Graduate School of Information, Production and Systems, Waseda University

6. Network Design Problems

6. Network Design Problems

Genetic Algorithms (GAs) are one of the most powerful and broadly applicable stochastic search and optimization techniques based on principles from evolution theory (Holland, 1976):
Michalewicz, Z. : Genetic Algorithm + Data Structure = Evolution Programs, 2nd ed., Springer-Verlag, New York, 1994 Gen, M. & R. Cheng: Genetic Algorithms & Engineering Design, John Wiley & Sons, New York, 1997.

Recent advances in evolutionary computation have made it possible to solve such practical network optimization problems:
Ali, M. & F. Kamoun Neural Networks for Shortest Path Computation and Routing in Computer Networks, IEEE Trans. on Neural Networks, vol.4, pp.941-954, 1993. Perfetti, R. : Optimization Neural Network for Solving Flow Problems, IEEE Trans. on Neural Network, Vol.6, No.5, pp.1287-1291, 1995. Gen, M. & K. Ida: Neural Networks and Optimization with Mathematica, Kyoritsu Shuppan, 1998 in Japanese. Ahn, C. W., R. Ramakrishna, C. Kang & I. Choi: Shortest Path Routing Algorithm using Hopfield Neural Network, Electronic Letter, Vol.37, No.19, pp.1176-1178, 2001.
Soft Computing Lab. WASEDA UNIVERSITY , IPS 2

6. Network Design Problems

In the past few years, the genetic algorithms community has turned much of its attention toward the optimization of network design problems:
Munakata, T. & D. J. Hashier: A genetic algorithm applied to the maximum flow problem, Proc. of the 5th Inter. Conf. on Genetic Algorithms, San Francisco, pp.488-493, 1993. Gen, M. & R. Cheng: Genetic Algorithms and Engineering Design, John Wiley & Sons, New York, 1997. Munetomo, M., Y. Takai & Y. Sato: A migration Scheme for the Genetic Adaptive routing Algorithm, Proc. of IEEE Int. Conf. Systems, Man, and Cybernetics, pp.2774-2779, 1998. Inagaki, J., M. Haseyama & H. Kitajima: A Genetic Algorithm for Determining Multiple Routes and Its Applications, Proc. of IEEE Int. Symp. Circuits and Systems, pp.137-140, 1999. Gen, M. & R. Cheng: Genetic Algorithms and Engineering Optimization, John Wiley & Sons, New York, 2000. Gen, M., R. Cheng & S.S. Oren: "Network design techniques using adapted genetic algorithms", Advances in Engineering Software, Vol.32, pp.731-744, 2001. Ahn, C.W. & R. Ramakrishna: A Genetic Algorithm for Shortest Path Routing Problem and the Sizing of Populations, IEEE Trans. on Evol. Comput., Vol.6, No.6, pp.566-579, 2002. Zhou, G. & M. Gen: A Genetic Algorithm Approach on Tree-like Telecommunication Network Design Problem, J. of Operational Research Society, Vol. 54, No. 3, pp.248-254, 2003.

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

vBNS Backbone Network Map

http://www.mci.com/index.jsp

Soft Computing Lab.

4 WASEDA UNIVERSITY ,very high speed Backbone Network Services vBNS: IPS

vBNS Logical Network Map

http://www.mci.com/index.jsp

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

6. Network Design Problems

1. Shortest Path Problem (SPP) 2. Maximum Flow (MXF) Problem 3. Minimum Cost Flow (MCF) Problem 4. Bicriteria Network Design Problem (BNP) 5. Multi-criteria Network Design Problem

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

6. Network Design Problems

1. Shortest Path Problem (SPP)
1.1 Basic Concept of Shortest Path Problem 1.2 Application of Shortest Path Problem 1.3 Methods for solving SPP 1.4 Genetic Approach for solving SPP
1.4.1 Reviewing Encoding Methods 1.4.2 Priority-based Genetic Algorithm 1.4.3 Genetic Operators

1.5 Numerical Examples

2. 3. 4. 5.

Maximum Flow (MXF) Problem Minimum Cost Flow (MCF) Problem Bicriteria Network Design Problem (BNP) Multi-criteria Network Design Problem
WASEDA UNIVERSITY , IPS 7

Soft Computing Lab.

1. Shortest Path Problem (SPP)

1.1 Basic Concept of Shortest Path Problem
SPP is perhaps the simplest of all network design problems. For this problem, the object is to find a path of minimum cost (or length) from a specified source node s to another specified sink node t, assuming that each arc (i, j)A has an associated cost (or length) cij.
2 36 s 18 16 13 27 3 12 23 6 5 12 13 38 9 4 11 7 32 20 15 8 24 t 10

Data table of example network i j cij 1 1 2 3 3 3 4 4 5 6 6 7 8 8 9 2 3 4 2 5 6 7 8 4 7 9 8 9 10 10 36 27 18 13 12 23 11 32 16 12 38 20 15 24 13

cij

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

1. Shortest Path Problem (SPP)

1.1 Basic Concept of Shortest Path Problem
Directed graph G=(V, A) where V is a set of nodes, A is a set of links. cij is a cost associated with each arc(i, j) Source node: node 1 Destination node: node n Indicator variable:
1 , i ln fik i , 0 , j i i cu e i t ept s nl dd n h ah ohr i e t ew s x ij = {

Data table of example network

i 1 1 2 3 3 3 4 4 5 6 6 7 8 8 9 j 2 3 4 2 5 6 7 8 4 7 9 8 9 10 10 cij 36 27 18 13 12 23 11 32 16 12 38 20 15 24 13

2 36 s

18 16 13 5

4 11

32 20 7 12

8 24 t 15 13 10

1 27 3

12 23 6

cij

38

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

1. Shortest Path Problem (SPP)

1.1 Basic Concept of Shortest Path Problem
SPP can be formulated as follows:
min z = cij xij
i =1 j =1 n n

s.t.

x x
j =1 ij k =1

ki

(i 1= 1) = 0 = 2,3, , n 1) (i 1= n) (i xij = 0 1 j = 1, 2, , n) or (i,

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

10

1. Shortest Path Problem (SPP)

1.2 Application of Shortest Path Problem
This basic model can be applied in many applications such as:
Evans, J. R. and E. Minieka: Optimization Algorithms for Networks and Graphs. New York: Marcel-Dkker, 1992. Transportation Planning
How to determine the route road that have prohibitive weight restriction so that the driver can reach the destination within the shortest possible time. Suppose that a sales person want to go to Los Angeles from Boston and stop over in several city to get some commission. How can she determine the route? How to determine the invest strategy to get an optimal investment plan. The Routing algorithm computes the shortest (least cost) path between the router and all the networks of the internetwork. It is one of the most important issues that has a significant impact on the networks performance.

Salesperson Routing

Investment Planning

Message routing in communication systems

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

11

1. Shortest Path Problem (SPP)

1.2 Application of Shortest Path Problem
With the growth of the Internet, Internet Service Providers (ISPs) try to meet the increasing traffic demand with new technology and improved utilization of existing resources. Routing of data packets can affect network utilization. Packets are sent along network paths from source to destination following a protocol. Open Shortest Path First (OSPF) is the most commonly used protocol.
Ericsson, M., M.G.C. Resende & P.M. Pardalos: A Genetic Algorithm for the Weight Setting Problem in OSPF Routing, J. of Combinatorial Optimization, No.6, pp.299333, 2002.

OSPF is designed for exchanging routing information within a large or very large internetwork. The biggest advantage of OSPF is that it is efficient.
OSPF requires very little network overhead even in very large internetworks. OSPF requires proper planning and is more difficult to configure and administer.
WASEDA UNIVERSITY , IPS 12

The biggest disadvantage of OSPF is its complexity.

Soft Computing Lab.

1. Shortest Path Problem (SPP)

1.2 Application of Shortest Path Problem
OSPF uses a Shortest Path Routing (SPR) algorithm to compute routes in the routing table.
The SPR algorithm computes the shortest (least cost) path between the router and all the networks of the internetwork. Memory requirements and route computation times increase.

As the size of the link state database increases:

Genetic Algorithm (GA) approaches to the SPR problem in OSPF.

Ahn, C.W. & R. Ramakrishna: A Genetic Algorithm for Shortest Path Routing Problem and the Sizing of Populations, IEEE Trans. on Evol. Comput., Vol.6, No.6, pp.566-579, 2002. Lin, L., M. Gen & R. Cheng: Priority-based Genetic Algorithm for Shortest Path Routing Problem in OSPF, Proc. of 3rd Inter. Conf. on Information and Management Sciences, Dunhuang, China, June 5-10, 2004. The objective of this research considers the quality of solution (path optimality) within the shortest route computation times.
WASEDA UNIVERSITY , IPS 13

Soft Computing Lab.

1. Shortest Path Problem (SPP)

1.3 Methods for Solving SPP
Dijkstra Shortest Path Algorithm
Dijkstra, E. W.: "A Note on Two Problems in Connection with Graphs",
Numerische Math., No.1, pp.269-271, 1959. Dijkstra's algorithm can be implemented efficiently by storing the graph in the form of adjacency lists and using a heap as priority queue to implement the Extract-Min function. Computes shortest paths in a graph with non-negative edge weights.

Bellman-Ford Algorithm
Bellman-Ford algorithm computes single-source shortest paths in a weighted graph (where some of the edge weights may be negative). Bellman-Ford is usually used only when there are negative edge weights.

Floyd-Warshall Algorithm
Floyd-Warshall algorithm is an algorithm to solve the all pairs shortest path problem in a weighted, directed graph by multiplying an adjacency-matrix representation of the graph multiple times.

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

14

1. Shortest Path Problem (SPP)

1.4 Genetic Approach for Solving SPP
How to encode a path in a network is critical for designing a GA. Special difficulties:
a path contains variable number of nodes. a random sequence of edges usually does not correspond to a path.
2 36 s 18 4 16 13 27 3 12 23 6 5 12 13 38 9 11 7 32 8 20 15 24 t 10

Path 1 124810 Objective function value z=110 Path 2 1247810 Objective function value z=109 Path 3 13547810 Objective function value z=110
Soft Computing Lab.

WASEDA UNIVERSITY , IPS

15

1. Shortest Path Problem (SPP)

1.4.1 Reviewing Encoding Methods
How to encode a solution of the problem into a chromosome is a key issue for GAs. For the nonstring coding approach, three critical issues emerged concerning with the encoding and decoding between chromosomes and solutions:
The feasibility of a chromosome
Feasibility refers to the phenomenon of whether a solution decoded from a chromosome lies in the feasible region of a given problem. Legality refers to the phenomenon of whether a chromosome represents a solution to a given problem. The illegality of chromosomes originates from the nature of encoding techniques. Repairing techniques are usually adopted to convert an illegal chromosome to a legal one. The mapping from chromosomes to solutions (decoding) may belong to one of the following three cases: (a) 1-to-1 mapping; (b) n-to-1 mapping; (c) 1-to-n mapping. The 1-to-1 mapping is the best one among three cases And 1-to-n mapping is the most undesired one.
WASEDA UNIVERSITY , IPS 16

The legality of a chromosome

The uniqueness of mapping

Soft Computing Lab.

1.4.1 Reviewing Encoding Methods

a. Priority-based Chromosome (Cheng & Gen, 1997)
Cheng & Gen proposed a priority-based encoding method for solving resourceconstrained project scheduling problem (rcPSP) first. And also adopted this method for solving SPP in 1997.
Cheng, R. & M. Gen: Resource Constrained Project Scheduling Problem using Genetic Algorithm, Inter. J. of Intelligent Auto. and Soft Comput., Vol.3, pp.273-286, 1997. Gen, M., R. Cheng & D. Wang: Genetic Algorithms for Solving Shortest Path Problems, Proc. of IEEE Int. Conf. on Evol. Comput., Indianapolis, Indiana, pp.401-406, 1997.

They adopted an indirect approach: The path is generated by sequential node appending procedure with beginning from the specified node 1 and terminating at the specified node n. At each step, there are usually several nodes available for consideration. They gave each node a priority with a random mechanism and add the one with the highest priority into path.
As we know, a gene in a chromosome is characterized by two factors: locus, i.e., the position of gene located within the structure of chromosome, allele, i.e., the value which the gene takes.

In the priority-based encoding method, the position of a gene is used to represent node ID and its value is used to represent the priority of the node for constructing a path among candidates. A path can be uniquely determined from this encoding.

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

17

1.4.1 Reviewing Encoding Methods

a. Priority-based Chromosome (Cheng & Gen, 1997)
Example: An example of generated chromosome and its decoded path as follows:
node ID :1 2 3 4 5 6 7

2 1 s 1 3 4

5 t 1 7 6

priority :2

path :

Advantage:
Any permutation of the encoding corresponds to a path (legality). Most existing genetic operators can be easily applied to the encoding. Any path has a corresponding encoding (completeness); any point in solution space is accessible for genetic search.

Disadvantage:
At some case, n-to-1 mapping may occur for the encoding.
Soft Computing Lab. WASEDA UNIVERSITY , IPS 18

1.4.1 Reviewing Encoding Methods

b. Variable-length Chromosome (Munemoto et al., 1998)
Munemoto et. al. (1998) proposed a variable-length encoding method for network routing problems in a wired or wireless environment. Ahn et. al. (2002) also used the encoding method for solving the shortest path routing (SPR) problem.
Munetomo, M., Y. Takai & Y. Sato: A migration Scheme for the Genetic Adaptive routing Algorithm, Proc. of IEEE Int. Conf. Systems, Man, and Cybernetics, pp.2774-2779, 1998. Ahn, C.W. & R. Ramakrishna: A Genetic Algorithm for Shortest Path Routing Problem and the Sizing of Populations, IEEE Trans. on Evol. Comput., Vol.6, No.6, pp.566-579, 2002.

The proposed encoding method consists of sequences of positive integers that represent the IDs of nodes through which a path passes.
Each locus of the chromosome represents an order of a node (indicated by the gene of the locus) in a path. The length of the chromosome is variable, but is should not exceed the maximum length n, where n is the total number of nodes in the network, since it never needs more than n number of nodes to form a path. The gene of the first locus encodes the source node, and the gene of second locus is randomly or heuristically selected from the nodes connected with the source node. Soft Computing Lab. WASEDA UNIVERSITY , IPS 19

1.4.1 Reviewing Encoding Methods

b. Variable-length Chromosome (Munemoto et al., 1998)
Example: An example of generated chromosome and its decoded path as follows:
locus :1 2 3 4

2 1 s 1 3 4

5 t 1 7 6

node ID :1

path :

Advantage: The mapping from any chromosome to solution (decoding) belongs to 1to-1 mapping (uniqueness). Theoretically, convergence performance is better than the priority-based encoding method. Disadvantage: In general, the genetic operators may generate infeasible chromosomes (illegality) that violate the constraints, generating loops in the paths. Repairing techniques are usually adopted to convert an illegal chromosome to a legal one.
Soft Computing Lab. WASEDA UNIVERSITY , IPS 20

1.4.1 Reviewing Encoding Methods

c. Fixed-length Chromosome (Inagaki et al., 1999) Inagaki et al. (1999) proposed a fixed-length encoding method determining multiple routes in routing applications.
Inagaki, J., M. Haseyama & H. Kitajima: A Genetic Algorithm for Determining Multiple Routes and Its Applications, Proc. of IEEE Int. Symp. Circuits and Systems, pp.137-140, 1999.

