This action might not be possible to undo. Are you sure you want to continue?
Miller Reviewed work(s): Source: Operations Research, Vol. 22, No. 2 (Mar. - Apr., 1974), pp. 340-349 Published by: INFORMS Stable URL: http://www.jstor.org/stable/169591 . Accessed: 14/06/2012 10:40
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact email@example.com.
INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Operations Research.
In contrast. A single vehicle may not be able to supply every demand because of load and distance constraints.A Heuristic Algorithm for the Vehicle-Dispatch Problem Billy E.as well as large-scale vehicle-dispatchproblems with load and distance constraints for each vehicle. The algorithmhas the feature that the amount of computationrequired increases linearly with the number of locations if the average number of locations for each route remains relatively constant.for solving medium. A number of customers located at different locations desire a certain quantity of goods that can be delivered from a single depot. Miller Bowling Green State University. Exact algorithms are ones that. 1971) This paper introduces and illustrates an efficient algorithm. will yield an exact solution in a finite number of ANDEILON1I2 developed an exact algorithm based on a branchsteps. called the sweep algorithm. HERE ARE many problems that fall in the general category of vehicle-dispatch or delivery problems. the distance constraint could be replaced by a time constraint without changing the problem. Likewise. For example. subject to the load and distance constraints on each vehicle. in the absence of round-off or other errors. however. Ohio (ReceivedApril 11.5. if the average number of locations per route is 7. Bowling Green. The locations that are used to make up each route are determined accordingto the polar-coordinate angle for each location. however. Rolla. This is generally true of all exact algorithms that solve the vehicle-dispatch problem. Of course. The problem then is to determine the number of routes and the paths in each route that will minimize the total distance traveled by all vehicles in supplying all demands. Basically there are two types of algorithms that can be used to solve the vehicledispatch problem: exact and heuristic. the sweep algorithm is not as computationallyefficient as Gaskell's and is slightly less so than Christofides and Eilon's. the algorithmtakes approximately75 seconds to solve a 75-location problem on an IBM 360/67 and approximately115 seconds to solve a 100-location problem. the time to solve a problem with a fixed number of locations increases quadraticallywith the average number of locations per route. Missouri and Leland R. this paper specifically considers the following type of vehicle-dispatch problem. They are all restricted by comT 340 . An iterative procedure is then used to improve the total distance traveled over all routes. the problem could be thought of in terms of picking up customers or merchandise at each location. The sweep algorithm generally produces results that are significantly better than those producedby GASKELL'Ssavings approachand are generally slightly better than CHRISTOFIDES AND EILON'S results. Gillett University of Missouri. CHRISTOFIDES and-bound approach that was limited to problems involving relatively few locations because of the computer time involved.
C: capacity of each vehicle. A(I. . assume An(I)<An(I+ ) for all I. on the other hand. unique ordering. forall I. problems with many more than 100 locations are well within the computational limits of the sweep algorithm. Consequently. for all I. whereas the sweep algorithm divides the locations into a number of routes and then operates on the individual routes until an optimum or near-optimum solution is obtained. ))/(X(I)-X(1))]. If there exist an I and J then I<J if A (1. When the problem is broken into a number of smaller subproblems. Heuristic algorithms. are quite often faster and capable of obtaining optimum or near-optimum solutions to much larger problems in a reasonable amount of computer time. J): distance between locations I and J. * . 3. The constraints on the problem are Q(I) <C. forallI#J. J)>0. * *XN) Q (I): demand at location I. Y (I)): rectangular coordinates of Ith location. Heuristic methods developed to date 2'67'141 have been limited to problems with no more than 100 locations. I)< D Assume the locations are renumbered according to the size of their polar-coordinate angle. A (I. 3. This determines a such that An(I)=An(J). 2. I)<A (1. . A (I. J).The Vehicle-Dispatch Problem 341 puter time and/or storage. D: maximum distance each vehicle can travel. if Y(I)-Y(1)<0. I)=0. The following notation and definitions will aid in its development. SWEEP ALGORITHM THE ALGORITHM DISCUSSED in this section will yield an optimum or near-optimum solution to the single-depot vehicle-dispatch problem. and 0<An(I)?7r (I = 2. N) defined as An(I) =arctan[(Y(I)-Y(1 where -7r<An(I)<O Y(I)-Y(1)?0. N: number of locations including the depot (the depot is always location 1). This concept will be treated in detail in the next section. This point is illustrated by methods B and C in Christofides and Eilon[2I and by the sweep algorithm presented in this paper. That is. Each method attempts to solve the vehicle-dispatch problem as one big problem. N) (X (I). the computation time involved increases somewhat in a linear rather than in an exponential manner as more locations are added to a given problem. (I = 1. (I = 2. if N) R (I): radius from location 1 to location I. 3 . 1) +A (1. An (I): polar coordinate angle of Ith location (I = 2. A (I.
