this has more excluded and included arcs, and thus is more
selected subproblem is a complete tour, and its cost is constrained than its parent, less choices of arcs are left
less than the cost of the current best tour found, then for the AP to the subproblem than for the AP to its
the current best tour is replaced by the AP solution to parent. Therefore, the AP cost is monotonically nonde-
this subproblem, and another subproblem is selected. creasing, in the sense that the AP costs of subproblems
along a path starting at the root can only increase.
(j)
Two common strategies can be used to select a subprob- i
lem from all active ones. The best- rst strategy chooses
a subproblem with the minimum AP cost among all ac- (i) j
tive subproblems. The depth- rst strategy selects a sub-
problem which was most recently generated, and has a Figure 2: An (i; j )-patching operation.
minimum AP cost among its siblings. One important
feature of the best- rst strategy is that all subproblems
selected have costs less than or equal to the cost of the tree links between a node and its children represent the
optimal solution to the entire problem. The best- rst decomposition of the parent, and leaf nodes are sub-
strategy generates the fewest subproblems, among all problems whose AP solutions are complete tours.
strategies using the same monotonic cost function [4].
To this end, it needs to store all active subproblems in
A, whose size is usually exponential in the maximum 2.2 Truncated BnB for Approximate
search depth. Because of this exponential space require- Solutions
ment, the best- rst strategy is only applicable to small The BnB subtour elimination algorithm can also be
problems in practice. In contrast, the depth- rst strat- augmented by an upper-bound function. This func-
egy only maintains the subproblems on the path from tion is applied to a subproblem whose AP solution is
original problem to the subproblem that is now being not a single tour, and constructs a single tour based on
considered, plus their immediate siblings, and thus the the subproblem. The tour cost thus obtained is greater
space used is only linear in the maximum search depth. than or equal to the cost of an optimal tour for the sub-
It may, however, decompose subproblems whose costs problem, and thus gives an upper bound on the optimal
are greater than the cost of the solution to the initial tour cost to this subproblem.
problem.
Which subtour to eliminate, and how to expand a prob- Karp and Steele's patching algorithm [10, 11] can serve
lem constitute the decomposition rule. One such rule is as an upper-bound function. This algorithm also uses
due to Carpaneto and Toth [3]. A subtour is chosen the AP algorithm, plus a post-processing operation.
such that the number of newly generated subproblems Let i and j be two cities that appear in two disjoint
is as small as possible, in order to reduce the total num- subtours, as shown in Figure 2. The (i; j )-patching op-
ber of subproblems generated. Since excluding one arc eration joins the subtours into one by inserting the arcs
of a subtour is sucient for eliminating the subtour, a (i;  (j )) and (j;  (i)) and deleting the arcs (i;  (i)) and
subtour with the minimum number of free arcs, ones (j;  (j )), where  (i) is the successor of i in the subtour.
neither included nor excluded in the corresponding AP This operation changes the total cost by
solution, is selected for elimination. Subproblems are (; i; j ) = ci; (j ) + cj; (i) ci; (i) cj; (j ):
generated in such a way that they are mutually exclu-
sive, so that there are no duplicate subproblems. Let I (; i; j ) is called the cost of the (i; j )-patching oper-
be the included set and E the excluded set of the sub- ation. Using this operation, the patching algorithm
problem to be expanded, and let X = fx1; x2; :::; xtg works as follows. First solve the AP of the n cities,
be the arcs that are neither included nor excluded from resulting in a permutation  . If  is a single tour, it is
the subproblem. The k-th newly generated subproblem also a solution to the ATSP. If  is not a single tour,
then has included set Ik and excluded set Ek satisfying select two subtours with the largest numbers of arcs,
 and join them by an (i; j )-patching operation that has
Ek = E [ fxk g; minimum cost. Recursively apply the above process to
Ik = I [ fx1; x2; :::; xk 1g; the remaining subtours until  becomes a tour. This al-
for k = 1; 2; :::; t, where [ stands for the set union oper- gorithm has the same complexity as the AP algorithm,
ation. Since xk is an excluded arc of the k-th subprob- which is O(n3 ).