The proposed method are sequences of integers and each gene represents the node ID through which it passes.
To encode a route from node 1 to node n, put i in the jth locus of the chromosome. This process is reiterated from the specified node 1 and terminating at the specified node n. If the route does not pass through a node x, select one node randomly from the set of nodes which are connected with node x, and put it in the xth locus.
Soft Computing Lab. WASEDA UNIVERSITY , IPS 21

1.4.1 Reviewing Encoding Methods

c. Fixed-length Chromosome (Inagaki et al., 1999)
Example: An example of generated chromosome and its decoded path as follows:
locus : 1 2 3 4 5 6 7

2 1 s 1 3 4

5 t 1 7 6 node ID :3 1 4 7 2 4 6

path :

Advantage: Any path has a corresponding encoding (completeness). Any point in solution space is accessible for genetic search. Any permutation of the encoding corresponds to a path (legality) using the special genetic operators. Disadvantage: At some case, n-to-1 mapping may occur for the encoding. Furthermore the probability of occurrence of n-to-1 mapping is higher than the priority-based encoding method. In the special genetic operator phase, some offspring may generate new chromosomes that resemble the initial chromosomes in fitness, thereby retarding the process of evolution.
Soft Computing Lab. WASEDA UNIVERSITY , IPS 22

1.4.1 Reviewing Encoding Methods

Compared with the Performance of Different Encoding Methods:
Variable-length encoding method Convergence performance is best than others. However, the genetic operators may generate infeasible chromosomes (illegality). Repairing techniques have to be adopted to convert an illegal chromosome to a legal one. For the computation times, variable-length encoding method may be slow in several large network design problems. Fixed-length encoding method n-to-1 mapping may occur for the encoding. The special genetic operators have to been adopted; thereby some offspring may generate new chromosomes that resemble the initial chromosomes in fitness.

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

23

1.4.2 Priority-based Genetic Algorithm

Priority-based Encoding Method
procedure 1: Priority-based Encoding input: number of nodes n output: chromosome vk begin for j=1 to n vk(j) j; // step 0 // step 1

for i=1 to n / 2 repeat jrandom[1, n]; lrandom[1, n]; until lj swap (vk(j), vk(l)); output the chromosome vk; end
Soft Computing Lab.

// step 2
24

WASEDA UNIVERSITY , IPS

1.4.2 Priority-based Genetic Algorithm

Decoding Method
procedure 2: One Path Growth input: number of nodes n, chromosome vk , the set of nodes Si with all nodes adjacent to node i. output: path Pk begin initial source node i1, Pk ; // step 0 while Si do // step 1 select l from Si with the highest priority; if vk(l)0 then vk(l)=0; Pk Pk{xil }; il; else Si Si \{l} end output the complete path Pk ; // step 2 end
Soft Computing Lab. WASEDA UNIVERSITY , IPS 25

1.4.2 Priority-based Genetic Algorithm

Illustration of Priority-based GA
2 36 1 s 1 27 3 13 5 12 23 12 6 38 13 9 18 16 4 11 7 32 20 15 Data table of example network 8 24 t 1 10
i 1 1 2 3 3 3 4 4 5 6 6 7 8 8 9 j 2 3 4 2 5 6 7 8 4 7 9 8 9 10 10 cij 36 27 18 13 12 23 11 32 16 12 38 20 15 24 13
26

Chromosome:
node ID: j

10

priority: v(j)

8 10 1

Path:

1367810

Objective function value z=106

Soft Computing Lab. WASEDA UNIVERSITY , IPS

1.4.3 Genetic Operators --- Crossover

It operates on two parents (chromosomes) at a time and generates offspring by combining both chromosomes features.
In network design problem, crossover plays the role of exchanging each partial route of two chosen parents in such a manner that the offspring produced by the crossover represents.

In this study, the nature of the priority-based encoding is a kind of permutation representation.
Generally, this representation will yield illegal offspring by one-point crossover or other simple crossover operators.

During the past decade, several crossover operators have been proposed for permutation representation, such as:
Partial-mapped crossover (PMX) Goldberg, D. & R. Lingle, Alleles: loci and the traveling salesman problem, Proc. of the 1st Inter. Conf. on GA, pp.154-159, 1985. Order crossover (OX):

Davis, L. : Applying adaptive algorithms to domains, Proc. of the Inter. Joint Conf. on Artificial Intelligence, pp.1162-164, 1985.
Position-based crossover (PX)

Davis, L. : Applying adaptive algorithms to domains, Proc. of the Inter. Joint Conf. on Artificial Intelligence, pp.1162-164, 1985.
Cycle crossover (CX) Oliver, I. & J. Holland: A study of permutation crossover operators on the traveling salesman problem, Euro. J. of OR, vol.26, pp.187-210, 1986. Heuristic crossover, and so on. Soft Computing Lab. WASEDA UNIVERSITY , IPS

27

1.4.3 Genetic Operators --- Crossover

Partial-Mapped Crossover (PMX)
PMX was proposed by Goldberg and Lingle.
Goldberg, D. & R. Lingle, Alleles: loci and the traveling salesman problem, Proc. of the 1st Inter. Conf. on GA, pp.154-159, 1985.

PMX can be viewed as an extension of two-point crossover for binary string to permutation representation. It uses a special repairing procedure to resolve the illegitimacy caused by the simple two-point crossover. step 3 : determine mapping relationship step 1 : select the substring at random
substring selected
parent 1: parent 2: 1 4 7 6 2 3 3 5 4 7 6 1 5 8 8 2 3 5 7 2 3 4

235 47

step 2 : exchange substrings between

parent 1: parent 2: 1 4 7 6 3 2 5 3 7 4 6 1 5 8 8 2

step 4 : legalize offspring with mapping relationship

offspring 1: offspring 2: 1 7 4 6 3 2 5 3 7 4 6 1 2 8 8 5
28

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

1.4.3 Genetic Operators --- Crossover

Order Crossover (OX)
OX was proposed by Davis.
Davis, L. : Applying adaptive algorithms to domains, Proc. of the Inter.
Joint Conf. on Artificial Intelligence, pp.1162-164, 1985.

It can be viewed as a kind of variation of PMX with a different repairing procedure.

parent 1: offspring: 1 6 7 5 2 2 3 3 4 4 6 7 5 1 8 8

substring selected

parent 2:

Fig. 6.1 Illustration of the OX operator.

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

29

1.4.3 Genetic Operators --- Crossover

Position-based Crossover (PX)
PX was proposed by Syswerda.
Davis, L. : Applying adaptive algorithms to domains, Proc. of the Inter.
Joint Conf. on Artificial Intelligence, pp.1162-164, 1985.

It is essentially a kind of uniform crossover for permutation representation together with a repairing procedure. It also can be viewed as a kind of variation of OX in which the nodes are selected inconsecutively.
parent 1: offspring: 1 3 7 7 2 5 3 1 4 4 6 6 5 2 8 8

parent 2:

Fig. 6.2 Illustration of the PX operator.

Soft Computing Lab. WASEDA UNIVERSITY , IPS 30

1.4.3 Genetic Operators --- Crossover

However, in all of above approaches:
the mechanism of the crossover is not the same as that of the conventional one-point crossover. Some offspring may generate new chromosomes that are not possible to succeed the character of the parents. thereby retarding the process of evolution.

We proposed a new crossover operator, Weight Mapping Crossover (WMX).

WMX can be viewed as an extension of one-point crossover for permutation representation. As one-point crossover:
Two chromosomes (parents) would be to choose a random cut-point. Generate the offspring by using segment of own parent to the left of the one-cut point Then remapping the right segment that base on the weight of other parent of right segment .
Soft Computing Lab. WASEDA UNIVERSITY , IPS 31

1.4.3 Genetic Operators --- Crossover

Weight Mapping Crossover (WMX)
procedure : Weight Mapping Crossover input : v1,v2 , n output : v1', v2' begin p random[1, n]; l n p; for i = 1 to l s1[i ] v1[ p + i ]; s2 [i ] v2 [ p + i ]; s1[] sorting ( s1[]); s2 [] sorting ( s2 []); v1' v1[1 : p ] // v2 [ p + 1 : n]; v2' v2 [1 : p ] // v1[ p + 1 : n];
Soft Computing Lab.

for i = 1 to l for j = 1 to l if v1'[ p + i ] = s2 [ j ] then v1'[ p + i ] s1[ j ]; for j = 1 to l if v2'[ p + i ] = s1[ j ] then v2'[ p + i ] s2 [ j ]; output v1' , v2' ; end

WASEDA UNIVERSITY , IPS

32

1.4.3 Genetic Operators --- Crossover

Weight Mapping Crossover (WMX)
As shown Fig., first we choose a random cut-point p. calculate l that is the length of right segments of chromosomes, where n is number of nodes in the network. Then get mapping relationship by sorting the weight of the right segments s1[] and s2[]. As one-point crossover, generate the offspring v1, v2 by exchange substrings between parents v1, v2; legalize offspring with mapping relationship.
step 1: select a cut-point
parent 1 : parent 2 : 2 3 1 7 7 2 4 6 3 1 cut-point 5 5 5 4 3 1 6 5 6 4

parent 1 : 1 parent 2 : 1

3 2

4 4

7 5 7

step 2: mapping the weight of the right segment

5 5 3 1 6 4

offspring 1 : 1 offspring 2 : 1
3 1 5 5

3 2

4 4

5 7

step 3: generate offspring with mapping relationship

offspring 1 : offspring 2 : 2 3 1 7 7 2 4 6 6 4

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

33

1.4.3 Genetic Operators --- Mutation

It is relatively easy to produce some mutation operators for permutation representation. During the past decade, several mutation operators have been proposed for permutation representation, such as:
Inversion Insertion Displacement Swap mutation.

Insertion Mutation
Selects a gene at random and inserts it in a random position as follows:
select a gene at random parent : 2 1 7 4 5 3 6 parent : 1 3 4 7

insert it in a random position offspring : Soft Computing Lab. 2 5 1 7 4 3 6 offspring : 1 4 7 34

WASEDA UNIVERSITY , IPS

1.4.3 Genetic Operators --- Immigration

The trade-off between exploration and exploitation in serial GAs for function optimization is a fundamental issue.
If a GA is biased towards exploitation:
highly fit members are repeatedly selected for recombination. Although this quickly promotes better members, the population can prematurely converge to a local optimum of the function.

If a GA is biased towards exploration: Large numbers of schemata are sampled which tends to inhibit premature convergence. Unfortunately, excessive exploration results in a large number of function evaluations, and defaults to random search in the worst case.

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

35

1.4.3 Genetic Operators --- Immigration

To search effectively and efficiently, a GA must maintain a balance between these two opposing forces.
Michael, C.M., C.V. Stewart & R.B. Kelly: Reducing the Search Time of A Steady State Genetic Algorithm using the Immigration Operator, Proc. of IEEE Int. Conf. on Tools for AI San Jose, CA, pp.500-501, 1991. Michael et. al. (1991) proposed an immigration operator which, for certain types of functions, allows increased exploration while maintaining nearly the same level of exploitation for the given population size.

Immigration operator
step 1: The algorithm is modified to include immigration, with each generation generated. step 2: Evaluate random members (, called the immigration rate). step 3: Replace the worst members of the population with the random members.

This study experimentally examines the immigration operator, and present the effectiveness of this approach for solving network design problems in next section.
Soft Computing Lab. WASEDA UNIVERSITY , IPS 36

1.4.3 Genetic Operators --- Selection

Selection operators: two basic types of selection scheme used commonly in current practice.
Proportionate selection: picks out chromosomes based on their fitness values relative to the fitness of the other chromosomes in the population.
Roulette wheel selection Stochastic remainder selection Stochastic universal selection

Ordinal-based selection: upon their rank within the population. The chromosomes are ranked according to their fitness values.
Tournament selection ( , ) selection Truncation selection Linear ranking selection

In this study, the roulette wheel selection, a type of Proportionate selection, is adopted.
Soft Computing Lab. WASEDA UNIVERSITY , IPS 37

1.4.4 Overall Procedure

GA Procedure for Shortest Path Problem
procedure: Priority-based GA for Shortest Path Problem input: network data (V, A, C), GA parameters output: best shortest path begin t 0; initialize P(t) by priority-based encoding; fitness eval(P); while (not termination condition) do crossover P(t) to yield C(t) by weight mapping crossover; mutation P(t) to yield C(t) by insertion mutation; immigration operation to yield C(t) fitness eval(C); select P(t+1) from P(t) and C(t) by roulette wheel selection; t t + 1; end output best shortest path; end Soft Computing Lab. WASEDA UNIVERSITY , IPS

38

1.5 Numerical Examples

Test Problems:
For examining the effect of different encoding methods, we applied Ahn et als method and priority-based encoding method to the 6 test problems:
Ahn, C.W. & R. Ramakrishna: A Genetic Algorithm for Shortest Path Routing Problem and the Sizing of Populations. IEEE Trans. Evol. Comput., Vol.6, No.6, pp.566-579, 2002. OR-Notes. [Online]. Available: http://mscmga.ms.ic.ac.uk/jeb/or/orweb.html

Using the following parameter specifications.

Population size: popSize =20 Crossover probability: pC =0.70 Mutation probability: pM =0.50 Immigration rate: =3 Maximum generation: maxGen =1000 Terminating condition: 100 generations with same fitness.

Each solution is compared with Dijkstras algorithm that provides a reference point (optimal solution). Each algorithm was applied to each test problem 20 times (i.e., 20 runs) using different initial populations. All the simulations were performed with Java on Pentium 4 processor (1.5-GHz clock).
Soft Computing Lab. WASEDA UNIVERSITY , IPS 39

1.5 Numerical Examples

The first numerical example, presented by Ahn et als was adopted. The problem comprises 20 nodes and 49 arcs. It is given as follows:
(Ahn, C.W. & R. Ramakrishna: A Genetic Algorithm for Shortest Path Routing Problem and the Sizing of Populations, IEEE Trans. on Evol. Comput., Vol.6, No.6, pp.566579, 2002.)

1 1

Soft Computing Lab.

WASEDA UNIVERSITY , IPS Fig.6.3 Example of the first numerical example

40

1.5 Numerical Examples

Convergence property of each algorithm for a Fixed Network With 20 Nodes
(Ahn, C.W. & R. Ramakrishna: A Genetic Algorithm for Shortest Path Routing Problem and the Sizing of Populations, IEEE Trans. on Evol. Comput., Vol.6, No.6, pp.566-579, 2002.)
2.5

Objective Function Values

2.0

Dijkstras Algorithm Munemotos Algorithm Inagakis Algorithm Ahns Algorithm

1.5

1.0

0.5 0 2 4 6 Generations 8 10

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

41

1.5 Numerical Examples

Convergence property of Ahn et al.s algorithm and proposed algorithm for a Fixed Network With 20 Nodes
et al.

260
Fig. 6.4 Convergence property of Ahn et al.s algorithm and proposed algorithm.
Soft Computing Lab. WASEDA UNIVERSITY , IPS 42

240

1.5 Numerical Examples

Comparison with results
(Ahn, C.W. & R. Ramakrishna: A Genetic Algorithm for Shortest Path Routing Problem and the Sizing of Populations, IEEE Trans. on Evol. Comput., Vol.6, No.6, pp.566-579, 2002.)
1200

Inagakis Algorithm Munemotos Algorithm Ahns Algorithm

Objective Function Value

1000

800

600

400

200

0 15 20 25 30 35 40 The Number of Nodes 45 50

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

43

1.5 Numerical Examples

Discussion of the Results:
Table 6.1 Performance comparisons with different genetic operators
Test Problems (# of nodes/ # of arcs) Optimal Solutions Alg. 1 Alg. 2 Best Solutions Alg. 3 Alg. 4 143.93 396.52 297.21 382.48 629.81 552.71 Alg. 5 142.00 389.00 291.62 284.69 395.01 331.09 Alg. 6 142.00 389.00 291.00 284.00 394.00 288.00

20/49 142.00 148.35 148.53 147.70 80/120 389.00 423.53 425.33 418.82 80/632 291.00 320.06 311.04 320.15 160/2544 284.00 429.55 454.98 480.19 320/1845 394.00 754.94 786.08 906.18 320/10208 288.00 794.26 732.72 819.85 Alg. 1: FMX+Swap; Alg. 2: OX+Swap; Alg. 3: PX+Swap; Alg. 4: WMX+Swap; Alg. 5: WMX+Swap+Immigration(3); Alg. 6: WMX+Insertion+Immigration(3).