namely. 2. The procedure is then modified to consider replacing one location in route K with one or more locations in route K+ 1 for K = i. The process of adding one or more locations to route K and deleting another location continues until no improvement is found (K= 1. The sweep algorithmconsists of two parts. a forward sweep and a backward sweep. route 2 contains locations L-1. This process is continued until J is the last location in route K? 1 that can be added to route K. The X and Y axes are then rotated counterclockwise so the first location becomes the last. The second location considered for inclusion in route K is the location in route K+ 1 that is nearest to location p. to be picked up by a later formed route. algorithm. Recall that the locations were renumbered according to the size of their polar-coordinate angle and the depot is location 1. L. say location p. *. *. the locations are partitioned into routes beginning with the location that has the smallest angle. Initially route 1 contains locations N. The remaining routes are formed in the same manner. where J is the last location that can be added without exceeding the vehicle capacity or distance constraint. location 2. The second route contains locations J +1. J. then the next location in route K+ 1 is also checked to see if it can be included in route K. The total distance traveled then is just the sum of the distances for each route. Gillettand LelandR. where m is the number of routes formed. The process of rotating the X and Y axes is continued until all possibilities have been exhausted. The time to solve a . The sweep algorithm was used to solve a large number of vehicle-dispatch problems where the number of locations ranged from 10 to 250. the second becomes the first. Miller The vehicle-dispatch problem is to minimize the total distance traveled in supplying all demands while satisfying all constraints. * . Of course. This provides a location that is close to the depot and also close to the next route. N-i. however. N-2. A step procedure for the forward-sweep algorithm is given in Appendix A. it is a very fast scheme for selecting the locations and it produces good results. Choosing locations in this manner may not give the best solution. 3. Again a minimum total distance is calculated. *. rn-1. and so forth. L . 2. Of course. the smallest of these two minimums is the best approximate solution. In the forward-sweepalgorithm. The first location. a replacement is made only if the total distance is decreased. The first route consists of locations 2. m7). Each time a minimum total distance is calculated. A function that works very well is RI(I)+An(I) -A VR. In most cases the two procedures produce different routes and consequently in some cases different minimum total distances. The replaced location is left unassigned. is exactly like the forA second algorithm. The smallest of these minimums provides a good heuristic solution for the vehicle-dispatch problem.342 BillyE. L. J + 2. ** M. and so forth. **. that is considered for inclusion in route K is the location in route I+1 that is nearest to the last location that was added to route K. where L is the last location that can be added to the second route without exceeding the vehicle capacity or distance constraint.2. Location J+2 is then checked to see if it can be included in route K. The above procedure of partitioning routes and interchanging locations between routes is then repeated. called the backward-sweep ward-sweep algorithm except it forms the routes in reverse order. If one or more locations are added to route K by this scheme. where A VR is the average of the radii for all locations. The location to be deleted from route K is obtained by minimizing a function of the radius R(I) and the angle An(I) of each location in route K.
