Optimization with Big Data: Network Flows

Transcription

1 Optimization with Big Data: Network Flows Sertac Karaman Assistant Professor of Aeronautics and Astronautics Laboratory for Information and Decision Systems Institute for Data, Systems, and Society Massachusetts Institute of Technology

2 Recall Linear Programming Maximize Linear (or minimize) Programming an objective function with respect to constraints General Formulation Maximize subject to: n c j x j j 1= n a x ij j b i for i = 1, 2,, m j 1= for j = 12,,, n x j 0 objective function constraints

3 Linear Programming Formulation Linear Programming LP as a Decision Making Problem: Maximize Z = c 1 x 1 + c 2 x c n x n Subject to: a 11 x 1 + a 12 x a 1n x n b 1 a 21 x 1 + a 22 x a 2n x n b 2... a m1 x 1 + a m2 x a mn x n b m and x 1 0 x 2, 0, x n 0 NEXTOR/Virginia Tech - Air Transportation Systems Lab Given by the problem instance Value maximization subject to resource constraints x j c j b i a ij Decision variable - we would like to determine (decide) this The unit value for the jth decision variable The available amount for the ith resource The unit of ith resource required for one unit of 6 the jth decision variable

4 Linear Programming Formulation Linear Programming Maximize Z = c 1 x 1 + c 2 x c n x n Subject to: a 11 x 1 + a 12 x a 1n x n b 1 a 21 x 1 + a 22 x a 2n x n b 2... a m1 x 1 + a m2 x a mn x n b m and x 1 0 x 2, 0, x n 0 Feasible Solution: A set of values for the decision variables (x1,, xn) that NEXTOR/Virginia Tech - Air Transportation Systems Lab satisfies all of the constraints. Optimal Solution: A feasible solution that gives the best value ( c ix i ) j=1 among all feasible solutions. X n 6

5 Linear Programming Formulation Linear Programming Linear Programming Maximize Z = c 1 x 1 + c 2 x c n x n Subject to: a 11 x 1 + a 12 x a 1n x n b 1 a 21 x 1 + a 22 x a 2n x n b 2... a m1 x 1 + a m2 x a mn x n b m and x 1 0 x 2, 0, x n 0 Maximize Z = c 1 x 1 + c 2 x c n x n Minimize Subject to: a 11 x 1 + a 12 x a 1n x n b 1 a 21 x 1 + a 22 x a 2n x n b 2... a m1 x 1 + a m2 x a mn x n b m and x 1 0 x 2, 0, x n 0 NEXTOR/Virginia Tech - Air Transportation Systems Lab NEXTOR/Virginia Tech - Air Transportation Systems L The problem can be cast either as a maximization or a minimization problem. (Just multiply all the value constants ci by -1. ) 6

6 Linear Programming Formulation Cost minimization subject to resource constraints x j Decision variable Subject to: Linear Programming Minimize Maximize Z = c 1 x 1 + c 2 x c n x n Given by the problem instance c j b i a ij The unit cost for the jth decision variable The available amount for the ith resource The unit of ith resource required for one unit of the jth decision variable a 11 x 1 + a 12 x a 1n x n b 1 a 21 x 1 + a 22 x a 2n x n b 2... a m1 x 1 + a m2 x a mn x n b m and x 1 0 x 2, 0, x n 0 NEXTOR/Virginia Tech - Air Transportation Systems L The problem can be cast either as a maximization or a minimization problem. (Just multiply all the value constants ci by -1. ) In the new problem, ci can be interpreted as cost.

7 LP Example: Maximizing Capacity with Constrained Crew and Vehicle Supply Freight needs to be carried with trucks. Two types of trucks are available in limited numbers, each with different capacity and crew requirements: Capacity Crew required Number available Anadol BMC We have a limited amount of crew: Exactly 180 number of personnel available. All personnel can operate either truck. How many trucks of each type would you utilize?

8 Some Important Linear Programming Problems Transportation problem Assignment problem Maximum flow problem on a network Minimum-cost flow problem on a network

9 The Transportation Problem Need to ship goods peas from canneries to warehouses. warehouses for the P & T Co. problem THE TRANSPORTATION AND ASSIGNMENT PROBLEMS CANNERY 1 Bellingham CANNERY 2 Eugene WAREHOUSE 2 Salt Lake City WAREHOUSE 3 Rapid City CANNERY 3 Albert Lea How can you formulate a linear program for this shipment problem? WAREHOUSE 1 Sacramento WAREHOUSE 4 Albuquerque TABLE 8.2 Shipping data for P & T Co. Shipping Cost ($) per Truckload Warehouse FIGURE 8.1 Location of canneries and warehouses for the P & T Co. problem Output TABLE 8.2 Shipping data for P & T Co Cannery Allocation Shipping Cost ($) per Truckload Warehouse Output Cannery Allocation