lem, xk 2 Ek , and it is an included arc of the k + 1-st By applying an upper-bound function to every gener-
subproblem, xk 2 Ik+1 , a tour obtained from the k-th ated internal node of search tree, best- rst search al-
subproblem does not contain arc xk , but a tour ob- ways has a feasible solution available. The quality of the
tained from the k + 1-st subproblem must have arc xk . best solution obtained so far increases with the num-
Thus a tour from the k-th subproblem cannot be gen- ber of nodes generated up to that point. Moreover, the
erated from the k + 1-st one, and vice versa. In other upper-bound function signi cantly impacts depth- rst
words, all generated subproblems are mutually exclu- BnB. Although depth- rst search may nd feasible so-
sive. Therefore, the state space of the ATSP under lutions during the search without an upper-bound func-
Carpaneto and Toth's decomposition rule can be rep- tion, the quality of the best solution found so far can
resented by a search tree, in which the root represents be improved by the upper-bound function. In addition,
the initial problem, non-root nodes are subproblems, an upper bound on a node can reduce the overall up-
per bound for searching the subproblem space, which in has the minimum AP cost as the next current problem,
turn reduces the number of subproblems decomposed. and repeats the above process. Overall, this algorithm
One common feature of best- rst search and depth- rst searches along one particular path from the root of the
search augmented by an upper-bound function is that search tree to a leaf node. This stopping criterion is
there exists a feasible solution at any point during the also called nonbacktrack BnB and was implemented on
search. Therefore, when augmented by an upper-bound two-machine scheduling problems [1, 12].
function, BnB is an anytime algorithm, meaning that
it can stop at any time and provide a solution whose
quality improves with the total amount of computa-
3 Experimental Results
tion time used. The BnB subtour elimination algorithm We compared TBnB with the local search algorithm
augmented by the patching algorithm is an anytime al- of [9], the best existing approximation algorithm for
gorithm for the ATSP. the ATSP we found in literature. The local search al-
Our approximation scheme is very simple. It executes gorithm repeatedly improves a given initial tour, which
the BnB subtour elimination algorithm, augmented by may not be optimal, by local perturbations. As a start-
the patching algorithm which is applied to a generated ing point, we rst examined all polynomial-time tour
subproblem whose AP solution is not a complete tour, construction algorithms we found in the literature, in-
and terminates after an optimal tour is found, or when cluding nearest neighbor [5, 8], nearest insertion [5, 8],
the total available computation time runs out. The best greedy [5, 8], repeated assignment [5] and Karp and
tour thus obtained is then used as an approximate solu- Steele's patching algorithm [10, 11]. The best algo-
tion. Since this method is simply BnB subtour elimina- rithm was then used to build the initial tour for the
tion, truncated at di erent stopping points depending local search1 , so that our experimental comparison was
upon the total available computation, we call this ap- set on a fair basis. From our experiments, the best
proximation scheme truncated BnB, or TBnB for short. polynomial-time initial tour construction algorithm was
The patching algorithm and the original BnB subtour the patching algorithm.
elimination algorithm are special cases of TBnB. If the We used the following di erent cost matrix structures:
algorithm stops without decomposing the original prob- (a) random matrices with ci;j from f0; 1; 2; :::231 1g;
lem only after applying the patching algorithm, it de- (b) matrices with the triangle inequality; (c) random
generates to the patching algorithm. If the algorithm matrices with ci;j from f0; 1; :::; i  j g, which are known
terminates after all subproblems whose AP costs are to be more dicult for many methods based on AP
less than the current best tour cost are decomposed, algorithms [19]; (d) matrices converted from no-wait
it becomes the original BnB subtour elimination algo- workshop scheduling for 4 machines; and (e) matri-
rithm. If the algorithm stops in the middle of these ces from some actual problems encountered in industry
two extreme cases, it provides a tour whose cost lies in [18]. We use a large cost range for the rst structure be-
between the tour cost from the patching algorithm on cause there exists an average-case complexity transition
the original problem, and the optimal tour cost. The for BnB subtour elimination [23, 24, 25]. In particular,
tour cost decreases with the total computation time, the ATSP is relatively easy to solve when the range r of
and converges to the optimal tour cost if the total com- intercity costs is small, with respect to the number of
putation time is unconstrained. cities n. Since r  n, the problem instances we chose
We use the depth- rst search strategy, and choose a are also relatively dicult for our new approximation
particular stopping point to implement TBnB. It rst method, since it is based on BnB subtour elimination.