However. the only significant amount of time is the sum of the times associated with each route. for the 75-location problem. The time TKassociated with the Kth route is the time required to check to see if a location should be deleted from route K and/or if other locations from route K+1 should be added to route K. where x is the average number of locations per route and y is the total computer time in seconds. If the route contains relatively few locations. This explains why one 40-location problem may take 10 seconds to solve while another 40-location problem may take 50 seconds to solve. locations were generated that met these three constraints. When the sweep algorithm is used to solve a vehicle-dispatch problem.TheVehicle-Dispatch Problem 343 given problem is strictly a function of the number of routes and the number of locations per route. The demand constraint was varied so that the average number of locations per route was 5. Y) plane. In fact. K= 1. Similar results hold for the 100-location problem. Note that. while the second may have only two routes with 20 locations each. if h additional locations are added to the original problem it only requires approximately B additional units of time to solve the new problem. Thus. (2) the demand per location is uniformly distributed. . the total computer time increases quadratically as the average number of locations per route increases from 5 to 10 to 15. the computer time per route remains relatively constant and the total computer time increases linearly as the number of locations increases from 75 to 100. the number of locations per route has the greatest effect on the total computer time. if the average number of locations per route remains constant. TKincreases rapidly as the number of locations per route increases. If the number of locations per route is relatively constant. The multiple correlation coefficient R2 was equal to 0. The computer time needed to solve the traveling-salesman problem increases quadratically with the number of locations. where m is the number of routes. then TK= B = . and (3) the locations are uniformly distributed in the (X. if any number of locations. say h. In fact. and 15 for each case. Table I gives the results. say p. mi. A leastsquares equation to predict the total computer time needed to solve the 75-location problem as a function of the average number of locations per route was calculated: y= 227. OK=1 TK. even with the best algorithm. m. One case involved 75 locations and another 100 locations. *. the approximate time to solve the new problem is the time to solve the original problem plus the time to solve a separate vehicle-dispatch problem with p locations. and the total time for the sweep algorithm is constant for all K= I. To illustrate. the first problem may have approximately five locations per route. Also note that. The initial allocation of locations to routes takes a very insignificant amount of time. the time to solve the vehicle-dispatch problem increases linearly with the number of locations if the number of locations for each route is approximately the same.1. whereas the computer time increases quadratically with the average number of locations per route if the total number of locations remains relatively constant. The computer time for the sweep algorithm increases quadratically with the average number of locations per route if (1) the total number of locations remains relatively constant. 2.33x2. 2.2-68. On the contrary.96 for this model. TK is fairly small. This process involves solving the traveling-salesman problem each time a check is made. 10.57x+6. is added to a given problem.
cities.711 10.461 12. a traveling-salesman problem must be solved to determine the minimum path to service each of the locations in the route. LIN AND KERNIGHAN 'I1 have developed a heuristic algorithm that will produce an optimum or near-optimum solution to 100-location problems in reasonable times. Gillettand LelandR. Consequently. their revised algorithm is faster in most cases and is always at least as accurate as the original Lin[12I procedure. cities. The HELD AND KARPf'01 procedure is an exact method that must rely on a branch-and-bound procedure in many cases.2 Held and Karp Distance Time (sec) 12.0 107. whereas exact algorithms may be extremely slow but always get the optimum solution if time permits.95 23.344 BillyE. Held and Karp['l Karg and Thompson["] Dantzig and Fulkerson Held and Karp['1 Karg and Thompson["l] 1.Miller TABLE I COMPARISONOF COMPUTERTIMES FOR 75 AND 100 LOCATIONS Average number of locations/route Number of locations Total computer time (Sec) Computer time per route (sec) 5 5 10 10 15 15 75 100 75 100 75 100 41 59 174 233 622 856 2.0 780.7 36. cities. cities. In addition. but the IBM 360/91 is several times faster.Y'2] Table II shows the comparisons.955 54.52 TRAVELING-SALESMAN ALGORITHMS set of locations is considered for a given route. Highly successful heuristic algorithms as well as exact ones are available to solve the traveling-salesman problem. but sometimes sacrifice accuracy. the computer times reported are excessive even when compared to heuristic results reported by LIN.30 124.985 GE 635 Time (sec) 5.20 23. . Specifically. Lin's112] algorithm failed to give the optimal solution in only one case.3 63.861 699 11. However.73 2.0 84.0 Computer(a) IBM 360/91 (a) Exact comparison between the IBM 360/91 and the GE 635 cannot be made.40 128.461 12. The heuristic algorithms are generally considerably faster.0 699 11.24 10. EACH TIME A TABLE II COMPARISON OF THE LIN AND HELD-KARP PROCEDURES Lin procedures Problem Distance 25 33 42 48 57 cities. Lin and Kernighan report that their revised algorithm gets the optimum solution for each of the cases in Table II.