The quality of solution with different genetic operators is investigated in Table 1. The path optimality is defined in all test problems, by Alg.6 (WMX+Insertion+ Immigration) that the GA finds the global optimum (i.e., the shortest path). The path optimality is defined in #1, #2 test problems, by Alg.5 (WMX+Swap+ Immigration), The near optimal result is defined in other test problems. By Alg.1 ~ Alg.4, the path optimality is not defined. Since the number of possible alternatives become to very large in test problems, the population be prematurely converged to a local optimum of the function.
WASEDA UNIVERSITY , IPS 44

Soft Computing Lab.

1.5 Numerical Examples

Comparison results of Ahns algorithm and Proposed algorithm
Table 6.2 Performance comparisons with Ahns algorithm and Proposed algorithm.
Test Problems (# of nodes/ # of arcs) Optimal Solutions Best Solutions Ahns Alg. 20/49 80/120 80/632 160/2544 320/1845 320/10208 142.00 389.00 291.00 284.00 394.00 288.00 142.00 389.00 291.00 286.20 403.40 288.90 Prop. Alg. 142.00 389.00 291.00 284.00 394.00 288.00 CPU Times (ms) Ahns Alg. 40.60 118.50 109.50 336.20 779.80 1028.30 Prop. Alg. 23.37 96.80 118.50 490.50 1062.50 1498.50 Generation Num. of Obtained best result Ahns Alg. Prop. Alg. 2 4 19 31 44 38 9 4 10 26 11 26

Best solutions

Soft Computing Lab.

20/49

80/120

WASEDA UNIVERSITY , IPS Problem size

80/630

160/2544

320/1845

320/10208

45

1.5 Numerical Examples

Different Parameter Settings:
Table 6.3 Performance comparisons with different parameter settings
Parameter Settings ( popSize / pC : pM ) Test Problems (# of nodes/ # of arcs) Optimal Solutions Best Solutions Ahns Alg. 10 / 0.3 : 0.1 20/49 80/120 80/632 160/2544 320/1845 320/10208 20/49 80/120 80/632 160/2544 320/1845 320/10208 20/49 80/120 80/632 160/2544 320/1845 320/10208 142.00 389.00 291.00 284.00 394.00 288.00 142.00 389.00 291.00 284.00 394.00 288.00 142.00 389.00 291.00 284.00 394.00 288.00 156.20 389.00 313.20 320.90 478.70 444.00 145.23 389.00 303.10 298.70 465.70 373.10 142.00 389.00 291.00 286.20 403.40 288.90 Prop. Alg. 142.00 389.00 291.00 284.20 394.00 288.30 142.00 389.00 291.00 284.20 394.00 288.60 142.00 389.00 291.00 284.00 394.00 288.00 CPU Times ( ms ) Ahns Alg. 10.42 32.80 29.40 67.10 120.30 126.40 22.36 56.30 50.10 122.10 213.90 311.00 40.60 118.50 109.50 336.20 779.80 1028.30 Prop. Alg. 8.37 31.10 34.40 106.30 250.20 400.20 13.34 51.50 56.30 181.20 496.70 631.10 23.37 96.80 118.50 490.50 1062.50 1498.50 Generation Num. of Obtained best result Ahns Alg. 38 5 43 48 68 25 27 4 18 44 32 61 6 1 19 31 44 38 Prop. Alg. 27 1 16 37 18 59 24 1 10 35 17 35 9 1 10 26 11 26

20 / 0.3 : 0.1

20 / 0.7 : 0.5

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

46

1.5 Numerical Examples

Different Parameter Settings with Ahns algorithm and Proposed algorithm
Parameter Settings ( popSize / pC : pM ) 10 / 0.3 : 0.1 20 / 0.3 : 0.1 30 / 0.3 : 0.1 50 / 0.3 : 0.1 100 / 0.3 : 0.1 10 / 0.7 : 0.5 20 / 0.7 : 0.5 30 / 0.7 : 0.5 50 / 0.7 : 0.5 100 / 0.7 : 0.5 Probability of obtaining the optimal solutions Ahns Alg. Proposed Alg. 16.67% 16.67% 33.33% 50.00% 33.33% 33.33% 50.00% 50.00% 83.33% 83.33% 66.67% 66.67% 83.33% 100.00% 100.00% 83.33% 100.00% 100.00% 100.00% 100.00%

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

47

1.5 Numerical Examples

Simulation (# of nodes: 100, # of arcs: 859)

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

48

6. Network Design Problems

1. 2. Shortest Path Problem (SPP) Maximum Flow (MXF) Problem
2.1 Basic Concept of Maximum Flow Problem 2.2 Application of Maximum Flow Problem 2.3 Methods for solving MXF Problem 2.4 Genetic Approach for solving MXF Problem 2.4.1 Genetic Representation 2.4.2 Genetic Operators 2.5 Numerical Examples

1. 2. 3.

Minimum Cost Flow (MCF) Problem Bicriteria Network Design Problem (BNP) Multi-criteria Network Design Problem
WASEDA UNIVERSITY , IPS 49

Soft Computing Lab.

2. Maximum Flow (MXF) Problem

2.1 Basic Concept of Maximum Flow Problem
[Online]. Available: http://www-b2.is.tokushima-u.ac.jp/ ~ikeda/suuri/maxflow/Maxflow.shtml.en MXF is in a sense a complementary model to SPP. MXF seeks a feasible solution that sends the maximum amount of flow from a specified source node s to another specified sink node t. If we interpret uij as the maximum flow rate of arc (i, j), MXF identifies the maximum steady-state flow that the network can send from node s to node t per unit time.
2 60 f s 1 60 60 30 3 30 4 Soft Computing Lab. 40 30 50 30 40 5 20 6 20 7 60 30 40 30 40 30 9 30 10 8 60 50 50 i t 11 f Data table of example network
i 1 1 1 2 2 2 3 3 3 4 5 6 6 6 6 7 7 8 8 9 9 10 j 2 3 4 3 5 6 4 6 7 7 8 5 8 9 10 6 10 9 11 10 11 11 uij 60 60 60 30 40 30 30 50 30 40 60 20 30 40 30 20 40 30 60 30 50 50

uij

WASEDA UNIVERSITY , IPS

50

2. Maximum Flow (MXF) Problem

2.1 Basic Concept of Maximum Flow Problem
Directed graph G=(V, A) where V is a set of nodes, A is a set of links. uij is a capacity associated with each link(i, j) Source node: node 1 Destination node: node n
2 60 f s 1 60 60 30 3 30 4 40 30 50 30 40 5 20 6 20 7 60 30 40 30 40 30 9 30 10 8 60 50 50 t 11 f Data table of example network
i 1 1 1 2 2 2 3 3 3 4 5 6 6 6 6 7 7 8 8 9 9 10 j 2 3 4 3 5 6 4 6 7 7 8 5 8 9 10 6 10 9 11 10 11 11 uij 60 60 60 30 40 30 30 50 30 40 60 20 30 40 30 20 40 30 60 30 50 50

i Soft Computing Lab.

uij

WASEDA UNIVERSITY , IPS

51

2. Maximum Flow (MXF) Problem

2.1 Basic Concept of Maximum Flow Problem
MXF problem can be formulated as follows:

max z = f
f (i = 1) s. t. xij xki = 0 (i = 2,3, , n 1) j =1 k =1 f (i = n) 0 xij uij , (i, j ) A
n n

f 0

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

52

2. Maximum Flow (MXF) Problem

2.2 Application of Maximum Flow Problem
This basic MXF model can be applied in many applications such as:
Ahuj, R. K., T. L. Magnanti & J. B. Orlin: Network Flows. Prentice-Hall, Upper Saddle River, NJ, 1993. Scheduling on Uniform Parallel Machines
The feasible scheduling problem, described in the preceding paragraph, is a fundamental problem in this situation and can be used as a subroutine for more general scheduling problems, such as the maximum lateness problem, the (weighted) minimum completion time problem, and the (weighted) maximum utilization problem. Distributed computing on a two-processor computer concerns assigning different modules (subroutines) of a program to two processors in a way that minimizes the collective costs of interprocessor communication and computation.. A steamship company has contracted to deliver perishable goods between several different origin-destination pairs. Since the cargo is perishable, the customers have specified precise dates (i.e., delivery dates) when the shipments must reach their destinations..
WASEDA UNIVERSITY , IPS 53

Distributed Computing on a Two-Processor Computer

Tanker Scheduling Problem

Soft Computing Lab.

2. Maximum Flow (MXF) Problem

2.3 Methods for solving MXF Problem
Ford-Fulkerson Algorithm
It works by finding a flow augmenting path in the graph. By adding the flow augmenting path to the flow already established in the graph, the maximum flow will be reached when no more flow augmenting paths can be found in the graph.

Maximum Flow Algorithm

An incremental algorithm for max-flow problem that tries to find the max-flow in the network as an edge is deleted or inserted in the network, is presented. It has also been shown that other cases of a unit change can be considered as a special case of insertion and deletion of an edge in the network.

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

54

2. Maximum Flow (MXF) Problem

2.4 Genetic Approach for solving MXF Problem
Munakata, T. and D. J. Hashier: A genetic algorithm applied to the maximum flow problem, Proc. of 5th Int. Conf. on Genetic Algorithms, pp. 488-493, 1993.

The maximum flow problem appears to be more challenging in applying GAs than many other common graph problems (e.g., shortest path, minimum spanning tree) Its unique characteristic:
A flow at each edge can be anywhere between zero and its flow capacity, i.e., it has more "freedom" to choose. In many other problems, selecting an edge may mean to simply add a fixed distance. The flow at each edge must be between zero and its flow capacity. At each vertex, the incoming flow and outgoing flow must balance.

In the maximum flow problem, two conditions must be satisfied:

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

55

2.4 Genetic Approach for solving MXF Problem

2.4.1 Genetic Representation
procedure 1: Priority-based Encoding input: number of nodes n output: chromosome vk begin for j=1 to n vk(j) j; // step 0 // step 1

for i=1 to n / 2 repeat jrandom[1, n]; lrandom[1, n]; until lj swap (vk(j), vk(l)); end
Soft Computing Lab.

output the chromosome vk; // step 2

WASEDA UNIVERSITY , IPS 56

2.4 Genetic Approach for solving MXF Problem

The decoding procedure is a two-stage process.
First stage: the path is generated by one-path growth procedure
It is given in procedure 2 With beginning from the specified node 1 and terminating at the specified node n. At each step, add the one with the highest priority into path. procedure 2: One-path Growth
input: number of nodes n, chromosome vk , the set of nodes Si with all nodes adjacent to node i. output: path Pk step 0: the source node i1, Pk step 1: if Si= , goto step 3; otherwise, continue. step 2: select l from Si with the highest priority, and go back to step 1. if vk(l)0 then vk(l)=0; Pk Pk{xil }; il; else vk(l)=0

1 1 2 2 step 3: output the complete path3Pk ;

Pk = {x1l , xl ,l , xl

,l

, ..., xlm 1 ,lm }

57

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

2.4 Genetic Approach for solving MXF Problem

The decoding procedure is a two-stage process.
Second stage: overall paths are generated by overall paths growth procedure
For a given path, we can calculate its flow fk By removing the used capacity from uij of each arc, we have a new network with the new flow capacity ij. With the one-path growth procedure (procedure 2), we can obtain the second path. By repeating this procedure we can obtain the maximum flow for the given chromosome till no new network can be defined in this way. It is given in procedure 3.

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

58

2.4 Genetic Approach for solving MXF Problem

procedure 3: Overall-path Growth
input: network data (V, A, U), chromosome vk , the set of nodes Si with all nodes adjacent to node i output: number of paths Lk , the flow fik of each path, iLk step 0: number of paths l0 step 1: if S1= , go to step 7; otherwise, l l +1, continue. step 2: the implementation of path Plk growth is based on procedure 2. Select the sink node a of path plk. step 3: if the sink node a=n, continue; otherwise, perform the set of nodes Si update as follows, return to step S i 1.

S i {a}, i

step 4: calculate the kflow flk of the path Plk. k

f l f l 1 + min{uij | (i, j ) Plk }

~ step 5: perform the flow min{u uiji, j each arc update. Make a new flow capacity ij as follows: u u capacity ( of ) P k }
ij ij ij l

step 6: if the flow capacity ij =0, perform the set of nodes Si update which the node j adjacent to node i.

~ si si { jpaths iL, j ) l Pl k & u ijf = 0 path, iL step 7: output number of } , ( -1, the flow of each
k i k

. 59

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

2.4 Genetic Approach for solving MXF Problem

Illustration of Priority-based GA
2 60 f s 1 60 60 30 3 30 4 40 30 50 30 40 5 20 6 20 7 60 30 40 30 40 30 9 30 10 8 60 50 50 t 11 f Data table of example network
i 1 1 1 2 2 2 3 3 3 4 5 6 6 6 6 7 7 8 8 9 9 10 j 2 3 4 3 5 6 4 6 7 7 8 5 8 9 10 6 10 9 11 10 11 11 uij 60 60 60 30 40 30 30 50 30 40 60 20 30 40 30 20 40 30 60 30 50 50

Chromosome:
node ID :1 2 3 4 5 6 7 8 9 10 11

priority :2

11

10

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

60

2.4 Genetic Approach for solving MXF Problem

Illustration of Priority-based GA

Chromosome:
node ID :1 2 3 4 5 6 7 8 9 10 11

priority :2
k 1 i 0 1 2 3 4 5 0 1 2 3 4 Si

4
l 3 6 5 8 11 3 6 8 11

11 9
Pk

10 5

3
S1

7
fk

2, 3, 4 4, 6, 7 5, 8, 9, 10 8 9, 11 2, 3, 4 4, 6, 7 8, 9, 10 9, 11

1 1, 3 1, 3, 6 1, 3, 6, 5 1, 3, 6, 5, 8 1, 3, 6, 5, 8, 11 1 1, 3 1, 3, 6 1, 3, 6, 8 1, 3, 6, 8, 11

k : number of paths i : start node Si : the set of nodes l : sink node Pk : the kth path S1 : the set of nodes with all nodes adjacent to node 1 fk : maximum possible flow

2, 3, 4 20

60

2, 3, 4 50

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

3
61

2.4 Genetic Approach for solving MXF Problem

Illustration of Priority-based GA

Chromosome:
node ID :1 2 3 4 5 6 7 8 9 10 11

priority :2
k 3 i 0 1 2 3 4 5 0 1 2 3 4 5 Si

4
l

11 9
Pk

10 5

3
S1

7
fk

2, 3, 4 4, 7 6, 10 9, 10 10, 11 2, 4 7 6, 10 9, 10 10, 11

3 7 6 9 11 4 7 6 9 11

1 1, 3 1, 3, 7 1, 3, 7, 6 1, 3, 7, 6, 9 1, 3, 7, 6, 9, 11 1 1, 4 1, 4, 7 1, 4, 7, 6 1, 4, 7, 6, 9 1, 4, 7, 6, 9, 11

k : number of paths i : start node Si : the set of nodes

2, 4 60

l : sink node Pk : the kth path S1 : the set of nodes with all nodes adjacent to node 1 fk : maximum possible flow

