Proceedings of the 2013 Industrial and Systems Engineering Research Conference

A. Krishnamurthy and W.K.V. Chan, eds.

A Sweeping Cones Technique for Post Pareto Analysis

Victor M. Carrilloa, Heidi Taboada PhDb,*
a

Computational Science Program. The University of Texas at El Paso, 500 West University, El
Paso, TX 79968-0521, United States of America.
Department of Mathematics and Physics Sciences .Universidad Autnoma de Ciudad Jurez,Av
P.Elias Calles 1210,(52) 656-6884800, Juarez Chih ,Mxico.
b
Department of Industrial, Manufacturing and Systems Engineering. The University of Texas at
El Paso, 500 West University, El Paso, TX 79968-0521, United States of America.
Abstract
There exist two general approaches to solve multiple objective problems. The first approach belongs to the classical
mathematical methods. The weighted sum method, goal programming, or utility functions are methods that pertain
to this approach. The output of this method is a single optimal solution. On the other hand, we have the multiple
objective evolutionary algorithms that offer the decision maker a set of optimal solutions usually called non
dominated or, Pareto-optimal solutions. This set is usually very large and the decision maker faces the problem of
reducing the size of this set to a manageable number of solutions to analyze. This paper presents a post-Pareto
approach to prune the non-dominated set of solutions obtained by multiple objective evolutionary algorithms. The
proposed approach uses a non-uniform weight generator method with a sweeping cones technique to reduce the size
of the Pareto-front. This pruning system is not affected by the non convexity of the Pareto front. An example is
presented to show the performance of the method.

Keywords:
Post-Pareto optimality; decision making; multi-objective optimization problems.

1. Introduction
Most optimization problems in the real world are multi-objective and its functions are in conflict with each other.
Multi-objective optimization problems appear in economics, biology, engineering, military, etc. This is very easy to
observe in real life. For example, a high performance electronic product is also a high cost item, and every customer
is always looking for a device with high performance, but at the lowest rates. Therefore, development of a method
for finding a solution to several problems optimization goals has been the subject of numerous research and guides
documents [1]. Currently, two main approaches are known to solve multi-objective optimization problems: the first
is known as classic mathematical methods. Some examples of classical mathematical methods are: the weighted
sum, weighted metric, rotated weighted metric, constraint method, goal programming, the lexicographic method,
utility functions ,the weighted sum, weighted metric, rotated weighted metric, constraint method, goal programming,
the lexicographic method, utility functions among others . Meanwhile the second approach involves populating a
number of feasible solutions along the Pareto frontier, and the final solution is a set of Pareto-optimal solutions or
the set of all optimal solutions of the multi-objective optimization problem. Some evolutionary methods are: Tabusearch, neural networks, particle swarm, ant colony and genetic algorithms, among others. At this stage the
optimization procedure terminates, but if the Pareto optimal size is very large the decision maker won't be able to
select comfortably a manageable subset of the optimal solutions. The problem of obtaining a small subset of the
Pareto-optimal front led to the development of methods for reducing the size of the Pareto fronts. Most of
evolutionary methods obtain as a solution a set of Pareto optimal solutions based in the Pareto dominance concept.
The idea of the Pareto-Front is to compare all solutions against each other, where the best fitted solutions dominate
the weaker which in turn are said to be dominated [2]. The set of non-dominated solutions is known as the ParetoFront.

Proceedings of the 2013 Industrial and Systems Engineering Research Conference

A. Krishnamurthy and W.K.V. Chan, eds.

Definition: A solution x1 dominates a solution x2 if and only if the two following conditions are true:

For all objectives i.e. ( )

And for at least one
:

( )
( )

( )

The solution of a multiple objective optimization is a set of non-dominated solutions or Pareto-front. Sometimes,
the Pareto front can contain a large number (in some cases, thousands) of solutions. From the decision-makers
perspective , choosing some good optimal points from all the non-dominated solutions can be prohibitive and
inefficient. The selection of solutions from the Pareto set is called post-Pareto optimality analysis, in the present
work a non-uniform weight method[17] is used in the non-numerical ranking preferences method[3] with the
purpose to prune and obtain a smaller sub-set of preferred solutions from the large Pareto-optimal set. In section 2,
the literature review about post-Pareto optimality analysis is presented, in section 3 the proposed sweeping cones
method is developed, and an example applying the sweeping cones method is presented followed by the conclusions
in section 4.
In this analysis we scalarized the Pareto-optimal points obtained with the evolutionary algorithm NSGA[3] and
compute the function
,which afterwards is bounded using numerical thresholds to find
the minimum value of such aggregated values. The weight values were previously obtained utilizing the nonuniform weights algorithm according to the following non-numerical ranking f1 f 2
f n i.e. ,

w1 w2

wn .