The same is true with respect to problems 10 and 11. Seven of the problems proposed by Gaskell and Christofides and Eilon[21 along with five variations of these problems are presented in Table III.4 7.0 Christofides & Eilon New problem Christofides & Eilon[21 New problem New problem New problem Christofides & Eilon New problem Excludes depot.The Vehicle-Dispatch Problem 345 algorithm appears to be the best algorithm to use to solve The Lin-Kernighan[l31 the traveling-salesman phase of the sweep algorithm because of its near accuracy. 75. load 6000 4500 4500 8000 160 100 140 180 220 112 200 120 Max. and was able to obtain the same solution as Christofides and Eilon in problems 3 and 4. COMPUTATIONAL RESULTS 11 AND CHRISTOFIDES EILON solved ten vehicle-dispatch problems from the literaproture and compared their results with those obtained by one of GASKELL'St61 cedures for the same problems. This was done to illustrate that the time to solve a given problem is highly dependent on the average number of locations per route and much less on the total number of locations. Christofides and Eilon[21report that 10 runs of 36 seconds of IBM\ 7090 time each were needed to obtain their results for problem 1. which was just 0.7 percent better than that ob- . and its availability. Using best solution. loc.0 5.0 7. Problems 6. Table IV shows a comparison of results from the vehicle-dispatch algorithms mentioned in this paper.25 8.37 Ave. distance 210 240 240 240 Unlimited SC it Ca9.5 10. 1 2 3 4 5 6 7 8 9 10 11 12 (a) (b) Source Gaskell Gaskell Gaskell[6l] Gaskell No. except that the load constraint is changed. and 9 are the same as problem 7.25 4. no. no.71 7. The sweep-algorithm solutions for problems 1 and 2 were very close to the best available solutions. TABLE III LIST OF PROBLEMS Prob. (a) 21 22 29 32 50 75 75 75 75 100 100 250 Max. it produced significant decreases in the total distances traveled in most cases and was able to reduce the number of routes in three of the problems.5 10. loc. the sweep algorithm was able to produce better solutions for the larger problems (50. with much less computer time. Even though their method took approximately three to four times longer for most of the problems.14 12. In each case where a comparison with other algorithms was possible.0 10. 8. and 100 locations). per route(b) 5.
0 0. multiple approach.5 863 8 865 765 723 1176 10.21 0.?4 N no .0 864 5794 Excludes depot.1 0. 598 955 963 839 585 900 4 5 5 5 6 10 min(b) Sol.8 2. 591 956 888 817 546 1128 884 754 715 1170 862 5911 Rts.2 0. (e) Attribute (c) above could restrict the algorithm to problems as small as 60 locations if there are approximately 30 locations per route. Method C-forward-sweep algorithm.1 0.17 0.23 2. BillyE. Rts. Method D-backward-sweep algorithm. and 10 runs of 48 seconds each for problem 3.0 0. (c) The computer time increases quadratically with the average number of locations per route if the total number of locations remains relatively constant.346 R.23 3.62 2. (C) IBM 360/67-total time for Methods C and D. 585 949 875 810 556 876 4 5 4 4 5 10 min(b) Sol. (b) The computer time increases linearly with the total number of locations if the number of locations for each route remains relatively constant. 4 5 4 4 5 15 10 8 7 14 8 25 Sol. Based on (b) above.7 1 2 3 4 5 6 7 8 9 10 11 12 (a) (b) 21 22 29 32 50 75 75 75 75 100 100 250 0.83 6._ _ Best Sweep Sol. Rts. 591 956 875 810 546 1127 865 754 715 1170 862 5794 Rts.2 0. The 250-location problem with an average of 10 locations per route was solved in 9.0 0. Note: Method A-Gaskell's saving. _ _ _ _ _ _ Sol. it is practical timewise even up to 250 locations with 10 locations per route and 1000 locations with five locations per route.3 6.51 0.0 4. The literature is void of algorithms that can solve problems this large. 602 970 875 810 574 1127 4 5 4 4 5 15 10 8 7 14 8 25 Sol.0 9.6 1. min(C) 4 5 4 4 5 15 10 8 7 14 8 25 0. Gillettand Leland Miller tained using the sweep algorithm. (d) It is extremely useful for very large problems that have only a few locations per route on the average. CONCLUSIONS THE SWEEP ALGORITHM has the following attributes: (a) It is a new approach to the solution of the single-depot vehicle-dispatch problem.7 minutes of IBM 360/67 time. where the result was the same as the result from the sweep algorithm. (a Method A _ _ MethodB _ _ _ _ _ - Method C Method D _ _ _ _ _ _ _ .0 887 8 2.5 8. (f) For large problems from the literature (50 locations or more) it has proved to be better than the best known algorithms and possibly requires just slightly . IBM 7090.68 1. 1 oC ) . TABLE IV COMPARISON OF VEHICLE-DISPATCH ALGORITHMS Pro~b lo~. Rts. Method B-Christofides and Eilon's 3-optimal approach.68 1.