60

2,4

70

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

3
62

2.4 Genetic Approach for solving MXF Problem

Illustration of Priority-based GA

Chromosome:
node ID :1 2 3 4 5 6 7 8 9 10 11

priority :2

11 9

10 5

k 5

i 0 1 2 3 4 0 1 2 3 4

Si 2, 4 7 10 11 2 3, 5, 6 8 9, 11

l 4 7 10 11 2 5 8 11

Pk 1 1, 4 1, 4, 7 1, 4, 7, 10 1, 4, 7, 10, 11 1 1, 2 1, 2, 5 1, 2, 5, 8 1, 2, 5, 8, 11

S1

fk

k : number of paths i : start node Si : the set of nodes l : sink node Pk : the kth path S1 : the set of nodes with all nodes adjacent to node 1 fk : maximum possible flow

100

60

110

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

3
63

2.4 Genetic Approach for solving MXF Problem

Illustration of Priority-based GA

Chromosome:
node ID :1 2 3 4 5 6 7 8 9 10 11

priority :2
k 7 i 0 1 2 3 4 5 0 1 2 3 4 5

1
Si

11 9
l 2 5 8 9 11 2 6 9 10 11 Pk

10 5

3
S1

7
fk

2 3, 5, 6 8 9 10, 11 2 3, 6 9, 10 10 11

1 1, 2 1, 2, 5 1, 2, 5, 8 1, 2, 5, 8, 9 1, 2, 5, 8, 9, 11 1 1, 2 1, 2, 6 1, 2, 6, 9 1, 2, 6, 9, 10 1, 2, 6, 9, 10, 11

k : number of paths i : start node Si : the set of nodes

2 140

l : sink node Pk : the kth path S1 : the set of nodes with all nodes adjacent to node 1 fk : maximum possible flow

60

160

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

3
64

2.4 Genetic Approach for solving MXF Problem

Illustration of Priority-based GA
2 60/60 160 s 1 60/60 40/60 4 3 40/40 5 60/60 8 60/60 50/50 t 11 160
Data table of example network
i 1 1 1 2 2 2 3 3 3 4 5 6 6 6 6 7 7 8 8 9 9 10 j 2 3 4 3 5 6 4 6 7 7 8 5 8 9 10 6 10 9 11 10 11 11 uij 60 60 60 30 40 30 30 50 30 40 60 20 30 40 30 20 40 30 60 30 50 50 65

20/30 20/20 30/30 30/30 50/50 6 40/40 9 20/30 30/40

7 8

10/30 20/20 40/40

4 5

50/50

7
6

10
9 10 11

Chromosome:3 node ID :1 2
priority :2 1 6

11

10

Objective function value z=160

i Soft Computing Lab.

xij / uij

WASEDA UNIVERSITY , IPS

2.4 Genetic Approach for solving MXF Problem

2.4.2 Genetic Operators
Here the position-based crossover operator proposed by PMX (Partial Mapped Crossover) (Gen-Cheng97, pp.119-125) was adopted. It uses a special repairing procedure to resolve the illegitimacy caused by the simple two-point crossover as follows:
step 1 : select the substring at random
substring selected
parent 1: parent 2: 1 4 7 6 2 3 3 5 4 7 6 1 5 8 8 2 3 5 7

step 3 : determine mapping relationship

2 3 4

235 47

step 2 : exchange substrings between

parent 1: parent 2: 1 4 7 6 3 2 5 3 7 4 6 1 5 8 8 2

step 4 : legalize offspring with mapping relationship

offspring 1: offspring 2: 1 7 4 6 3 2 5 3 7 4 6 1 2 8 8 5
66

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

2.4 Genetic Approach for solving MXF Problem

2.4.2 Genetic Operators
Mutation: The swap mutation operator was used here, in which two positions are selected at random and their contents are swapped as follows:
exchanging points

parent: offspring:

1 1

7 7

2 6

3 3

4 4

6 2

5 5

8 8

Selection: The roulette wheel approach, a type of fitnessproportional selection, was adopted.

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

67

2.4 Genetic Approach for solving MXF Problem

GA Procedure for Maximum Flow Problem
procedure: Priority-based GA for Maximum Flow Problem input: network data (V, A, U), GA parameters output: best maximum flow begin t 0; initialize P(t) by priority-based encoding; fitness eval(P); while (not termination condition) do crossover P(t) to yield C(t) by partial mapped crossover; mutation P(t) to yield C(t) by swap mutation; fitness eval(C); select P(t+1) from P(t) and C(t) by roulette wheel selection; t t + 1; end output best maximum flow; end
Soft Computing Lab. WASEDA UNIVERSITY , IPS 68

2.5 Numerical Examples

Test Problems:
The numerical examples, presented by T. Munakata & D.J. Hashier, was adopted.
Munakata, T. and D. J. Hashier: A genetic algorithm applied to the maximum flow problem, Proc. of 5th Int. Conf. on Genetic Algorithms, pp. 488-493, 1993.

Using the following parameter specifications.

Population size: popSize =10 Crossover probability: pC =0.50 Mutation probability: pM =0.50 Maximum generation: maxGen =1000 Terminating condition: 100 generations with same fitness.

All the simulations were performed with Java on Pentium 4 processor (1.5-GHz clock).
Soft Computing Lab. WASEDA UNIVERSITY , IPS 69

2.5 Numerical Examples

Test Problem 1:
The first numerical example, presented by Munakata & Hashier, was adopted. The problem comprises 25 nodes and 49 arcs. It is given as follows:
2
6 20 20 10 8 5 4 5 8 10 10 20 4 10 10 10

15 15

12

20 20

17
8 25 25 20 10 25 10 25 25

15 15

13

20 20

18
8

22
30

20 20

15 15

14

20 20

19
8

23

30 30

25

10

15 15

15

20 20

20

24
i

uij

6
Soft Computing Lab.

11

16

20

21
70

WASEDA UNIVERSITY , IPS

2.5 Numerical Examples

2
4/6 18/20 13/20 10/10 8/8

14/15

12
9/20

17
5/25 4/8

5/5 4 /4 2/5

15/15 15/15

13

20/20 20/20

18
4/8

25/25 15/20

22
30/30

90

20/20 19/20

8/8 10/10

7/15 15/15

14

10/20 2/20

19

5/10 5/25

23 30/30
30/30

25

90

10/10 20/20

6/8 10/10

4/4 5/10 10/10 10/10

10

9/15

15

20/20 9/20

20 25/25

24

15/15

11

16

21

xij / uij

Objective function value: z=90 (optimal solution) Generation Num. of Obtained best result: 34 Best Chromosome:
node ID : 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

priority:

25

12 15 20

16 21 14

18 23

13

17

11

24 19 10 22

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

71

2.5 Numerical Examples

Process of Genetic Computing

Soft Computing Lab.

flow

WASEDA UNIVERSITY , IPS

72

2.5 Numerical Examples

Test Problem 2:
The first numerical example, presented by Munakata & Hashier, was adopted. The problem comprises 25 nodes and 56 arcs. It is given as follows:
2
10 20 20 5 10 8 5

11
2 18 7 5

11 8

20
9

7
6 8 6

16
10 7

12
12 8

8 10

21
20 15 8

15

8
9 8 7

17
10

20

4
7

13
20 6 9

10

22
8

30 20

25

20

9
10 7 12 5

18
6 5 6

5
20

14
22 6 8

23
15 9

10
15

19
10

uij

Soft Computing Lab.

15 WASEDA UNIVERSITY , IPS 24

73

2.5 Numerical Examples

2
10/10 8/8 10/10 18/20 20/20 5/5

11 7
18/18 4/5

8/11 2/2 7/8

20

16
6/10

9/9

6/6 4/8

12
9/12

4/8

21
5/7 20/20 8/15 8/8

15/15

8
7/9

17
6/10 1/10

91

20/20

8/8 7/7

13 9
15/20

22

22/30 20/20

25 91

20/20

18
6/6 4/5

8/8 6/6

5
13/20

10/10 3/7 7/12

14
7/22 3/6

23
14/15

10
6/15

19
9/10

5/9

uij

15

24

Objective function value: z=91 (optimal solution) Generation Num. of Obtained best result: 67 Best Chromosome:
node ID : 1 2 3 4 5 6 7 8 9 10 11 12

13

14

15

16

17

18

19

20

21

22

23

24

25

priority: 20 10

22 25 23 11

18

16 12

14 15 19

13 17

21 24

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

74

2.5 Numerical Examples

Process of Genetic Computing

Soft Computing Lab.

flow

WASEDA UNIVERSITY , IPS

75

2.5 Numerical Examples

Simulation (# of nodes: 80, # of arcs: 826)

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

76

6. Network Design Problems

1. 2. 3. Shortest Path Problem (SPP) Maximum Flow (MXF) Problem Minimum Cost Flow (MCF) Problem
3.1 Basic Concept of MCF Problem 3.2 Application of MCF Problem 3.3 Methods for solving MCF Problem 3.4 Genetic Approach for solving MCF Problem

3.4.1 Genetic Representation 3.4.2 Genetic Operators 3.5 Numerical Examples

4. 5.

Bicriteria Network Design Problem (BNP) Multicriteria Network Design Problem

WASEDA UNIVERSITY , IPS 77

Soft Computing Lab.

3. Minimum Cost Flow (MCF) Problem

3.1 Basic Concept of MCF Problem
MCF model is the most fundamental of all network design problems. The problem is to determine the minimum cost plan for sending flow through the network to satisfy supply and demand requirements. The problem is defined by a given set of arcs and a given set of nodes, where each are has a known capacity uij and unit cost cij and each node has a fixed external flow. SPP model is a special case of MCF when the flow q =1.
2
18, 60 16, 40 14, 30 13, 30 15, 20

Data table of example network

i 1 1 1 2 2 2 3 3 3 4 5 6 6 6 6 7 7 8 8 9 9 10 j 2 3 4 3 5 6 4 6 7 7 8 5 8 9 10 6 10 9 11 10 11 11 cij 18 19 17 13 16 14 15 16 17 19 19 15 16 15 18 15 13 17 18 14 19 17 uij 60 60 60 30 40 30 30 50 30 40 60 20 30 40 30 20 40 30 60 30 50 50

19, 60

8
18, 60 19, 50

s 1

16, 30 17, 30 15, 40 18, 30 13, 40

19, 60

16, 50 17, 30

t 11

6
15, 20

17, 60

15, 30

14, 30 17, 50

19, 40

10 i

cij , uij

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

78

3. Minimum Cost Flow (MCF) Problem

3.1 Basic Concept of MCF Problem
Directed graph G=(V, A) where V is a set of nodes, A is a set of links. uij is a capacity associated with each link(i, j) cij is unit cost associated with each link(i, j) Source node: node 1 Destination node: node n
2
18, 60 16, 40 14, 30 13, 30 15, 20

Data table of example network

i 1 1 1 2 2 2 3 3 3 4 5 6 6 6 6 7 7 8 8 9 9 10 j 2 3 4 3 5 6 4 6 7 7 8 5 8 9 10 6 10 9 11 10 11 11 cij 18 19 17 13 16 14 15 16 17 19 19 15 16 15 18 15 13 17 18 14 19 17 uij 60 60 60 30 40 30 30 50 30 40 60 20 30 40 30 20 40 30 60 30 50 50

19, 60

8
18, 60 19, 50

s 1

16, 30 17, 30 15, 40 18, 30 13, 40

19, 60

16, 50 17, 30

t 11

6
15, 20

17, 60

15, 30

14, 30 17, 50

19, 40

10

i Soft Computing Lab.

cij , uij

WASEDA UNIVERSITY , IPS

79

3. Minimum Cost Flow (MCF) Problem

3.1 Basic Concept of MCF Problem
MCF problem can be formulated as follows:

min z = cij xij

i =1 j =1

q (i = 1) s. t. xij xki = 0 (i = 2,3, , n 1) i =1 k =1 q (i = n)

n n

0 xij uij , (i, j ) A

q: total flow value
Soft Computing Lab. WASEDA UNIVERSITY , IPS 80

3. Minimum Cost Flow (MCF) Problem

3.2 Application of Minimum Cost Flow (MCF) Problem
This basic MCF model can be applied in many applications such as:
Ahuj, R. K., T. L. Magnanti & J. B. Orlin: Network Flows, Prentice-Hall, Upper
Saddle River, NJ, 1993. Transportation Problem There are a set of nodes called sources, and a set of nodes called destinations. All arcs go from a source to a destination. There is a per-unit cost on each arc. Each source has a supply of material, and each destination has a demand. It can be solved by applying Min-cost Flow Algorithm Distribution Problem The distribution of a product from manufacturing plants to warehouses, or from warehouses to retailers The flow of raw material and intermediate goods through the various machining stations in a production line The routing of automobiles through an urban street network The routing of calls through the telephone system.
WASEDA UNIVERSITY , IPS 81

Soft Computing Lab.

3. Minimum Cost Flow (MCF) Problem

3.3 Methods for solving MCF Problem
Ahuj, R. K., T. L. Magnanti & J. B. Orlin: Network Flows, Prentice-Hall, Upper Saddle River, NJ, 1993.

Successive Shortest Path Algorithm

The successive shortest path algorithm maintains optimality of the solution at every step and strives to attain feasibility. The primal-dual algorithm for the minimum cost flow problem is similar to the successive shortest path algorithm in the sense that it also maintains a pseudoflow that satisfies the reduced cost optimality conditions and gradually converts it into a flow by augmenting flows along shortest paths. The out-of-kilter algorithm, which satisfies only the mass balance constraints, so intermediate solutions might violate both the optimality conditions and the flow bound restrictions.

Primal-dual Algorithm

Out-of-Kilter Algorithm

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

82

3.4 Genetic Approach for solving MCF Problem

3.4.1 Genetic Representation
procedure 1: priority-based encoding input: number of nodes n output: chromosome vk begin for j=1 to n vk(j) j; // step 0 // step 1

for i=1 to n / 2 repeat jrandom[1, n]; lrandom[1, n]; until lj swap (vk(j), vk(l)); end
Soft Computing Lab.

output the chromosome vk; // step 2

WASEDA UNIVERSITY , IPS 83

3.4 Genetic Approach for solving MCF Problem

The decoding procedure is a two-stage process.
First stage: the path is generated by One-path growth procedure
It is given in procedure 2 With beginning from the specified node 1 and terminating at the specified node n. At each step, add the one with the highest priority into path. procedure 2: One-path Growth
input: number of nodes n, chromosome vk , the set of nodes Si with all nodes adjacent to node i. output: path Pk step 0: the source node i1, Pk step 1: if Si= , goto step 3; otherwise, continue. step 2: select l from Si with the highest priority, and go back to step 1. if vk(l)0 then vk(l)=0; Pk Pk{xil }; il; else =k{x , x P v (l)=0
k 1l1

step 3: output the complete path Pk . Soft Computing Lab.

l1 ,l2

, xl2 ,l3 , ..., xlm 1 ,lm }

84

WASEDA UNIVERSITY , IPS

3.4 Genetic Approach for solving MCF Problem

The decoding procedure is a two-stage process.
Second stage: overall paths are generated by overall paths growth procedure For a given path, we can calculate its flow fk and the cost ck. By removing the used capacity from uij of each arc, we have a new network with the new flow capacity ij. With the one-path growth procedure (procedure 2), we can obtain the second path. By repeating this procedure we can obtain the maximum flow for the given chromosome till the flow fk is larger than total flow value q. It is given in procedure 3.