solves the AP of the n cities. If the solution is a tour, For problems with cost matrices satisfying the triangle
it terminates with this tour as the nal solution, which inequality, we rst generated random matrices as above,
is optimal. If the solution is a set of subtours, it con- and then used a closure algorithm to enforce the trian-
structs a complete tour by the patching algorithm, and gle inequality, ci;j  ci;k + ck;j , for all i; j; k 2 (ci;j ).
takes its cost as an upper bound . It then generates For the ATSPs with ci;j in f0; 1; :::; i  j g, we chose
all children of the current subproblem using Carpaneto ci;j independently and uniformly from f0; 1; :::;i  j g.
and Toth's decomposition rule. It ignores those children For the ATSPs corresponding to no-wait scheduling, we
whose AP costs are greater than or equal to , since so- rst generated random scheduling problems, by setting
lutions from them will not be better than the best tour the required processing time of a job on a machine to
obtained so far. The remaining newly generated child be an integer31independently and uniformly chosen from
problems (not all active subproblems) constitute a set f0; 1; 2; :::; 2 1g, and then converted the scheduling
R. If R is empty, it stops with the best tour obtained so problems into ATSPs using the method suggested in
far as an approximate solution. If the AP solution of a [22].
problem in R is a complete tour, it terminates with this
tour. If none of the AP solutions of the subproblems in 1
Detailed descriptions of our experiments on the
R is a complete tour, it takes the subproblem in R which polynomial-time tour construction algorithms and the local
search algorithm are in [25].
We implemented the initial tour construction algo- Both algorithms were also tested on constructed ATSPs
rithms, local search algorithm, and our new approxima- that are dicult for the local search algorithm [21]. For
tion method in C and ran them on a Sun4/sparc-460 an ATSP with n = 6k cities, there is an optimal tour
with 32 Mbytes of memory. The problems were in the with cost 0, and (k 1)! local optimums that have arbi-
range of 100 to 1000 cities, in 100 city increments. For trarily large costs. Not surprisingly, the local search al-
100-city to 500-city instances, all results are averages of gorithm stops at a local minimum which is not a global
1000 trials each, and for 600- to 1000-cities, the results optimal, while TBnB nds the optimal tour by expand-
are averages of 100 trials each. ing only the original problem and generating four sub-
Figure 3 summarizes our results. Since the local search problems.
algorithm and TBnB use the patching algorithm as
starting point, we also include the performance of the
patching algorithm for comparison. The horizontal axes
4 Discussion
of the gures are the numbers of cities. For the gures The good performance of TBnB is due primarily to
in the left column, the vertical axes are the relative er- the AP lower-bound function. As observed in pre-
rors of tour costs with respect to the AP lower bounds, vious research [2] and in our experiments, the AP
which is less than the actual relative errors. For the lower-bound function is very e ective. In our exper-
gures in the right column, the vertical axes are the iments on the random ATSP with the elements of
CPU times of the algorithms. The relative tour error cost matrices independently and uniformly chosen from
of TBnB is less than 1:5% for 100-city problems, and f0; 1; 2; :::; 231 1g, the AP cost is 99.090% of the ATSP
decreases to less than 0:15% for 1000-city instances on cost on the average for 100-city instances, 99.816% for
all cost matrices. On random matrices and matrices 500-city instances, and 99.916% for 1000-city instances.
with ci;j chosen from f0; 1; :::;i  j g, the relative tour
errors of TBnB are signi cantly smaller than those of The eciency of TBnB is also due to the ATSP search
the local search, and the execution time of TBnB is tree under BnB subtour elimination, which has inter-
competitive. On matrices with the triangle inequality, nal nodes whose AP solutions are not complete tours.