Otherwise. N. or K(J+4). (g) For small problems it is comparable with the best known algorithms with respect to both time and accuracy. (The IBM 360/67 is roughly three to four times faster than the IBM 7090. Step 3. Step 12. go to step 10. Otherwise. Continue If the distance capacity is exceeded. Step 6. constraint. constraint this procedure until the distance Step 8. *. If D1+D3<D2+D4. K(J+1). K(J+1). Begin the first route with J =2 and SUMll Q(K(2)). Step 14. Step 2. then go to step 10. Calculate the minimum distance D1 for the route by means of a traveling-salesCheck the distance constraint again. then go to step 13. and ending at K(J+5). Step 4. 3. Step 10. Otherwise go to step 10. but may vary considerably depending on the program and programmer). from the route. K(3) denote the location with the second smallest angle. step 4. Make SUMI =SUM1J-Q(K(J-1)) and J=J-1. through the minimum distance D3 for starting at 1. Determine K(I) for I -2. Let An(I) represent the angle and R(I) the radius for location I. and so forth. eliminate K(J-1) Check the distance is satisfied. then go to step 16.. If J AN. Evaluate the minimum distance Db for the route with K(JJX) substituted for K(KII). or K(J+4).. K(J+4) and ending at K(J+5). 0). Determine JJX so that K(JJX) is the nearest location to K(J-1) Find JII so that K(JII) route. with K(JJX) and K(JII) ex- . Step 5. K(J+4). Let KII denote this I. If D2 ?D and the load constraint is satisfied.. If SUM+Q(K(J)) >C. traveling through locaDetermine the distance D4 for *.. then go to step 11. go to step 12. Instead of relabeling the locations. Step 9. ***. traveling Evaluate Go to tions K(J). J+2 is not checked. N such that An(K(I)) is less than or equal to An(K(I + 1)). Go to step 4. except eliminate K(JJX) and inject K(KII). I =2. If D5 < D and the load constraint is satisfied. . is the nearest Likewise determine I so that R(K(I)) +An(K(I)) an(I not in a location to K(JJX) and not in a route. Place K(JJX) in the route and remove location K(KII). . then go to step 4. then man algorithm. Evaluate the polar coordinates for each location with the depot at (0. Step 1. SUM =SUM +Q(K(J)). *. K(J+1). Determine the minimum distance D6 for starting at 1. then location The notation in the body of the paper is assumed. *. Increment the angle by making J =J+1. K(JJX) is not K(J). K(J+1). If the same locations. Determine the minimum distance D2 for the route with K(JJX) added to the route and K(KII) deleted from it. A IVR is a minmum for all locations in the route. Step 7. If K(JJX) and K(JII) are not K(J). we let K(2) denote the location with the smallest angle. then go to step 10. traveling through loca- tions K(J). Augment the route with location K(J) by making If J =N.TheVehicle-Dispatch Problem 347 more computer time. Record the route and start a new route by setting SUM =Q(K(J)). then go to step 14. then go to step 7. and K(JII) Step 13. 3. APPENDIX A A STEP-BY-STEP PRESENTATION OF THE FORWARD-SWEEP ALGORITHM IF LOCATION J is in a given route and location J +1 cannot be added to it. Step 11.