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

85

3.4 Genetic Approach for solving MCF Problem

procedure 3: Overall-path Growth input: network data (V, A, C, U), chromosome vk , the set of nodes Si with all nodes adjacent to node i output: number of paths Lk , the flow fik and the cost cik of each path, iLk step 0: number of paths l0 step 1: if flow fk >q, go to step 7; otherwise, l l +1, continue. step 2: the implementation of path Plk growth is based on procedure 2. Select the sink node a of path plk. step 3: if the sink node a=n, continue; otherwise, perform the set of nodes Sii update asifollows, return to step 1. S i {a},

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

86

3.4 Genetic Approach for solving MCF Problem

step 4: calculate the flow flk and the cost clk of the path Plk.
k f l k f l 1 + min{uij | (i, j ) Plk } k + cij ( f l k f l 1 ), (i, j ) Pl k i =1 j =1 n n

c c
k l

k l 1

step 5: perform the flow capacity uij of each arc update. Make a new flow capacity ij as follows:

~ u ij u ij min{u ij (i, j ) Pl k }

step 6: if the flow capacity ij =0, perform the set of nodes Si update which the node j adjacent to node i. ~ k

si si { j} , (i, j ) Pl & u ij = 0

step 7: output number of paths Lk l -1, the flow fik and the cost cik of
Soft Computing Lab.

each path, iLk .

WASEDA UNIVERSITY , IPS

87

3.4 Genetic Approach for solving MCF Problem

Illustration of Priority-based GA
2
18, 60 16, 40 14, 30 13, 30 15, 20

Data table of example network

i 1 1 1 2 2 2 3 3 3 4 5 6 6 6 6 7 7 8 8 9 9 10 j 2 3 4 3 5 6 4 6 7 7 8 5 8 9 10 6 10 9 11 10 11 11 cij 18 19 17 13 16 14 15 16 17 19 19 15 16 15 18 15 13 17 18 14 19 17 uij 60 60 60 30 40 30 30 50 30 40 60 20 30 40 30 20 40 30 60 30 50 50

19, 60

8
18, 60 19, 50

s 1

16, 30 17, 30 15, 40 18, 30 13, 40

19, 60

16, 50 17, 30

t 11

6
15, 20

17, 60

15, 30

14, 30 17, 50

19, 40

10

cij , uij

Chromosome:
node ID :1 2 3 4 5 6 7 8 9 10 11

priority :2

11

10

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

88

3.4 Genetic Approach for solving MCF Problem

Illustration of Priority-based GA

Chromosome:
node ID :1 2 3 4 5 6 7 8 9 10 11

priority :2

11 9

10 5

= 60
k 1 i 0 1 2 3 4 5 0 1 2 3 4 2, 3, 4 4, 6, 7 5, 8, 9, 10 8 9, 11 2, 3, 4 4, 6, 7 8, 9, 10 9, 11 3 6 5 8 11 3 6 8 11 Si l Pk 1 1, 3 1, 3, 6 1, 3, 6, 5 1, 3, 6, 5, 8 1, 3, 6, 5, 8, 11 1 1, 3 1, 3, 6 1, 3, 6, 8 1, 3, 6, 8, 11 S1 flk
clk

2, 3, 4 20

87

k : number of paths i : start node Si : the set of nodes l : sink node Pk : the kth path S1 : the set of nodes with all nodes adjacent to node 1 flk : the total flow clk : minimum possible cost
89

2, 3, 4 30

69

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

3.4 Genetic Approach for solving MCF Problem

Illustration of Priority-based GA

Chromosome:
node ID :1 2 3 4 5 6 7 8 9 10 11

priority :2

11 9

10 5

= 60
k 3 i 0 1 2 3 4 5 2, 3, 4 4, 7 6, 10 9, 10 10, 11 3 7 6 9 11 Si l Pk 1 1, 3 1, 3, 7 1, 3, 7, 6 1, 3, 7, 6, 9 1, 3, 7, 6, 9, 11 S1 flk
clk

2, 4

10

85

k : number of paths i : start node Si : the set of nodes l : sink node Pk : the kth path S1 : the set of nodes with all nodes adjacent to node 1 flk : the total flow clk : minimum possible cost
90

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

3.4 Genetic Approach for solving MCF Problem

Illustration of Prioritybased GA
2 s 1 5
15, 20/20 19, 60/60 19, 20/60 16, 30/30

i 8

cij , xij /uij

Data table of example network j

i 1 1 1 2 2 2 3 3 3 4 5 6 6 6 6 7 7 8 8 9 9 10 j 2 3 4 3 5 6 4 6 7 7 8 5 8 9 10 6 10 9 11 10 11 11 cij 18 19 17 13 16 14 15 16 17 19 19 15 16 15 18 15 13 17 18 14 19 17 uij 60 60 60 30 40 30 30 50 30 40 60 20 30 40 30 20 40 30 60 30 50 50 91

18, 50/60 19, 10/50

60

16, 50/50 17, 10/30

15, 10/40

t 11

60

15, 10/20

7
4 5 6 7 8 9

10
10 11

Chromosome:3 node ID :1 2
priority :2 1 6

11

10

Objective function value z=20*87+30*69+10*85 =4660

Soft Computing Lab. WASEDA UNIVERSITY , IPS

3.4 Genetic Approach for solving MCF Problem

3.4.2 Genetic Operators
Here the position-based crossover operator proposed by PMX (Partial Mapped Crossover) (Gen-Cheng97, pp.119-125) was adopted. It uses a special repairing procedure to resolve the illegitimacy caused by the simple two-point crossover as follows:
step 1 : select the substring at random
substring selected
parent 1: parent 2: 1 4 7 6 2 3 3 5 4 7 6 1 5 8 8 2 3 5 7

step 3 : determine mapping relationship

2 3 4

235 47

step 2 : exchange substrings between

parent 1: parent 2: 1 4 7 6 3 2 5 3 7 4 6 1 5 8 8 2

step 4 : legalize offspring with mapping relationship

offspring 1: offspring 2: 1 7 4 6 3 2 5 3 7 4 6 1 2 8 8 5
92

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

3.4 Genetic Approach for solving MCF Problem

3.4.2 Genetic Operators
Mutation: The swap mutation operator was used here, in which two positions are selected at random and their contents are swapped as follows:
exchanging points

parent: offspring:

1 1

7 7

2 6

3 3

4 4

6 2

5 5

8 8

Selection: The roulette wheel approach, a type of fitnessproportional selection, was adopted.

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

93

3.4 Genetic Approach for solving MCF Problem

GA Procedure for solving MCF Problem
procedure: Priority-based GA for Solving MCF Problem input: network data (V, A, C, U), GA parameters output: best minimum cost begin t 0; initialize P(t) by priority-based encoding; fitness eval(P); while (not termination condition) do crossover P(t) to yield C(t) by partial mapped crossover; mutation P(t) to yield C(t) by swap mutation; fitness eval(C); select P(t+1) from P(t) and C(t) by roulette wheel selection; t t + 1; end output best minimum cost; end
Soft Computing Lab. WASEDA UNIVERSITY , IPS 94

3.5 Numerical Examples

Test Problems:
The numerical examples, presented by Munakata & Hashier, was adopted.
Munakata, T. and D. J. Hashier: A genetic algorithm applied to the maximum flow problem, Proc. of 5th Int. Conf. on Genetic Algorithms, pp. 488-493, 1993.

Using the following parameter specifications.

Population size: popSize =10 Crossover probability: pC =0.50 Mutation probability: pM =0.50 Maximum generation: maxGen =1000 Terminating condition: 100 generations with same fitness.

All the simulations were performed with Java on Pentium 4 processor (1.5-GHz clock).
Soft Computing Lab. WASEDA UNIVERSITY , IPS 95

3.5 Numerical Examples

Test Problem 1:
The first numerical example, presented by Munakata & Hashier, was adopted. The problem comprises 25 nodes and 49 arcs. It is given as follows:
2
10, 10 13, 8 10, 6 10, 20 15, 5 33, 4 4, 5 15, 8 35, 10

10, 15 4, 15

12

8, 20 6, 20

17
11, 8 11, 25

3
13, 20

9, 15

13

7, 20 7, 20

18
5, 8

8, 25 34, 20 10, 10 35, 25

22
10, 30

11, 15

32, 20

9 12, 15
9, 15

14

7, 20 5, 20

19

23 2, 30
3, 30

25

135, 20 631, 20

3, 10 13, 4 33, 10 7, 10 7, 10

3, 8 3, 10

10

14, 15

15

14, 20 4, 20

20 14, 25
12, 25

24

5, 15

6
Soft Computing Lab.

11

16

14, 20

21

cij , uij

j 96

WASEDA UNIVERSITY , IPS

3.5 Numerical Examples

2
10, 10/10 13, 8/8 10, 5/6 10, 18/20 15, 5/5 33, 4/4

10, 15/15

12
6, 10/20

17
11, 10/25

8 9, 15/15

13

7, 20/20

18 8, 20/25 22
10, 30/30

13, 14/20 11, 15/15 4, 2/5 q=70 32, 20/20 15, 8/8 1 4 9 12, 7/15 35, 10/10 135, 14/20 3, 10/10 9, 3/15 13, 4/4 5 10 631, 4/20 14, 15/15 7, 4/10

14

7, 18/20

19

10, 10/10

q=70 23 2, 20/30 25

15

3, 8/8 3, 20/30 3, 10/10 14, 7/20 2014, 20/25 24 4, 15/20

11

16

21

Total flow value q = 70 Objective function value: z= 6969 Generation Num. of Obtained best result:863 Best Chromosome:
node ID : 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

priority:

16 11

15 10

12 13 21

22 14 18 20 24 17 25 23

19

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

97

3.5 Numerical Examples

Process of Genetic Computing

cost

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

98

3.5 Numerical Examples

Test Problem 2:
The first numerical example, presented by Munakata & Hashier, was adopted. The problem comprises 25 nodes and 56 arcs. It is given as follows:
2
3, 10 14, 8 8, 10 2, 20 3, 5 7, 18 12, 7 4, 6 11, 5

11

29, 11 13, 2 6, 8

20

16

26, 9

3
10, 20

13, 6 32, 8 7, 5 6, 12 12, 8 9, 7

12

30, 10 6, 8 9, 10 12, 7

21
10, 20 31, 15 8, 4

2,15

17

33, 20

12, 9 33, 8 4, 7 12, 10 30, 7 11, 12 8, 20 12, 6 3, 5

13

34, 10 15, 10 2, 9

22

6, 30

25

126, 20

18

7, 20 30, 8 11, 6

5
634, 20

14

35, 6 14, 5 7, 8

23
9, 15

10

11, 22 6, 6

19

35, 9

Soft Computing Lab.

12, 15

WASEDA UNIVERSITY , IPS

15

8, 10

24

cij , uij

j 99

3.5 Numerical Examples

2
3, 10/10 14, 8/8

11
7, 8/18

29, 10/11

20
26, 8/9

7
2, 18/20

16
30, 6/10

13, 6/6 32, 8/8 7, 3/5

12
6, 6/12

21
12, 6/7 10, 20/20 31, 3/15 8,4/ 4

2,10/15

10, 14/20

17
34,9/10

q=72

33, 20/20

12, 9/9 33, 8/8 4, 7/7 12, 10/10 30, 3/7

13 9
8, 15/20

22

6, 16/30

25

q=72

126, 20/20

2, 3/9 3, 5/5

18

7, 20/20 30, 8/8 11, 4/6

14

35, 6/6 14, 2/5

23
9, 6/15

10 6

11,2 /22 6, 6/6

19
8, 6/10

15

24

Total flow value q = 72 Objective function value: z=5986 Generation Num. of Obtained best result:132 Best Chromosome:
node ID : 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

Soft Computing Lab.

priority: 10 25 15 24 11

WASEDA UNIVERSITY , IPS

12

13 14

16 17 18

20 22 19 23 21

100

3.5 Numerical Examples

Process of Genetic Computing
7000 6900 6800 6700 6600

cost

6500 6400 6300 6200 6100 6000 5900 0 50 100 150 200 250

generation Soft Computing Lab. WASEDA UNIVERSITY , IPS 101

3.5 Numerical Examples

Simulation (# of nodes: 80, # of arcs: 857)

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

102

6. Network Design Problems

1. 2. 3. 4. Shortest Path Problem (SPP) Maximum Flow (MXF) Problem Minimum Cost Flow (MCF) Problem Bicriteria Network Design Problem (BNP)
4.1 Introduction of BNP
4.2 BNP Formulation 4.3 Genetic Approach for solving BNP 4.3.1 Genetic Representation 4.3.2 Decoding Method 4.3.3 Fitness Assignment 4.3.4 Genetic Operators 4.4 Numerical Examples

5. Multi-criteria Network Design Problem

Soft Computing Lab. WASEDA UNIVERSITY , IPS 103

4. Bicriteria Network Design Problem (BNP)

In real-life applications, network design is often the case that the network to be built is required to optimize multicriteria simultaneously.
Marathe, M. V., R. Ravi, R. Sundaram, S. S. Ravi, D. J. Rosenkrantz, and H. B. Hunt: Bicriteria network design problems, J. Algorithms, vol. 28, no. 1, pp. 142-171, Jul. 1998. Lo, C. and W. Chang: A multiobjective hybrid genetic algorithm for the capacitated multipoint network design problem, IEEE Trans. Syst., Man, Cybern. B, vol. 30, no. 3, pp. 461-470, Jun. 2000. Kim, J. R. and M. Gen: A genetic algorithm for bicriteria communication network topology design, Eng. Val. Cost Analysis, vol. 3, pp. 351-363, 2000. Raghavan, S., M. O. Ball, and V. S. Trichur: Bicriteria product design optimization, Institute for Systems Research, Tech. Rep. TR 2001-8, 2001. [Online]. Available: http://techreports.isr.umd.edu/ARCHIVE/ Kumar, R., P. P. Parida, and M. Gupta: Topological design of communication networks using multiobjective genetic optimization, Proc. Cong. Evol. Comput., May. 2002, vol. 1, pp. 425-430. Yuan, D.: A bicriteria optimization approach for robust OSPF routing, Proc. IPOM, 2003, pp. 91-98. Medaglia, A. L. and S. Fang: A genetic-based framework for solving (multi-criteria) weighted matching problems, Eur. J. Oper. Res., vol. 149, pp.77-101, Jan. 2003. Yang, H., M. Maier, M. Reisslein, and W. M. Carlyle: A genetic algorithm-based methodology for optimizing multiservice convergence in a metro WDM network, J. Lightwave Technol., vol. 21, no. 5, pp. 1114-1133, May. 2003. Zhou, G., H. Min, and M. Gen: A genetic algorithm approach to the bi-criteria allocation of customers to warehouses, Int. J. Production Economics, vol. 86, pp. 35-45, Oct. 2003.

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

104

4.1 Introduction of BNP

The problems may arise when designing:
In a communication network, find a set of links which consider the low cost (or delay) and the high throughput (or reliability) for increasing the network performance.
Marathe, M. V., R. Ravi, R. Sundaram, S. S. Ravi, D. J. Rosenkrantz, and H. B. Hunt: Bicriteria network design problems, J. Algorithms, vol. 28, no. 1, pp. 142-171, Jul. 1998. Yuan, D.: A bicriteria optimization approach for robust OSPF routing, Proc. IPOM, 2003, pp. 91-98. Yang, H., M. Maier, M. Reisslein, and W. M. Carlyle: A genetic algorithm-based methodology for optimizing multiservice convergence in a metro WDM network, J. Lightwave Technol., vol. 21, no. 5, pp. 1114-1133, May. 2003.