TBnB signi cantly outperforms the local search, both Thus BnB searches the space of all city permutations,
in terms of tour quality and execution time. For the while the local search explores only the space of cyclic
ATSP from no-wait workshop scheduling, the relative city permutations. In addition, by the patching algo-
tour errors of both TBnB and local search are less than rithm, TBnB considers only those subproblems whose
0:03% in all cases. Meanwhile, TBnB takes much less AP costs are less than the initial upper bound, which
CPU time than the local search. The cost matrices decreases during the search. Thus the subproblems ex-
from no-wait scheduling problems are also subject to plored by TBnB have small AP costs, and the nal tour
the triangle inequality [9]. This suggests that TBnB cost can be signi cantly smaller than the initial upper
is a better approximation algorithm when the triangle bound.
inequality is satis ed. To reiterate, the tour quality can be continuously im-
We also examined TBnB on two particular problems proved by TBnB if it keeps backtracking when more
from industry [18], a 23-city and a 43-city ATSP. These computation time is available.
two problems have AP costs 85 and 43, but optimal The average complexity of TBnB is determined by the
ATSP costs 259 and 5620, respectively. The BnB sub- average depth of the leaf nodes in the search tree. In
tour elimination algorithm is not ecient on these prob- our experiments, the average depth of leaf nodes was
lems. Although the fast algorithm in [19] nds an ap- less than ln(n) on average for all cost matrices used.
proximate tour of cost 5625 to the 43-city ATSP al- The AP of the initial problem can be solved in O(n3 )
most immediately, it cannot optimally solve this prob- time, and the AP of an internal tree node can be incre-
lem [18]. For the 43-city problem, our implementation mentally solved in O(n2) time [17]. Since a tree node
of BnB subtour elimination with Carpaneto and Toth's has at most n children, the average3complexity of TBnB
decomposition rules and the depth- rst strategy does appears to be approximately O(n ln(n)) from our ex-
not nd an optimal tour after generating 30 million sub- periments.
problems, and a best- rst strategy runs out of memory
on a 32 Mbyte machine. TBnB nds an optimal tour of
cost 259 on the 23-city problem, while the local search
stops at a non-optimal tour of cost 262. For the 43-city
5 Conclusions
ATSP, TBnB nds a tour of cost 5623 in 1 second of Branch-and-bound (BnB) is not only a general
CPU time by expanding 54 subproblems and generat- problem-solving technique for optimally solving most
ing 173 subproblems. This result is slightly better than NP-hard combinatorial problems, but can also be
the tour of cost 5627 from the local search algorithm, adapted to an ecient approximation algorithm. We
also with 1 second of CPU time. presented such an algorithm for the asymmetric travel-
ing salesman problem (ATSP). Our experiments indi-
error relative to AP (%) CPU time (s)
9.0
patch matrix with 140 TBnB
matrix with
31 31
7.0 c(i,j) from {0,1,2,...,2 -1} c(i,j) from {0,1,2,...,2 -1}
100
5.0 full local search
full local search 60
3.0

1.0 TBnB 20 patch


200 400 600 800 1000 200 400 600 800 1000
# of cities CPU time (sec.) # of cities
error relative to AP (%)
5.0 matrix with 250 matrix with TBnB
4.0 c(i,j) from {0,1,...,i*j} c(i,j) from {0,1,...,i*j}
200
patch full local search
3.0 150
2.0 100
full local search patch
1.0 50
TBnB
0 0
200 400 600 800 1000 200 400 600 800 1000
# of cities # of cities
error relative to AP (%) CPU time (sec.)
patch 120
1.2 matrix with triangle inequality
matrix with triangle inequality
80 full local search
0.8
TBnB
0.4 40
full local search patch
TBnB
0 0
200 400 600 800 1000 200 400 600 800 1000
error relative to AP (%) # of cities CPU time (sec.) # of cities
patch 800 full local search
0.04
nowait scheduling
0.03 600 nowait scheduling
full local search
0.02 400
TBnB
0.01 200
patch
TBnB
0 0
200 400 600 800 1000 200 400 600 800 1000
# of cities # of cities

Figure 3: Comparison of patching, local search, and truncated BnB.


cate that the new algorithm is superior to a local search [11] Karp, R.M., and J.M. Steele, \Probabilistic anal-
algorithm, which is the best existing approximation al- ysis of heuristic," in The Traveling Salesman Prob-
gorithm for the ATSP. lem, E.L. Lawler, et al. (eds.) John Wiley & Sons,
We are extending the idea of developing new approx- 1985, pp.181-205.
imation algorithms using BnB, to other NP-complete [12] Kohler, W.H., and K. Steiglitz, \Enumerative and
combinatorial problems, including the symmetric TSP iterative computational approaches," in Computer
and scheduling problems. and Job-Shop Scheduling Theory, E.G. Co man, Jr.
(ed.) John Wiley & Sons, 1976.