10 The Transportation Problem Consider the network formulation: 8.1 THE TRANSPORTATION PROBLEM 353 W1 [ 80] [75] C E 8.2 rk representation of T Co. problem. [125] [100] C2 C W2 W3 W4 [ 65] [ 70] [ 85] subject to the constraints x 11 x 12 x 13 x 14 x 21 x 21 x 21 x 21 x 21 x 21 x 21 x 21 75

11 warehouse has been allocated a certain amount from the total supply of peas. This information 8.1 (in THE units TRANSPORTATION of truckloads), along PROBLEM with the shipping cost per truckload for each cannery-warehouse combination, is given in Table 8.2. Thus, there are a total of 300 truck- 353 loads to be shipped. The problem now is to determine which plan for assigning these shipments to the various cannery-warehouse combinations would minimize the total ship- W1 [ 80] The Transportation Problem ping cost. Define the decision variables: x ij is the amount shipped from cannery i to warehouse j, where i = 1, 2, [125] 3 and C2 j = 1, 2, 3, 4. Formulate the (the objective x ij ) so as to function: FIGURE 8.2 Network representation of the P & T Co. problem. Formulate the constraints: By ignoring the geographical layout of the canneries and warehouses, we can provide a network [75] representation C1 of this problem in a simple way by lining up all the canneries in one column on the left and all the warehouses in one column on the right. This representation is shown in Fig The arrows show the possible routes for the truckloads, where the number next to each arrow is the shipping cost per truckload for that route. A square W2 [ 65] bracket next to each location gives the number of truckloads to be shipped out of that location (so that the allocation into each warehouse is given as a negative number). The problem depicted in Fig. 8.2 is actually a linear programming problem of the transportation problem type. To formulate the model, let Z denote total shipping cost, and let x ij (i 1, 2, 3; j 1, 2, 3, 4) be the number of truckloads to be shipped from cannery W3 [ 70] i to warehouse j. Thus, the objective is to choose the values of these 12 decision variables [100] C3 Minimize Z 464x x x x x x x x x x x x 34, W4 [ 85] subject to the constraints and x 11 x 12 x 13 x 14 x 21 x 21 x 21 x 21 x 21 x 21 x 21 x x 21 x 21 x 21 x 21 x 21 x 22 x 23 x 24 x 21 x 21 x 21 x x 21 x 21 x 21 x 21 x 21 x 21 x 21 x 21 x 31 x 32 x 33 x x 11 x 21 x 21 x 21 x 21 x 21 x 21 x 21 x 31 x 21 x 21 x x 11 x 12 x 21 x 21 x 21 x 22 x 21 x 21 x 21 x 32 x 21 x x 11 x 12 x 13 x 21 x 21 x 21 x 23 x 21 x 21 x 21 x 33 x x 11 x 12 x 13 x 14 x 21 x 21 x 21 x 24 x 21 x 21 x 21 x x ij 0 (i 1, 2, 3; j 1, 2, 3, 4).

12 The Transportation Problem Model The General Transportation Problem To describe the general model for the transportation problem, we need to use terms that are considerably less specific than those for the components of the prototype example. In particular, the general transportation problem is concerned (literally or figuratively) with distributing any commodity from any group of supply centers, called sources,to any group of receiving centers, called destinations, in such a way as to minimize the total distribution cost. The correspondence in terminology between the prototype example and the gen- Several similar problems are in fact source-to-destination transportation eral problem is summarized in Table 8.4. problem. As indicated by the fourth and fifth rows of the table, each source has a certain supply of units to distribute to the destinations, and each destination has a certain demand TABLE 8.4 Terminology for the transportation problem Prototype Example Truckloads of canned peas Three canneries Four warehouses Output from cannery i Allocation to warehouse j Shipping cost per truckload from cannery i to warehouse j General Problem Units of a commodity m sources n destinations Supply s i from source i Demand d j at destination j Cost c ij per unit distributed from source i to destination j