In a manufacturing system, the two criteria under consideration are minimizing cost and maximizing manufacturing.
Raghavan, S., M. O. Ball, and V. S. Trichur: Bicriteria product design optimization, Institute for Systems Research, Tech. Rep. TR 2001-8, 2001. [Online]. Available: http://techreports.isr.umd.edu/ARCHIVE/

In a logistic system, the main drive to improve logistics productivity is the enhancement of customer services and asset utilization through a significant reduction in order cycle time (lead time) and logistics costs.
Zhou, G. , H. Min, and M. Gen: A genetic algorithm approach to the bi-criteria allocation of customers to warehouses, Int. J. Production Economics, vol. 86, pp. 35-45, Oct. 2003. WASEDA UNIVERSITY , IPS 105

Soft Computing Lab.

4.1 Introduction of BNP

The Bicriteria Network Design Problem (BNP) is known as NP-hard (Garey and Johnson, 1979), it is not simply an extension from single objective to two objectives.
Garey, M. and D. Johnson: Computers and Intractability: A Guide to the Theory of NP-Completeness, W.H. Freeman, New York, 1979. In generally, we can not get the optimal solution of the problem because these objectives usually conflict with each other in practice. The real solutions to the problem are a set of Pareto optimal solutions (Chankong and Haimes, 1983).
Chankong, V. and Y.Y. Haimes: Multiobjective Decision Making Theory and Methodology. North-Holland, New York, 1983.

For solving the BNP, the set of efficient paths may be very large and possibly exponential in size.
Thus the computational effort required to solve it can increase exponentially with the problem size in the worst case.

While the tractability of the problem is of importance when solving large scale problems, the issue concerning with the size of the efficient set is important to a decision maker. Having to evaluate a large efficient set in order to select the best one poses a considerable cognitive burden on decision makers. Therefore, in such cases, obtaining the entire Pareto optimal set is of little interest to decision makers.
WASEDA UNIVERSITY , IPS 106

Soft Computing Lab.

4.1 Introduction of BNP

Recently, GAs have received considerable attention regarding their potential as a novel approach to multiobjective optimization problems, known as evolutionary or genetic multiobjective optimization.
Deb, K. : Multiobjective Optimization Using Evolutionary Algorithms, Wiley, Chichester, UK, 2001. The basic feature of GAs is the multiple directional and global search by maintaining a population of potential solutions from generation to generation. The population-to-population approach is hopeful to explore Pareto optimal or nondominated solutions. GAs do not have much mathematical requirements about the problems and can handle any kind of objective functions and constraints. Due to their evolutionary nature, GAs can search for solutions without regard to the specific inner workings of the problem. Therefore, GAs is possibly well suited to the multiobjective optimization problems.

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

107

4.1 Introduction of BNP

The bicriteria shortest path problem is one of BNPs, which of finding a diameter-constrained shortest path from a specified source node s to another specified sink node t.
This problem, termed the multi-objective shortest path problem (MOSP) in the literature is NP-complete.
Warburton (1987) presented the first fully polynomial approximation scheme (FPAS) for it.
Warburto, A.: Approximation of Pareto optima in multiple-objective, shortest path problems, Operations Research, vol. 35, no. 1, pp. 70-79, 1987.

Hassin (1992) provided a strongly polynomial FPAS for the problem which improved the running time of Warburton.
Hassin, R.: Approximation schemes for the restricted shortest path problem, Math. Of Operations Research, vol. 17, no. 1, pp. 36-42, Feb. 1992.

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

108

4.1 Introduction of BNP

The general classes of BNPs with minimum two objectives (under different cost functions) are defined and extended to the more multi-criteria network design problems.
Ravi et al. (1994) presented an approximation algorithm for finding good broadcast networks.
Ravi, R.: "Rapid rumor ramification: approximating the minimum broadcast time, Proc. 35th Annual IEEE Foundations of Comput. Sci., pp. 202-213, 1994.

Ganley et al. (1995) consider a more general problem with more than two objective functions.
Ganley, J. L., M. J. Golin, and J. S. Salowe: The multi-weighted spanning tree problem, Proc. 1st COCOON, pp. 141-150, Springer-Verlag, LNCS, 1995.

Marathe et al. (1998) consider three different criteria of network and presented the first polynomial-time approximation algorithms for a large class of BNP.
Marathe, M. V., R. Ravi, R. Sundaram, S. S. Ravi, D. J. Rosenkrantz, and H. B. Hunt: Bicriteria network design problems, J. Algorithms, vol. 28, no. 1, pp. 142-171, 1998.

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

109

4.1 Introduction of BNP

In this study, we dominated BNP with more complexity cases as two criteria problem that maximum flow and minimum cost considered.
Priority-based encoding method (Cheng and Gen, 1994) has been improved.
Cheng, R. and M. Gen: Evolution program for resource constrained project scheduling problem, Proc. of Int. Conf. Evol. Comput., pp.736-741, 1994. For maximizing flow, different form other genetic representation methods, such as path oriented encoding method, priority-based encoding method can represent various efficient paths by each chromosome.

Considering the characteristic of priority-based encoding method, we proposed a new crossover operator called as Weight Mapping Crossover (WMX) Insertion mutation operator and Immigration operator (Michael et al., 1991) was adopted.
Michael, C.M., C.V. Stewart and R. B. Kelly: Reducing the search time of a steady state genetic algorithm using the immigration operator, Proc. IEEE Int. Conf. Tools for AI, San Jose, CA, pp.500-501, 1991.

These methods provide a search capability that results in improved quality of solution and enhanced rate of convergence. For ensure the population diversity in MOGA, Adaptive Weight Approach (AWA) which is one of weighted-sum approach, was adopted.
Gen, M. and R. Cheng: Genetic Algorithms and Engineering Optimization, John Wiley & Sons, New York, 2000. Their elements represent that weights are adjusted adaptively based on the current generation to obtain search pressure toward the positive ideal point.
WASEDA UNIVERSITY , IPS 110

Soft Computing Lab.

4.2 BNP Formulation

4.2 BNP Formulation
In this study, we present a mathematical programming formulation of the bicriteria network design model including MXF model and MCF model. Different from the generic BNP, the problems efficient set of paths may be very large, possibly exponential in size. Thus the computational effort required to solve it can increase exponentially with the problem size in the worst case.

In a network with flow capacities and costs on the arcs, BNP is to determine both the maximum possible flow z1 and minimum cost z2 in the same time, from a source to a sink.

max z1 = f min z 2 = cij xij

i =1 j =1 n n n

s. t.

f (i = 1) xij xki = 0 (i = 2,3,, n 1) j =1 k =1 f (i = n)

n

0 xij uij , (i, j ) A f 0

Soft Computing Lab. WASEDA UNIVERSITY , IPS 111

4.3.1 Genetic Representation

Priority-based encoding method
procedure 1: Priority-based Encoding input: number of nodes n output: chromosome vk begin for j=1 to n vk(j) j; // step 0 // step 1

for i=1 to n / 2 repeat jrandom[1, n]; lrandom[1, n]; until lj swap (vk(j), vk(l)); end
Soft Computing Lab.

output the chromosome vk; // step 2

WASEDA UNIVERSITY , IPS 112

4.3.2 Decoding Method

The decoding procedure is a two-stage process.
First stage: the path is generated by One-path growth procedure
It is given in procedure 2 With beginning from the specified node 1 and terminating at the specified node n. At each step, add the one with the highest priority into path. procedure 2: One-path Growth
input: number of nodes n, chromosome vk , the set of nodes Si with all nodes adjacent to node i. output: path Pk step 0: the source node i1, Pk step 1: if Si= , goto step 3; otherwise, continue. step 2: select l from Si with the highest priority, and go back to step 1. if vk(l)0 then vk(l)=0; Pk Pk{xil }; il; else vk(l)=0 Soft Computing Lab.

step 3: output the complete path Pk ;

Pk = {x1l1 , xl1 ,l2 , xl2 ,l3 , ..., xlm 1 ,lm }

WASEDA UNIVERSITY , IPS 113

4.3.2 Decoding Method

The decoding procedure is a two-stage process.
Second stage: overall paths are generated by overall paths growth procedure
For a given path, we can calculate its flow fk and the cost ck. By removing the used capacity from uij of each arc, we have a new network with the new flow capacity ij. With the one-path growth procedure (procedure 2), we can obtain the second path. By repeating this procedure we can obtain the maximum flow for the given chromosome till no new network can be defined in this way. It is given in procedure 3.

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

114

4.3.2 Decoding Method

procedure 3: Overall-path Growth input: network data (V, A, C, U), chromosome vk , the set of nodes Si with all nodes adjacent to node i output: number of paths Lk , the flow fik and the cost cik of each path, iLk step 0: number of paths l0 step 1: if S1= , go to step 7; otherwise, l l +1, continue. step 2: the implementation of path Plk growth is based on procedure 2. Select the sink node a of path plk. step 3: if the sink node a=n, continue; otherwise, perform the set of S i S ii update i follows, return to step 1. nodes S {a}, as

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

115

4.3.2 Decoding Method

step 4: calculate the flow flk and the cost clk of the path Plk.
k f l k f l 1 + min{uij | (i, j ) Plk } k + cij ( f l k f l 1 ), (i, j ) Pl k i =1 j =1 n n

c c
k l

k l 1

step 5: perform the flow capacity uij of each arc update. Make a new flow capacity ij as follows:

~ u ij u ij min{u ij (i, j ) Pl k }

step 6: if the flow capacity ij =0, perform the set of nodes Si update which the node j adjacent to node i. ~ k

si si { j} , (i, j ) Pl & u ij = 0

step 7: output number of paths Lk l -1, the flow fik and the cost cik of
Soft Computing Lab.

each path, iLk .

WASEDA UNIVERSITY , IPS

116

Illustration of Decoding Method

Chromosome:

k 1

i 0 1 2 3 4 5 0 1 2 3 4

Si 2, 3, 4 4, 6, 7 5, 8, 9, 10 8 9, 11 2, 3, 4 4, 6, 7 8, 9, 10 9, 11

l 3 6 5 8 11 3 6 8 11

Pk 1 1, 3 1, 3, 6 1, 3, 6, 5 1, 3, 6, 5, 8 1, 3, 6, 5, 8, 11 1 1, 3 1, 3, 6 1, 3, 6, 8 1, 3, 6, 8, 11

S1

z1k

z2k

k : number of paths i : start node Si : the set of nodes l : sink node

2, 3, 4 20 1380

Pk : the kth path S1 : the set of nodes with all nodes

adjacent to node 1 z1k : maximum possible flow

2, 3, 4 50 5730

z2k : minimum possible cost

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

18, 60

117

Illustration of Decoding Method

Chromosome:

k 3

i 0 1 2 3 4 5 0 1 2 3 4 5

Si 2, 3, 4 4, 7 6, 10 9, 10 10, 11 2, 4 7 6, 10 9, 10 10, 11

l 3 7 6 9 11 4 7 6 9 11

Pk

S1

z1k

z2k

1 1, 3 1, 3, 7 1, 3, 7, 6 1, 3, 7, 6, 9 1, 3, 7, 6, 9, 11 2, 4 1 1, 4 1, 4, 7 1, 4, 7, 6 1, 4, 7, 6, 9 1, 4, 7, 6, 9, 11 2,4

k : number of paths i : start node Si : the set of nodes

60 6580

l : sink node Pk : the kth path S1 : the set of nodes with all nodes

adjacent to node 1 z1k : maximum possible flow

70 7430

z2k : minimum possible cost

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

18, 60

118

Illustration of Decoding Method

Chromosome:

k 5

i 0 1 2 3 4 0 1 2 3 4

Si 2, 4 7 10 11 2 3, 5, 6 8 9, 11

l 4 7 10 11 2 5 8 11

Pk 1 1, 4 1, 4, 7

S1

z1k

z2k

k : number of paths i : start node Si : the set of nodes l : sink node

1, 4, 7, 10 1, 4, 7, 10, 11 2 1 1, 2 1, 2, 5 1, 2, 5, 8 1, 2, 5, 8, 11 2

100

9410

Pk : the kth path S1 : the set of nodes with all nodes

110

10120

adjacent to node 1 z1k : maximum possible flow

z2k : minimum possible cost

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

18, 60

119

Illustration of Decoding Method

Chromosome:

k 7

i 0 1 2 3 4 5 0 1 2 3 4 5

Si 2 3, 5, 6 8 9 10, 11 2 3, 6 9, 10 10 11

l 2 5 8 9 11 2 6 9 10 11

Pk 1 1, 2 1, 2, 5 1, 2, 5, 8 1, 2, 5, 8, 9 1, 2, 5, 8, 9, 11 1 1, 2 1, 2, 6 1, 2, 6, 9 1, 2, 6, 9, 10 1, 2, 6, 9, 10, 11

S1

z1k

z2k

k : number of paths i : start node Si : the set of nodes

2 140 12790

l : sink node Pk : the kth path S1 : the set of nodes with all nodes

adjacent to node 1 z1k : maximum possible flow

160 14350

z2k : minimum possible cost

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

18, 60

120

Illustration of Decoding Method

Illustration of Priority-based GA
- cost
-0 -2000 -4000 -6000 -8000 -10000 -12000
140, 12790 50, 5730 60, 6580 70, 7430 100, 9410 110, 10120 40, 1380

Data table of example network

i 1 1 1 2 2 2 3 3 3 4 5 6 6 6 6 7 7 8 8 9 9 10 j 2 3 4 3 5 6 4 6 7 7 8 5 8 9 10 6 10 9 11 10 11 11 cij 18 19 17 13 16 14 15 16 17 19 19 15 16 15 18 15 13 17 18 14 19 17 uij 60 60 60 30 40 30 30 50 30 40 60 20 30 40 30 20 40 30 60 30 50 50 121

-14000 -16000
0 20 40 60 80 100 120 140 160, 14350 160 180

flow

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

4.3.3 Fitness Assignment

In this study, evaluate the fitness of each individual in the GA approach for BNP, we design an adaptive evaluation function based on the AWA as procedure 4.
procedure 4: Adaptive Weight Approach
input: chromosome vk , kpopSize, number of paths Lk , the flow fik and the cost cik of each path, iLk output: fitness value eval(vk), kpopSize step 1: define two extreme points: the maximum extreme point z+ and the minimum extreme point z- in criteria space as follows:
max z + = { z1max , z 2 } min z = { z1min , z 2 }

a where z1mx , z max , z1min and z2min are the maximal value and minimal value for

objective 1 and objective 2 in the current population. They are defined as follows:

z1max = max{ f i k | i Lk , k popSize} z1min = min{ f i k | i Lk , k popSize}

max z 2 = max{cik | i Lk , k popSize}

min z 2 = min{cik | i Lk , k popSize}

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

122

4.3.3 Fitness Assignment

step 2: The adaptive weight for objective 1 and objective 2 are calculated by the following equation: 1 w1 = max min z1 z1 w2 =
max z2

1 min z2
k min

step 3: Calculate the fitness value for each individual.

eval (vk ) =

( w ( f z ) w (c + z ) )
Lk k min i =1 1 i 1 2 i 2

Lk

, k popSize

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

123

4.3.4 Genetic Operators

Crossover Operator: Weight Mapping Crossover (WMX)
We proposed a new crossover operator: WMX. WMX can be viewed as an extension of one-point crossover for permutation representation.
step 1: select a cut-point
parent 1 :2 parent 2 :3 1 7 7 2 4 6 3 1 cut-point 5 3 6 5 5 4 1 6 5 4

parent 1 : 1 parent 2 : 1

3 2

4 4

7 5 7

step 2: mapping the weight of the right segment

5 5 3 1 6 4

offspring 1 : 1 offspring 2 : 1

3 2

4 4

5 7

step 3: generate offspring with mapping relationship

offspring 1 :2 offspring 2 :3 1 7 7 2 4 6 6 4 3 1 5 5

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

124

4.3.4 Genetic Operators

Mutation Operator: Insertion Mutation
Selects a gene at random and inserts it in a random position as follows:
select a gene at random parent :2 1 7 4 5

parent : 1

insert it in a random position offspring :2 5 1 7 4 3 6

offspring : 1

Selection:
the roulette wheel selection, a type of Proportionate selection, is adopted.
Soft Computing Lab. WASEDA UNIVERSITY , IPS 125

4.3.4 Genetic Operators

Immigration Operator:
Moed et. al. (1990) proposed an immigration operator which, for certain types of functions:
Allows increased exploration Maintaining nearly the same level of exploitation for the given population size.