Acknowledgements [13] Kumar, V., \Search, branch-and-bound," in En-
cyclopedia of Arti cial Intelligence, 2nd Ed, S.C.
I gratefully acknowledge helpful discussions with Shapiro (ed.) Wiley-Interscience, 1992, pp.1468-72.
Richard Karp, David Johnson, Donald Miller, Joseph [14] Lawler, E.L., J.K. Lenstra, A.H.G Rinnooy Kan,
Pekny and Bruno Repetto, comments from David John- and D.B. Shmoys, The Traveling Salesman Prob-
son and Donald Miller on an early draft. Special thanks lem, (eds.) John Wiley & Sons, 1985.
are due to my advisor, Rich Korf, for support, discus- [15] Lawler, E.L., and D.E. Wood, \Branch-and-bound
sions, comments and many draft readings. methods: A survey," Operations Research, 14
(1966) 699-719.
References [16] S. Lin, and B.W. Kernighan, \An e ective heuris-
tic algorithm for the traveling salesman problem,"
[1] Ashour, S., \An experimental investigation and Operations Research, 21 (1973) pp.498-516.
comparative evaluation of ow-shop scheduling [17] Martello, S., and P. Toth, \Linear assignment
techniques," Operations Research, 18 (1970) 541-9. problems," Annals of Discrete Mathematics, 31
[2] Balas, E., and P. Toth, \Branch and bound (1987) 259-82.
methods," The Traveling Salesman Problems, E.L.
Lawler, et al. (eds.) John Wiley and Sons, 1985, [18] Miller, D.L., personal communications, 1992.
pp.361-401. [19] Miller, D.L., and J.F. Pekny, \Exact solution
[3] Carpaneto, G., and P. Toth, \Some new branch- of large asymmetric traveling salesman problems,"
ing and bounding criteria for the asymmetric trav- Science, 251 (1991) 754-61.
eling salesman problem," Management Science, 26 [20] Papadimitriou, C.H., and P.C. Kanellakis, \Flow-
(1980) 736-43. shop scheduling with limited temporary storage," J.
[4] Dechter, R., and J. Pearl, \Generalized best- rst ACM, 27 (1980) 533-49.
search strategies and the optimality of A ," J. of [21] Papadimitriou, C.H., and K. Steiglitz, \Some ex-
ACM, 32 (1985) 505-36. amples of dicult traveling salesman problems,"
[5] Frieze, A., G. Galbiati, and F. Maoli, \On the Operations Research, 26 (1978) 434-443.
worst-case performance of some algorithms for the [22] Reddi, S.S., and C.V. Ramamoorthy, \On the ow-
asymmetric traveling salesman problem," Network, shop sequencing problem with no wait in process,"
12 (1982) 23-39. Operational Research Quarterly, 23 (1972) 323-31.
[6] Garey, M.R., and D.S. Johnson, Computers and [23] Zhang, W., and R. Korf, \An average-case analysis
Intractability: A Guide to the Theory of NP- of branch-and-bound with applications: Summary
Completeness, Freeman, 1979 of results," Proc. 10-th National Conf. on Arti cial
[7] Ibaraki, T., S. Muro, T. Murakami, and T. Intelligence, AAAI-92, San Jose, CA, July 12-17,
Hasegawa \Using branch-and-bound algorithms to 1992, pp.545-50.
obtain suboptimal solutions," Zeitchrift fur Opera- [24] Zhang, W., and R. Korf, \Performance of linear-
tions Research, 27 (1983) 177-202. space branch-and-bound algorithms," submitted to
[8] Johnson, D.S., \Local optimization and the trav- Arti cial Intelligence, 1992.
eling salesman problem," Proc 17th Intern. Collo- [25] Zhang, W., and R. Korf, \On the asymmetric trav-
quium on Automata, Languages and Programming, eling salesman problem under subtour elimination
England, July 16-20, 1990. and local search," Manuscript in preparation.
[9] Kanellakis, P.C., and C.H. Papadimitriou, \Local
search for the asymmetric traveling salesman prob-
lem," Operations Research, 28 (1980) 1086-99.
[10] Karp, R.M., \A patching algorithm for the non-
symmetric Traveling-Salesman Problem," SIAM J.
Comput. 8 (1979) 561-73.

Das könnte Ihnen auch gefallen