21 233. Go to step 10. along with the corresponding sweep-algorithm solutions can be obtained from the authors. Go to step 2.e. Step 18. Allow 10 miles extra for each location visited. Step 19. If satisfied. 1 2 3 4 Route 0-18-10-11-12-9-17-7-13-16-15-0 0-26-28-27-25-24-29-0 0-20-3-6-1-4-5-2-22-0 0-23-8-14-21-19-0 Total Miles 227. Maximum distance: 240. Step 17. then that set of routes is complete. Depot coordinates: (162. Miller TABLE V 3 OF TABLE X 126 125 116 126 125 119 115 153 175 180 Y 347 346 355 335 355 357 341 351 363 360 PROBLEM III No. then go to step 17. . Evaluate the minimum distance for the route and check the distance constraint. Stop.354). start with K(3) for the second set of routes]. go to step 15. 11 12 13 14 15 16 17 18 19 20 Demand 950 125 150 150 550 150 100 150 400 300 No.00 Note: The depot is location 0. APPENDIX B THE DATA FOR problem 3 in Table III and the corresponding routes using the sweep algorithm are given in Tables V and VI.00 2800. Place K(JJY) and K(JII) in the route and eliminate K(KII) from the route. 1 2 3 4 5 6 7 8 9 10 X 218 218 201 214 224 210 104 126 119 129 Y 382 358 370 371 370 382 354 338 340 349 Demand 300 3100 125 100 200 150 150 450 300 100 No.59 177. Otherwise. TABLE VI THE FORWARD-SWEEP ALGORITHM SOLUTION OF PROBLEM 3 OF TABLE III No. Step 15. Check to see if another set of routes is needed.95 236. Otherwise go to step 18. Delete one from the route (J =J-1). Increment the angle by one location [i. Step 16.24 874.99 Load 2725. Gillettand LelandR. If no more are needed. then go to step 10.348 BillyE.00 4375.00 2850. If not satisfied. eluded and K(KII) included. If D1+D3<D5 +Dc. Go to step 4. then go to step 19.. The data for the other problems in Table III. 21 22 23 24 25 26 27 28 29 X 159 188 152 215 212 188 207 184 207 Y 331 357 349 389 394 393 406 410 392 Derand 1500 100 300 500 800 300 100 150 1000 Maximum load: 4500.
. M. Res. 11. 7. 106 (1967)." Opns." Mathematical Programming 1. M.The Vehicle-Dispatch Problem REFERENCES 349 1."A Heuristic Approach to Solving Traveling Salesman Problems.. 20. Management Science Research Report No. N. EILON. 10. Quart. Res. "A Dynamic Programming Approach to Sequencing Problems. R. 16. SHEN LIN. 1138-1162 (1970). Res." J. R. Opns. 3. of Tech. "The Traveling Salesman Problem: A Survey. 2245-2269 (1965). 5. Problem. .AND J. 196-210 (1962). 19. Indust. CLARKAND J. . FULKERSON." J. 225-247 (1964). W. L. 6-25 (1971). D. BELLMOREAND G. 14." 9. Res. AND B. Res. Res. AND ." Opns. ." Management Sci. JOHNSON." Bell Syst. 11. G. 393-410 (1954). "The Traveling Salesman Problem and Minimum Spanning Trees: Part II. "The Delivery Problem. Scheduling of Vehicles from a Central Depot to a Number of Delivery Points. 354-358 (1968). 8. 44. COCHRAN. A. W. Report ? 1. KARG AND G. April 1972. Quart. L. TILLMAN AND H. THOMPSON. NEMHAUSER. AND S. GASKELL. L. 80-91 (1959)." Opns. F. M. M. Bell Labs. "Solution of a Large-Scale Traveling Salesman Problem. G. CHRISTOFIDESAND S. ROBERT HAYES. KARP. J. HELD AND R. 2. 12. Eng. "Computer Solutions of the Traveling Salesman Problem." Opnal. 6. 2. 4. 12. T. Res. B. "A Heuristic Algorithm for the Traveling Salesman 13." Opns. "An Algorithm for the Vehicle Dispatching Problem. 309-318 (1969). 538-558 (1968). RAMSER." Opnal. 18. WRIGHT. SIAi1J 10. H. "The Truck Dispatching Problem. 18. KERNINGHAN. AND 10. 281-295 (1967). DANTZIG."A Heuristic Approacb for Solving the Delivery Problem. "The Traveling Salesman Problem and Minimum Spanning Trees. 568-581 (1963). Tech J." Carnegie Inst. "Bases for Vehicle Fleet Scheduling." Computer Science Tech.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue listening from where you left off, or restart the preview.