Immigration operator procedure:

step 1: The algorithm is modified to include immigration, with each generation generated. step 2: Evaluate random members (, called the immigration rate). step 3: Replace the worst members of the population with the random members.

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

126

4.3.5 GA Procedure for BNP

GA Procedure for BNP
procedure: Priority-based GA for BNP input: network data (V, A, C, U), GA parameters output: Pareto optimal solution E(t) begin t 0; initialize P(t) by priority-based encoding; objectives z1(P), z2(P); create Pareto E(P); fitness eval(P) by adaptive weight approach; while (not termination condition) do crossover P(t) to yield C(t) by weight mapping crossover; mutation P(t) to yield C(t) by insertion mutation; immigration operation to yield C(t) ; objectives z1(C), z2(C); update Pareto E(P, C); fitness eval(P, C) by adaptive weight approach; select P(t+1) from P(t) and C(t) by roulette wheel selection; t t + 1; end output Pareto optimal solution E(t); end Soft Computing Lab. WASEDA UNIVERSITY , IPS

127

4.4 Numerical Examples

Test Problems:
The numerical examples, presented by Munakata & Hashier, was adopted.
Munakata, T. and D. J. Hashier: A genetic algorithm applied to the maximum flow problem, Proc. of 5th Int. Conf. on Genetic Algorithms, pp. 488-493 , 1993.

Using the following parameter specifications.

Population size: popSize =20 Crossover probability: pC =0.40 Mutation probability: pM =0.60 Maximum generation: maxGen =1000 Terminating condition: 100 generations with same fitness.

All the simulations were performed with Java on Pentium 4 processor (1.5-GHz clock).

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

128

4.4 Numerical Examples

Test Problem 1:
The first numerical example, presented by Munakata & Hashier, was adopted. The problem comprises 25 nodes and 49 arcs. It is given as follows:
2
10, 10 13, 8 10, 6 10, 20 15, 5 33, 4 4, 5 15, 8 35, 10

10, 15 4, 15

12

8, 20 6, 20

17
11, 8 11, 25

3
13, 20

9, 15

13

7, 20 7, 20

18
5, 8

8, 25 34, 20 10, 10 35, 25

22
10, 30

11, 15

32, 20

9 12, 15
9, 15

14

7, 20 5, 20

19

23 2, 30
3, 30

25

135, 20 631, 20

3, 10 13, 4 33, 10 7, 10 7, 10

3, 8 3, 10

10

14, 15

15

14, 20 4, 20

20 14, 25
12, 25

24

5, 15

6
Soft Computing Lab.

11

16

14, 20

21

cij , uij

j 129

WASEDA UNIVERSITY , IPS

4.4 Numerical Examples

Table 6.4 The Pareto optimal solutions of test problem 1
z1
4 5 8 10 12 15 18 20 21 23 28

z2
300 345 600 696 993 1001 1226 1568 1629 1833 2178

z1
30 33 38 40 43 47 52 56 59 66 69

z2
2470 2786 2926 3046 3274 3674 4074 4830 5406 6575 7145

z1
72 73 75 78 80 82 85 87 88 89 90

z2
7703 8382 9762 11799 13147 14531 17115 17941 19254 19333 20007

Gen, M., L. Lin & R. Cheng: Bicriteria Network Optimization Problem using Prioritybased Genetic Algorithm, IEEJ Trans. on Elect., Info. & Sys., Oct. 2004.
Soft Computing Lab. WASEDA UNIVERSITY , IPS 130

4.4 Numerical Examples

0ideal point: z1=90, z2=300

-5000 -

z1=66, z2=6575
-10000 cost -15000 -

-20000 -

- 25000 -

20

40

60

80

100

flow
Soft Computing Lab.

Fig. 6.5 The Pareto optimal solutions of test problem 1

WASEDA UNIVERSITY , IPS

131

4.4 Numerical Examples

Test Problem 2:
The first numerical example, presented by Munakata & Hashier, was adopted. The problem comprises 25 nodes and 56 arcs. It is given as follows:
2
3, 10 14, 8 8, 10 2, 20 3, 5 7, 18 12, 7 4, 6 11, 5

11

29, 11 13, 2 6, 8

20

16

26, 9

3
10, 20

13, 6 32, 8 7, 5 6, 12 12, 8 9, 7

12

30, 10 6, 8 9, 10 12, 7

21
10, 20 31, 15 8, 4

2,15

17

33, 20

12, 9 33, 8 4, 7 12, 10 30, 7 11, 12 8, 20 12, 6 3, 5

13

34, 10 15, 10 2, 9

22

6, 30

25

126, 20

18

7, 20 30, 8 11, 6

5
634, 20

14

35, 6 14, 5 7, 8

23
9, 15

10

11, 22 6, 6

19

35, 9

Soft Computing Lab.

12, 15

WASEDA UNIVERSITY , IPS

15

8, 10

24

cij , uij

j 132

4.4 Numerical Examples

Table 6.5 The Pareto optimal solutions of test problem 2
z1
2 8 10 15 18 19 20 21 25 26 27 28

z2
52 248 340 495 692 1012 1111 1220 1292 1406 1457 1475

z1
32 34 36 38 40 41 43 47 49 51 52

z2
1633 1909 1937 2077 2485 2581 2731 3080 3302 3551 3739

z1
54 55 58 61 63 65 66 67 68 71 72

z2
3872 3990 4146 4671 5153 5463 5704 6323 6422 6537 6748

z1
73 74 75 76 78 80 82 83 85 86 91

z2
6944 7192 7402 7532 7847 9228 10395 12508 12610 13151 16752

Gen, M., L. Lin & R. Cheng: Bicriteria Network Optimization Problem using Prioritybased Genetic Algorithm, IEEJ Trans. on Elect., Info. & Sys., Oct. 2004.
Soft Computing Lab. WASEDA UNIVERSITY , IPS 133

4.4 Numerical Examples

0 -2000 -4000 -6000 -8000 cost -10000 -12000 -14000 -16000 -18000 0
Pa too timl s lu n re p a o tio id a p in el o t

ideal point:

z1=91, z2=52

z1=61, z2=4671

20

40

60

80

100

flow
Soft Computing Lab.

Fig. 6.6 The Pareto optimal solutions of test problem 2

WASEDA UNIVERSITY , IPS

134

4.4 Numerical Examples

Simulation (# of nodes: 25, # of arcs: 56)

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

135

6. Network Design Problems

1. Shortest Path Problem (SPP) 2. Maximum Flow (MXF) Problem 3. Minimum Cost Flow (MCF) Problem 4. Bicriteria Network Design Problem (BNP) 5. Multi-criteria Network Design Problem
5.1 Introduction of Multi-criteria Network Design Problem 5.2 Reviewing Solution Approaches for MNP 5.3 Numerical Examples

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

136

5. Multi-criteria Network Design Problem (MNP)

With the information superhighway fast becoming a reality, the problem of designing networks capable of accommodating multimedia (both audio and video) traffic in a multicast (simultaneous transmission of data to multiple destinations) environment has come to assume paramount importance

Chow, C.-H.: On multicast path finding algorithms, Proceedings of IEEE

INFOCOM, pp.1274-1283, 1991. Frank, A., L. Wittie, and A. Bernstein: Multicast communication in network computers, IEEE Software, Vol. 2, No. 3, pp. 49-61,1985. Kadaba, B. and J. Jaffe: Routing to multiple destinations in computer networks, IEEE Transactions on Communications, Vol. COM-31, pp. 343351,1983. Kompella, V.P., J.C. Pasquale and G.C. Polyzos: Multicasting for multimedia applications, Proceedings of IEEE INFOCOM, 1992. Kompella, V.P., J.C. Pasquale and G.C. Polyzos: Multicast routing for multimedia communication, IEEE/ACM Transactions on Networking, pp. 286-292, 1993.
WASEDA UNIVERSITY , IPS 137

Soft Computing Lab.

5.1 Introduction of MNP

Network design problems where even one cost measure must be minimized, are often NP-hard. But, in real-life applications, it is often the case that the network to be built is required to minimize multiple cost measures simultaneously, with different cost functions for each measure. For example, in the problem of finding good multicast trees, each edge has associated with it two edge costs:
The construction cost: It is typically a measure of the amount of buffer space or channel bandwidth used The delay cost: It is a combination of the propagation, transmission and queuing delays.

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

138

5.1 Introduction of MNP

Multi-criteria network design problems, with separate cost functions for each optimization criterion, also occur naturally in Information Retrieval and VLSI designs.

Bookstein, A. & S.T. Klein: Construction of Optimal Graphs for Bit-Vector Compression, Proc. 13th ACM-SIGIR, vol. 16, pp. 387-400, 1990. Zhu, Q., M. Parsa & W.W.M. Dai: An iterative approach for delay-bounded minimum Steiner tree construction, Technical Report UCSC-CRL-94-39, UC Santa Cruz, 1994.

With the advent of deep micron VLSI designs, the feature size has shrunk to sizes of 0.5 microns and less. As a result, the interconnect resistance, being proportional to the square of the scaling factor, has increased significantly. An increase in interconnect resistance has led to an increase in interconnect delays thus making them a dominant factor in the timing analysis of VLSI circuits. Therefore VLSI circuit designers aim at finding minimum cost (spanning or Steiner) trees given delay bound constraints on sourcesink connections.
Soft Computing Lab. WASEDA UNIVERSITY , IPS 139

5.1 Introduction of MNP

For example, the problem of finding low-cost and low-transmission-delay multimedia networks can be modeled as the (Diameter, Total cost, Spanning tree)-bicriteria problem:
given an undirected graph G = (V,E) with two weight functions ce and de for each edge eE modeling construction and delay costs respectively, and a bound D (on the total delay), find a minimum c-cost spanning tree such that the diameter of the tree under the d-costs is at most D. It is easy to see that the notion of bicriteria optimization problems can be easily extended to the more general multicriteria optimization problems.

The applications set the stage for the formal definition of multicriteria network design problems. Marathe et al. explain this concept by giving a formal definition of a bicriteria network design problem.
Marathe, M. V., R. Ravi, R. Sundaram, S. S. Ravi, D. J. Rosenkrantz, and H. B. Hunt: Bicriteria network design problems, J. Algorithms, vol. 28, no. 1, pp. 142-171, 1998. Marathe et al. study the complexity and approximability of a number of bicriteria network design problems. The three objectives considered:
total cost diameter degree of the network.
WASEDA UNIVERSITY , IPS 140

Soft Computing Lab.

5.2 Reviewing Solution Approaches for MNP

a. AWA (Gen et al., 1998)
Gen, M. & R. Cheng: Genetic Algorithms and Engineering Optimization, John Wiley & Sons, New York, 2000.

b. RWA (Murata et al., 1998)

Gen, M. & R. Cheng: Genetic Algorithms and Engineering Optimization, John Wiley & Sons, New York, 2000.

c. SPEA (Zitzler et al., 1999)

Zitzler, E. & L. Thiele: Multiobjective Evolutionary Algorithms: A
Comparative Case Study and the Strength Pareto Approach, IEEE Trans. on Evol. Comput., Vol.3, No.4, pp.257-271, 1999.

d. NSGA-(Deb et al., 2000)

Deb, K., A. Pratap, S. Agarwal and T. Meyarivan: A Fast and Elitist Multiobjective Genetic Algorithm: NSGA-, IEEE Trans. on Evol. Comput., Vol.6, No.2, 182-197, 2002.
Soft Computing Lab. WASEDA UNIVERSITY , IPS 141

5.2 Reviewing Solution Approaches for MNP

a. AWA (Gen & Cheng, 1998)
Gen & Cheng (1998) proposed an Adaptive Weight Approach (AWA) which utilizes some useful information from the current population to readjust weights to obtain a search pressure toward a positive ideal point.
Gen, M. & R. Cheng: Genetic Algorithms and Engineering Optimization, John Wiley & Sons, New York, 2000.

For the examined solutions at each generation, they define two extreme points (maximum: z+, minimum: z-)
max max z + = [ z1max z2 z q ] min min z = [ z1min z2 z q ]

where zkmax and zkmin are the maximal and minimal values for the kth objective as defined by the following equations:

zkmax = max{ f k ( x ) | x P}, zkmin = min{ f k ( x ) | x P},

k = 1, 2, , q

k = 1, 2, , q
P: set of solution candidates.

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

142

5.2 Reviewing Solution Approaches for MNP

a. AWA (Gen & Cheng, 1998)
The weighted-sum objective function for a given chromosome x is given by the following equation:

z( x) =

w (z
k =1 k

min k

zk zkmin ) = max min = zk k =1 z k

q

k =1

f k ( x ) z kmin zkm ax zkm in

where wk is adaptive weight for objective k :

wk =

1 , max min zk zk

k = 1, 2, , q

The equation driven above is a hyperplane defined by the following extreme points in current solutions:

[ z1max [ z1min [ z1min [ z1min

Soft Computing Lab.

min min z 2 z kmin z q ] max min z 2 z kmin z q ] min min z 2 z kmax z q ] min max z 2 z kmin z q ]

WASEDA UNIVERSITY , IPS

143

5.2 Reviewing Solution Approaches for MNP

a. AWA (Gen & Cheng, 1998)
Adaptive moving line defined by the extreme points (z1max, z2min) and (z1max, z2min) are shown as follows:

z2
z
max 2

minimal rectangle containing all current solutions

positive ideal point whole criteria space Z

(z

min 1

,z

max 2

z+

maximum
extreme point

min z2

minimum

min ( z1max , z 2 )

adaptive
moving line

extreme point corresponding to

subspace

current solutions

z1min
Soft Computing Lab.

z1max
WASEDA UNIVERSITY , IPS

z1
144

Fig.6.7 Adaptive weights and adaptive hyperplane

5.2 Reviewing Solution Approaches for MNP

b. RWA (Murata et al., 1998)
Murata, Ishibuchi & Tanaka (1998) proposed a Random-Weight Approach (RWA) to obtaining a variable search direction toward the Pareto frontier.
Gen, M. & R. Cheng: Genetic Algorithms and Engineering Optimization, John Wiley & Sons, New York, 2000. Ishibuchi, H., T. Yoshida and T. Murata: Balance Between Genetic Search and Local Search in Memetic Algorithms for Multiobjective Permutation Flowshop Scheduling, IEEE Trans. on Evol. Comput., Vol.7, No.2, pp.204-223, 2003.