13 nes. Numerous such transportation problems best new production and distribution system. l The General Transportation Problem The general transportation problem is formulated as a linear program as follows: portation problem, we need to use terms that the components of the prototype example. In m is concerned (literally or figuratively) with of supply centers, called sources,to any group such a way as to minimize the total distribuy between the prototype example and the gens of the table, each source has a certain sups, and each destination has a certain demand tation problem m c m1 c m2 Demand d 1 d 2 Therefore, formulating a problem as a trans a parameter table in the format of Table 8.5. A provided by using the network representation o necessary to write out a formal mathematical m However, we will go ahead and show yo portation problem just to emphasize that it is in problem. Letting Z be the total distribution cost and the number of units to be distributed from sou ming formulation of this problem is 8 THE TRANSPORTATION AND ASSIGNMENT PROBLEMS TABLE 8.5 Parameter table for the transportation problem Source General Problem Units of a commodity m sources n destinations Supply s i from source i Demand d j at destination j Cost c ij per unit distributed from source i to destination j Cost per Unit Distributed Destination 1 2 n Supply 1 c 11 c 12 c 1n s 1 2 c 21 c 22 c 2n s 2 m c m1 c m2 c mn s m Demand d 1 d 2 d n Therefore, formulating a problem as a transportation problem only requires filling out a parameter table in the format of Table 8.5. Alternatively, the same information can be Minimize subject to and n j 1 m i 1 Z m i 1 n j 1 c ij x ij, x ij s i for i 1, 2,..., m, x ij d j for j 1, 2,..., n, x ij 0, for all i and j. Note that the resulting table of constraint coef Table 8.6. Any linear programming problem t transportation problem type, regardless of its numerous applications unrelated to transportatio ture, as we shall illustrate in the next example l lem described in Sec. 8.3 is an additional exam transportation problem is considered such an ming problem.

14 c 1n The General Transportation Problem The same transportation problem can be considered 8.1 THE TRANSPORTATION PROBLEM 357 as a network flow problem as follows: [s 1 ] [ d 1 ] S 1 c 11 D 1 c 12 c 21 [s 2 ] S 2 D 2 [ d 2 ] c 22 c 2n c m1 c m2 FIGURE 8.3 Network representation of the transportation problem. [s m ] [ dn] S m c mn D n

15 The General Transportation Problem 8.1 THE TRANSPORTATION PROBLEM 359 In Excel the transportation problem can be written down as follows: 8.4 dsheet formulation of T Co. problem as a rtation problem, ows 3 to 9 show the ter table and rows 12 isplay the solution ter using the Excel o obtain an optimal g plan. Both the s for the output cells specifications to set up the Solver n at the bottom.

16 Another Transportation Problem in Excel 8.3 THE ASSIGNMENT PROBLEM 387 Consider a special case, where we assign products to plants for production: TABLE 8.27 Data for the Better Products Co. problem Unit Cost ($) for Product Capacity Available Plant Production rate ucts, except that Plant 2 cannot produce product 3. However, the variable costs per unit of each product differ from plant to plant, as shown in the main body of Table Management now needs to make a decision on how to split up the production of the products among plants. Two kinds of options are available. Option 1: Permit product splitting, where the same product is produced in more than one plant. Option 2: Prohibit product splitting. This second option imposes a constraint that can only increase the cost of an optimal solution based on Table On the other hand, the key advantage of Option 2 is that it eliminates some hidden costs associated with product splitting that are not reflected in Table 8.27, including extra setup, distribution, and administration costs. Therefore, management wants both options analyzed before a final decision is made. For Option 2, man-

17 s a t to and :G26) Another Transportation Problem in Excel 8 THE Consider TRANSPORTATION a AND special ASSIGNMENT case, PROBLEMSwhere we assign products to plants for production: n of d by

18 The Assignment Problem Goal: Match assignees to tasks, so that: The number of assignees and the number of tasks are the same. (This number is denoted by n.) Each assignee is to be assigned to exactly one task. Each task is to be performed by exactly one assignee. There is a cost cij associated with assignee i (i = 1, 2,..., n) performing task j (j = 1, 2,..., n). The objective is to determine how all n assignments should be made to minimize the total cost. 8.3 THE ASSIGNMENT PROBLEM TABLE 8.25 Cost table for the Job Shop Co. assignment problem The Assignment Problem Model and Solutio The mathematical model for the assignment problem uses x ij 1 0 Task (Location) Assignee 2 15 M (Machine) (D) if assignee i performs task j, if not, for i 1, 2,..., n and j 1, 2,..., n. Thus, each x ij 0 or 1). As discussed at length in the chapter on integer variables are important in OR for representing yes/no d decision is: Should assignee i perform task j? By letting Z denote the total cost, the assignment p Minimize Z n i 1 n j 1 c ij x ij, subject to n

19 (Location) Task (Location) Assignee M The Assignment Problem (Machine) Assignee M (Machine) (D) (D) However, large assign cialized solution procedure eral simplex method for bi Because the assignmen venient and relatively fast w transportation simplex met the cost table to a paramet Table 8.26a. The assignment problem can be formulated as an integer program as follows: The Assignment The Problem Model and Solution Procedures Define the decision variables: The mathematical model for the assignment problem uses the following decision variables: The mathematical model for the assignment problem uses the following decision variables: x ij for i 1, 2,..., n and j 1, 2,..., n. Thus, each x ij is a binary variable (it has value 0 or 1). As discussed at length in the chapter on integer programming (Chap. 12), binary variables are important in OR for representing yes/no decisions. In this case, the yes/no decision is: Should assignee i perform task j? Source By letting Z denote the total cost, the assignment problem model is Write down the objective function and the constraints: Minimize subject to and x ij 0 n j 1 n i 1 Z n n i 1 j 1 n j 1 c ij x ij, x ij 1 for i 1, 2,..., n, x ij 1 for j 1, 2,..., n, x ij 0, for all i and j (x ij binary, for all i and j). if assignee i performs task j, if not, for i 1, 2,..., n and j 1, 2,..., n. Thus, each x ij is a binary variable (it has value 0 or 1). As discussed at length in the chapter on integer programming (Chap. 12), binary variables are important in OR for representing yes/no decisions. In this case, the yes/no decision is: Should assignee i perform task j? By letting Z denote the total cost, the assignment problem model is Minimize subject to and n j 1 n i 1 Z n i 1 c ij x ij, x ij 1 for i 1, 2,..., n, x ij 1 for j 1, 2,..., n, The first set of functional constraints specifies that each assignee is to perform exactly one task, whereas for the all second i and set j requires each task to be performed by exactly one assignee. x ij 0, if assignee i performs task j, if not, TABLE 8.26 Parameter table for the assignment proble transportation problem, illustrated by the J (a) General Case Cost per Unit Distributed Destination 1 2 n Supply 1 c 11 c 12 c 1n 1 2 c 21 c 22 c 2n 1 m n c n1 c n2 c nn 1 Demand 1 1 1

20 c 1n Now focus on the integer solutions property in the subsection on the transportation problem model. Because s i and d j are integers ( 1) now, this property implies that every BF solution (including an optimal one) is an integer solution for an assignment problem. The functional constraints of the assignment problem model prevent any variable from being greater than 1, and the nonnegativity constraints prevent values less than 0. Therefore, by deleting the binary restriction to enable us to solve an assignment problem as a linear programming problem, the resulting BF solutions obtained (including the final optimal solution) automatically will satisfy the binary restriction anyway. Just as the transportation problem has a network representation (see Fig. 8.3), the assignment problem can be depicted in a very similar way, as shown in Fig The first column now lists the n assignees and the second column the n tasks. Each number in a The Assignment Problem The assignment problem can be expressed in the network flow as follows: FIGURE 8.5 Network representation of the assignment problem. c 11 [1] [ 1] A 1 c 12 T 1 c 21 [1] A 2 T 2 [ 1] c 22 c2n c n1 c n2 [1] [ 1] A n c nn T n

21 The Maximum Flow Problem 9.5 THE MAXIMUM FLOW PROBLEM All flow in a network originates from node (called the source) and terminates in another node (called the destination). All remaining nodes are the transshipment nodes. FIGURE 9.6 The Seervada Park maximum flow problem. 5 O 7 4 A 1 3 C 2 4 B 4 5 E 1 D 6 9 T Flow between two nodes only allowed in the direction of the arrow, and at most at the rate of the given capacity. How can we maximize the flow from the source to the destination? the analysis focuses on outgoing trips only.) To avoid unduly disturbing the wildlife of the region, strict upper limits have been imposed on the number trips allowed Maximize per day the in flow the outbound through direction a company s on each individual distribution road. Fo the direction network of travel from for its outgoing factories trips to is indicated its customers. by an arrow in Fig at the base of the arrow gives the upper limit on the number of outgoing trips day. Given Maximize the limits, the one flow feasible through solution a company s is to send 7 trams supply per day, with route Onetwork B Efrom T, its 1 using vendors O to B its Cfactories. E T, and 1 using O E D T. However, because this solution blocks the use of any routes s O C Maximize (because the the E flow T and of Eoil through D capacities a system are fully of used), it is easy ter feasible pipelines. solutions. Many combinations of routes (and the number of trips each one) need to be considered to find the one(s) maximizing the number o Maximize the flow of water through a system of per day. This kind of problem is called a maximum flow problem. In general aqueducts. terms, the maximum flow problem can be described as follow Applications: Maximize the flow of vehicles through a 1. All flow through a directed and connected network originates at one nod source, transportation and terminates network. at one other node, called the sink. (The source and Seervada Park problem are the park entrance at node O and the scenic wo T, respectively.) 2. All the remaining nodes are transshipment nodes. (These are nodes A, B, in the Seervada Park problem.)