Fixed-weight approach gives the GAs a tendency to sample the area toward a fixed point in the criterion space. Random-weight approach gives the GAs a tendency to demonstrate a variable search direction, therefore, the ability to sample the area uniformly over the entire frontier. f1 f1

fixed search direction f2

Soft Computing Lab.

multiple search direction f2

145

WASEDA on a fixed direction Fig. 6.8 SearchUNIVERSITY , IPS in criterion space

5.2 Reviewing Solution Approaches for MNP

b. RWA (Murata et al., 1998)
For a problem to maximize q objective functions, weighted-sum objective is given as the follows:

z = wk f k ( x )
k =1

Random-weight wk is calculated by the following equation:

wk =

rk
q

j =1 j

k = 1, 2, , q

where rj are non-negative random number between [0, 1]. Before selecting a pair of parents for crossover operation, a new set of random weights is specified. The selection probability pi for individual i is then defined by the following linear scaling function:

pi =

zi z min

popSize j =1

(z

z min )

where zmin is the worst fitness value in the current population.

Soft Computing Lab. WASEDA UNIVERSITY , IPS 146

5.2 Reviewing Solution Approaches for MNP

c. SPEA (Zitzler et al., 1999) Zitzler & Thiele (1999) proposed a new evolutionary approach to multicriteria optimization, the Strength Pareto Evolutionary Algorithm (SPEA), that combines several features of previous multiobjective EAs in a unique manner.
Zitzler, E. & L. Thiele: Multiobjective Evolutionary Algorithms: A Comparative Case Study and the Strength Pareto Approach, IEEE Trans. on Evol. Comput., Vol.3, No.4, pp.257-271, 1999.

It is characterized by:
Storing nondominated solutions externally in a second, continuously updated population. Evaluating an individuals fitness dependent on the number of external nondominated points that dominate it. Preserving population diversity using the Pareto dominance relationship. Incorporating a clustering procedure in order to reduce the nondominated set without destroying its characteristics.

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

147

5.2 Reviewing Solution Approaches for MNP

c. SPEA (Zitzler et al., 1999)
The fitness assignment procedure is a two-stage process.
First, the individuals in the external nondominated set P are ranked.

n si = N +1

where si is proportional to the number of population members jP for which i j. n is the number of individuals in P that are covered by i and N is the size of P. f1
3/8 5/8

3/8

f2 Fig. 6.9 Two scenarios for a maximization problem with two objectives.
Soft Computing Lab. WASEDA UNIVERSITY , IPS 148

5.2 Reviewing Solution Approaches for MNP

c. SPEA (Zitzler et al., 1999)
The fitness assignment procedure is a two-stage process.
Afterwards, the individuals in the population P are evaluated.

f j = 1+

i, i j

s ,
i

where f j [1, N )

where the fitness of an individual jP is calculated by summing the strengths of all external nondominated solutions iP that cover j. f1
11/8 16/8 13/8 3/8 5/8 13/8

3/8 11/8

19/8 16/8

f2 Fig. 6.10 Two scenarios for a maximization problem with two objectives.
Soft Computing Lab. WASEDA UNIVERSITY , IPS 149

5.2 Reviewing Solution Approaches for MNP

d. NSGA-(Deb et al., 2000)
Deb, Pratap, Agarwal & Meyarivan (2000) suggest a nondominated sorting-based Multiobjective Evolutionary Algorithm (MOEA), called Nondominated Sorting Genetic Algorithm(NSGA-), which alleviates the three difficulties:
Deb, K., A. Pratap, S. Agarwal and T. Meyarivan: A Fast and Elitist Multiobjective Genetic Algorithm: NSGA-, IEEE Trans. on Evol. Comput., Vol.6, No.2, 182-197, 2002. Computational complexity Nonelitism approach The need for specifying a sharing parameter

They presented the proposed NSGA- approach:

Fast nondominated sort They first sorting a population into different nondomination levels. Nondomination rank (Ri); Crowding distance assignment The crowding-distance computation requires sorting the population according to each objective function value in ascending order of magnitude. Crowding distance (Di)
Soft Computing Lab. WASEDA UNIVERSITY , IPS 150

5.2 Reviewing Solution Approaches for MNP

d. NSGA-(Deb et al., 2000)
The new population Pt+1 is now used for selection, crossover, and mutation to create a new population Qt+1. It is important to note that they used a binary tournament selection operator, but the selection criterion is based on the crowded-comparison operator n . Crowded-comparison operator is defined as follows:

if ( Ri < R j ) or ( ( Ri = R j ) and ( Di > D j ) ) then in j

where, Ri is nondomination rank and Di is crowding distance

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

151

5.3 Numerical Examples

Test Problems:
For examining the effect of different solution approaches of multiobjective GAs, we applied SPEA, NSGA, RWA and AWA to the 2 test problems.
Munakata, T. & D. J. Hashier: A genetic algorithm applied to the maximum flow problem. Proc. of the 5th Inter. Conf. on Genetic Algorithms, San Francisco, pp.488-493, 1993.

Using the following parameter specifications.

Population size: Crossover probability: Mutation probability: Immigration rate: Stopping conditions: popSize =20 pC =0.70 pM =0.50 =3 Evaluation of 5000 solutions.

Each algorithm was applied to each test problem 20 times (i.e., 20 runs) using different initial populations. All the simulations were performed with Java on Pentium 4 processor (1.5-GHz clock).
Soft Computing Lab. WASEDA UNIVERSITY , IPS 152

5.3 Numerical Examples

Test Problems:
The first numerical example, presented by Munakata & Hashier, was adopted. The problem comprises 25 nodes and 49 arcs. It is given as follows:
2
10, 10 13, 8 10, 6 10, 20 15, 5 33, 4 4, 5 15, 8 35, 10

10, 15 4, 15

12

8, 20 6, 20

17
11, 8 11, 25

3
13, 20

9, 15

13

7, 20 7, 20

18
5, 8

8, 25 34, 20 10, 10 35, 25

22
10, 30

11, 15

32, 20

9 12, 15
9, 15

14

7, 20 5, 20

19

23 2, 30
3, 30

25

135, 20 631, 20

3, 10 13, 4 33, 10 7, 10 7, 10

3, 8 3, 10

10

14, 15

15

14, 20 4, 20

20 14, 25
12, 25

24

5, 15

6
Soft Computing Lab.

11

16

14, 20

21

cij , uij

j 153

WASEDA UNIVERSITY , IPS

5.3 Numerical Examples

Test Problems:
The second numerical example, presented by T. Munakata & D.J. Hashier, was adopted. The problem comprises 25 nodes and 56 arcs. It is given as follows: 2
3, 10 14, 8 8, 10 2, 20 3, 5 7, 18 12, 7 4, 6 11, 5

11

29, 11 13, 2 6, 8

20

16

26, 9

3
10, 20

13, 6 32, 8 7, 5 6, 12 12, 8 9, 7

12

30, 10 6, 8 9, 10 12, 7

21
10, 20 31, 15 8, 4

2,15

17

33, 20

12, 9 33, 8 4, 7 12, 10 30, 7 11, 12 8, 20 12, 6 3, 5

13

34, 10 15, 10 2, 9

22

6, 30

25

126, 20

18

7, 20 30, 8 11, 6

5
634, 20

14

35, 6 14, 5 7, 8

23
9, 15

10

11, 22 6, 6

19

35, 9

6
Soft Computing Lab.

12, 15

15

8, 10

24

cij , uij

WASEDA UNIVERSITY , IPS

154

5.3 Numerical Examples

Reference solution set S* :
The reference solution set S* of each test problem was found using the SPEA, NSGA-2, RWA, and AWA. Each algorithm was applied to each test problem with much longer computation time and larger memory storage than the other computational experiments in this study. More specifically, we used the following parameter specifications in all the three algorithms for finding the reference solution set of each test problem.
Population size: popSize =30 Crossover probability: pC =0.70 Mutation probability: Immigration probability: Stopping conditions: pM =0.80 =5 Evaluation of 100000 solutions.

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

155

5.3 Numerical Examples

Reference solution set S* to test problems:
0

5000

cost

10000

15000

20000 0 20 40 60 80 100

flow

Fig. 6.11 The Reference solution set of Example 1 (|S*|=69)

Soft Computing Lab. WASEDA UNIVERSITY , IPS 156

5.3 Numerical Examples

Reference solution set S* to test problems:
Table 6.6 The Reference solution set of Example 1 (|S*|=69)
z1
5 6 8 10 11 12 13 16 18 20 23 25 28 30 32 33 34 35 37 38

z2
260 318 376 510 590 644 744 808 918 1030 1206 1320 1496 1650 1796 1875 1955 2058 2162 2262

z1
39 40 41 42 43 45 46 47 48 49 50 51 52 53 54 55 56 57 58 60

z2
2380 2447 2531 2648 2696 2866 3042 3049 3151 3322 3393 3527 3530 3932 4066 4072 4364 4479 4747 5010

z1
62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81

z2

z1
82 83 84 85 86 87 88 89 90

z2
14088 14841 15561 16198 16842 17528 18332 18988 19597

5339 5743 5936 5962 6162 6382 6714 6846 6952 7182 7364 8043 8707 9375 10057 10750 11438 12123 12735 13523

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

157

5.3 Numerical Examples

Reference solution set S* to test problems:
0

5000

cost

10000

15000

20000 0 20 40

flow

60

80

100

Fig.6.12 The Reference solution set of Example 2 (|S*|=77)

Soft Computing Lab. WASEDA UNIVERSITY , IPS 158

5.3 Numerical Examples

Reference solution set S* to test problems:
Table 6.7 The Reference solution set of Example 2 (|S*|=77)
z1
2 5 7 8 10 11 13 15 17 18 20 22 23 24 25 26 27

z2
52 115 167 248 275 306 388 495 608 653 784 872 953 965 1035 1168 1217 1275 1332

z1
31 32 33 34 35 36 37 38 40 42 43 44 45 47 48 49 50

z2
1408 1488 1595 1640 1753 1818 1892 1942 2171 2319 2405 2587 2653 2701 2897 2974 3077

z1
55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 73

z2
3525 3610 3669 3732 3987 4189 4458 4551 4751 4867 5069 5341 5517 5583 5809 5941 6128

z1
75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91

z2
6815 7081 7227 7308 8002 8697 9319 9997 10667 11341 11949 12623 13455 14116 14735 15551 16100

Soft Computing Lab.

28 29

51 3106 WASEDA 52 3198

72 6319 UNIVERSITY , IPS 6487

159

5.3 Numerical Examples

Performance Measures:
We mainly use a performance measure based on:
The number of obtained solutions |Sj| The ratio of nondominated solutions RNDS(Sj)
The RNDS(Sj) measure can be written as follows:
RNDS ( S j ) = S j { x S j r S * : r x} Sj

The distance D1R

The D1R measure can be written as follows:

D1R =

1 S*

rS *

min{d

rx

x S j}

where S* is a reference solution set for evaluation the solution set Sj. dxr is the distance between a solution x and2 a reference solution r. 2 d rx = f1 ( r ) f1 ( x) + f 2 (r ) f 2 ( x)

) (

[Ref.] Ishibuchi, H., T. Yoshida & T. Murata: Balance Between Genetic Search and Local Search in Memetic Algorithms for Multiobjective Permutation Flowshop Scheduling. IEEE Trans. On Evol. Comp., Vol. 7, No. 2, pp. 204-223, 2003. Soft Computing Lab. 160 WASEDA UNIVERSITY , IPS

5.3 Numerical Examples

Table 6.8 Comparison with the four approaches using the |Sj| measure.
Test Problems (# of nodes/ # of arcs) 25/49 25/56 RWA 52 43 |Sj| SPEA NSGA- 57 44 43 55 AWA 49 43 RWA CPU Times SPEA NSGA- 15693 15981 AWA 14170 14961

15122 17635 11918 16684

Table 6.9 Comparison with the four approaches using the RNDS (Sj) measure.
Test Problems (# of nodes/ # of arcs) 25/49 25/56 RWA 0.57 0.41 RNDS (Sj) SPEA NSGA- 0.54 0.34 0.39 0.36 AWA 0.61 0.53 RWA CPU Times SPEA NSGA- 15693 15981 AWA 14170 14961

15122 17635 11918 16684

Table 6.10 Comparison with the four approaches using the D1R measure.
Test Problems (# of nodes/ # of arcs) 25/49 25/56 RWA 191.21 203.96 D1R measure SPEA NSGA- 315.61 224.40 228.65 185.89 AWA 143.58 141.43 RWA CPU Times SPEA NSGA- 15693 15981 AWA 14170 14961

15122 17635 11918 16684

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

161

5.3 Numerical Examples

Different Parameter Settings:
Comparison with different approaches using the stopping conditions: under the same computation time: 10,000 ms. Table 6.11 Comparison with the four approaches using the |Sj| measure.
Test Problems (# of nodes/ # of arcs) 25/49 25/56 |Sj| AWA 49 52 RWA 50 50 SPEA 53 40 NSGA- 50 34

Table 6.12 Comparison with the four approaches using the RNDS (Sj) measure.
Test Problems (# of nodes/ # of arcs) 25/49 25/56 AWA 0.57 0.51 RNDS (Sj) RWA SPEA 0.44 0.32 0.56 0.60 NSGA- 0.48 0.41

Table 6.13 Comparison with the four approaches using the D1R measure.
Test Problems (# of nodes/ # of arcs) 25/49 25/56 AWA 191.17 147.07 D1R measure RWA SPEA 203.72 219.59 222.28 279.60 NSGA- 239.99 433.43

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

162

5.3 Numerical Examples

Different Parameter Settings:
Comparison with different approaches using the stopping conditions: under the same computation time: 10,000 ms. Table 6.11 Comparison with the four approaches using the |Sj| measure.
Test Problems (# of nodes/ # of arcs) 25/49 25/56 RWA 50 50 |Sj| SPEA NSGA- 53 40 50 34 AWA 49 52

Table 6.12 Comparison with the four approaches using the RNDS (Sj) measure.
Test Problems (# of nodes/ # of arcs) 25/49 25/56 RWA 0.44 0.32 RNDS (Sj) SPEA NSGA- 0.56 0.60 0.48 0.41 AWA 0.57 0.51

Table 6.13 Comparison with the four approaches using the D1R measure.
Test Problems (# of nodes/ # of arcs) 25/49 25/56 RWA 203.72 219.59 D1R measure SPEA NSGA- 222.28 279.60 239.99 433.43 AWA 191.17 147.07

Soft Computing Lab.

WASEDA UNIVERSITY , IPS

163

5.3 Numerical Examples

Discussion of the Results:
1 D1R 1 D1R

Sj

R NDS ( S j )

Sj

R NDS ( S j )

Test problem 1 (25/49)

Test problem 1 (25/56)

D1R: distance RNDS (Sj): ratio of nondominated solutions |Sj|: No. of obtained solutions Soft Computing Lab. WASEDA UNIVERSITY , IPS 164

Conclusion
In this study, we presented a GA approach used a priority-based chromosome for solving the network design problems. It is easy to verify that any permutation of the encoding corresponds to the paths. So that most existing genetic operators can easily be applied to the encoding. Also, any path has a corresponding encoding. Therefore, any point in solution space is accessible for genetic search. For solving the MXF/MCF, and Multi-criteria Network Design Problem, we also combines an adaptive evaluation function based on the AWA. The fitness values of all individuals are calculated according to this adaptive evaluation function. In each generation, the set of Pareto solutions is updated by deleting all dominated solutions and adding all newly generated Pareto solutions. Computer simulations show the several numerical experiments by using several network optimization problems, and show the effectiveness of the proposed method.
WASEDA UNIVERSITY , IPS 165

Soft Computing Lab.