22 played in Fig The arcs are listed in columns B and C, and the corresponding arc capacities are given in column F. Since the decision variables are the flows through the re- The Maximum FIGURE 9.11 Flow Problem in Excel A spreadsheet formulation for the Seervada Park maximum flow problem, where the changing cells (D4:D15) show the optimal solution obtained by the Excel Solver and the target cell (D17) gives the resulting maximum flow through the network.

23 its sources (factories, etc.) to intermediate storage facilities (as needed) and then on to the customers. For example, consider the distribution network for the International Paper Company (as described in the March April 1988 issue of Interfaces). This company is the world s The Minimum-Cost Flow Problem At least one of the nodes is a supply node. At least one of the other nodes is a demand node. All the remaining nodes are transshipment nodes. Flow through an arc is allowed only in the direction indicated by the arrowhead, where The network has enough arcs with sufficient capacity to enable all the flow generated at the supply nodes to reach all the demand nodes. The cost of the flow through each arc is proportional to the amount of that flow, where the cost per unit flow is known. The objective is to minimize the total cost of sending the available supply through the network to satisfy the given demand NETWORK OPTIMIZATION MODELS TABLE 9.3 Typical kinds of applications of minimum cost flow problems Kind of Application Supply Nodes Transshipment Nodes Demand Nodes Operation of a Sources of goods Intermediate storage Customers distribution network facilities Solid waste Sources of solid Processing facilities Landfill locations management waste Operation of a supply Vendors Intermediate warehouses Processing network facilities Coordinating product Plants Production of a specific Market for a mixes at plants product specific product Cash flow Sources of cash at Short-term investment Needs for cash at management a specific time options a specific time

24 need to draw on its cash reserves. The demand at each such tion node (e.g., is purchasing the amount a certificate of cash of deposit from a b that will be needed then. The objective is to maximize the resulting company s network income will have from a investing the cash between each time it becomes available and available, when it being will be invested, used. There- and then being used after t succession of flows repre fore, each transshipment node represents the choice of a specific short-term investment option (e.g., purchasing a certificate of deposit from a bank) over Formulation a specific time of interval. the Model The resulting network will have a succession of flows representing Consider a schedule a directed for cash and becoming connected network where t available, being invested, and then being used after the maturing ply node of the and investment. at least one demand node. The decisio The Minimum-Cost Flow Problem x Formulation of the Model ij flow through arc i j, and the given information includes Consider a directed and connected network where the n nodes include at least one supply node and at least one demand node. The decision variables care ij cost per unit flow through arc i j, The objective is to minimize the total cost of sending the uavailable ij arc capacity supply for through arc i thej, network x ij to flow satisfy through the given arc i demand. j, b i net flow generated at node i. and the By given using information the convention includes that summations are taken The only value over of existing b i depends arcs, on the the nature linear programming formulation of this problem is of node i, wh c ij cost per unit flow through arc i j, b i 0 if node i is a supply node, u ij arc capacity for arc i j, b i 0 if node i is a demand node, Minimize b i net flow generated Z n at node c ij x ij, i. b i 0 if node i is a transshipment node. The minimum-cost 9 NETWORK OPTIMIZATION flow problem MODELS can be formulated as follows: Define the decision variables: Write down the objective and constraints: i 1 n j 1 The value of b i depends on the nature of node i, where subject to b i 0 if node i is a supply node, b i 0 if node i is a demand node, b i x ij 0 n if node x ji ib is i, a transshipment for each node node. i, and n j 1 j 1 0 x ij u ij, for each arc i j. The first summation in the node constraints represents the total flow out of node i,whereas the second summation represents the total flow into node i, so the difference is the net flow generated at this node. In some applications, it is necessary to have a lower bound L ij 0 for the flow through

25 x CE x DE x ED 60 and x AB 10, x CE 80, all x ij 0. The Minimum-Cost Problem as a Network Flow 9.12 ibution Unlimited lem formulated as a cost flow problem. b A [50] A c AD 9 [ 30] D 2 (u AB 10) 4 C [0] (u CE 80) B E [40] [ 60]

26 Excel provides a convenient way of formulating and solving small minimum cost flow problems like this one, as well as somewhat larger problems. Figure 9.13 shows how this can be done. The format is almost the same as displayed in Fig for a maximum flow problem. One difference is that the unit costs (c ij ) now need to be included (in column The Minimum-Cost Problem in Excel ation for ited w 10) ution l Solver 12) tal cost nts.