2. Literature review
Some methods developed to reduce the number of solutions in the Pareto-set in the literature are: Korhonen &
Halme (1990) [4] whom suggested a value efficiency approach to help the decision-maker identify the most
preferred solution in multi-objective optimization problems. Venkat, et al. (2004) [5] addressed the problem
introducing a Greedy Reduction (GR) approach. Their method works by obtaining a sub-set of Pareto optima from a
large set and is based on maximizing a scalarized function of the vector of percentile ordinal rankings of the Pareto
optima within the large Pareto set. Padhye, et al.(2009) [6] proposed a mutation driven hill climbing local search
using
achievement
scalarizing
function
to
refine
the
solutions
from
the
Pareto
set.
Carrillo,Aguirre,Taboada(2011)[7]proposed a method based on preferences rankings provided by the decision maker
to bound an scalarized function made out of Pareto-optimal values for Pareto-optimal set reduction purposes.
Carrillo,Taboada(2012)[8] developed an algorithm that generates n non-uniformly distributed weights for the
method based on preferences rankings for any MOOP problem size. Carrillo,Taboada(2012)[9] presented the
generalization of a post-Pareto optimality method known as the non-numerical ranking preferences (NNRP) method
originally proposed by Taboada & Coit [10].
Other known pruning techniques that do not rely precisely on scalarizing are for instance: the k-means clustering
methods which groups solutions with similar characteristics, for example , Taboada et al [11] ; Another similar
procedure was presented by Runkler[12] using a fuzzy c-means clustering using the k-means technique; Another
approach based in clustering was proposed by Li,Liao,Coit[13 ] where a data envelopment analysis is performed, by
comparing the relative efficiency of the solutions to determine the final representative solutions for the overall
problem. Aguirre,Taboada [14] developed an algorithm that offers two main advantages: there is no need to provide
an initial number of clusters, and at each hierarchical level, the algorithm optimizes the number of clusters, and can
reassign data from previous hierarchical levels in order to rearrange missed clustered data. The clustering method is
the dynamically growing self organizing tree (DGSOT) algorithm.
The method proposed in this work is based on a non-numerical ranking of the objective functions based on their
relative importance . The main strength of this method is precisely that the decision-maker only ranks nonnumerically (in order of relative importance) the objective functions but does not have to select specific weight
values ; the sweeping cone method does not depend on the non-convexity of the Pareto front and that can be applied
to any multi-objective optimization problem size . Unlike other methods ,Carrillo,Taboada (2011-2012)[7], [ 8], the
pruning method uses lower boundaries to help the decision-maker select in a flexible way solutions that reflect

Proceedings of the 2013 Industrial and Systems Engineering Research Conference

A. Krishnamurthy and W.K.V. Chan, eds.

his/her preferences. In next section are explained all the details of the proposed algorithm and how the pruning of
solution is performed.

3. Sweeping Cones Post-Pareto Method

This method requires normalizing the vectors
(
) of the Pareto front and the weights obtained to be
used in the non-numerical ranking preferences method calculating
and
Remark: These transformations sends all of the

and

vectors over the unitary sphere surface.

Sweeping cones method code

a) Select a lower boundary threshold of = 0.742, which is approximately equal to
"An initial

angle is chosen for the first sweeping cone".

b) Randomly generate a collection of weights

"Vector

) is the axis of the cone"

c) Compute the linear combination

d) Choose the

that satisfy

e) Increase gradually the threshold value (reduce the cone angle) and repeat steps a, b, c and
d, until there is no

that satisfy

Summary: Start with a 42 degrees cone around each one of the weight vectors is used to seize those of the
(
) optimal values on the unitary sphere with a sweeping search as with a lantern and gradually
diminishing the cone opening to separate clustered points and reducing the size of the Pareto front. To reduce the
cone opening the threshold value must be increased to be hold in the relation
until no points are
trapped by the cones as shown in Figures 1, 2, 3, 4 to end the process.

Figure 1: Sweeping cones

Figure 2: Sweeping cones

Proceedings of the 2013 Industrial and Systems Engineering Research Conference

A. Krishnamurthy and W.K.V. Chan, eds.

Figure 3: Sweeping cones

Figure 4: Sweeping cones

Example. The performance of the method was tested over a Pareto set obtained in the work presented by Taboada
and Coit (2007) [10]. The problem solved in such work is known as a multi-objective redundancy allocation
problem (MORAP). The RAP refers to a system of s subsystems in series. For each subsystem, there are mi
functionally equivalent components, with different levels of cost, weight, reliability and other characteristics, which
may be selected. There is an unlimited supply of each of the mi choices. The objective of the problem is to find how
many components to set in parallel in each subsystem and of which supplier in order to optimize three different
objectives in this specific case the objectives considered were: reliability cost and weight. Table 1 shows the Paretoset of solutions obtained in [10]. The Pareto sets consists of 75 solutions.
Table 1 Non dominated Solutions
Solution
Reliability
Cost
Weight
Solution
Reliability
Cost
Weight
Solution
Reliability
Cost
Weight
Solution
Reliability
Cost
Weight
Solution
Reliability
Cost
Weight

1
0.6820
13
13
16
0.9016
35
22
31
0.9730
42
34
46
0.9911
58
41
61
0.9996
81
77

2
0.7204
16
16
17
0.9156
33
28
32
0.9734
43
32
47
0.9921
59
37
62
0.9996
64
85

3
0.7571
19
19
18
0.9172
34
27
33
0.9746
42
41
48
0.9923
57
50
63
0.9997
70
89

4
0.7886
17
17
19
0.9196
36
28
34
0.9797
44
38
49
0.9943
58
45
64
0.9997
74
91

5
0.8432
19
19
20
0.9262
37
26
35
0.9808
43
47
50
0.9949
60
49
65
0.9998
68
92

6
0.8596
21
21
21
0.9342
35
32
36
0.9822
52
35
51
0.9980
68
45
66
0.9998
76
94

7
0.8604
21
21
22
0.9360
36
31
37
0.9824
45
45
52
0.9981
69
43
67
0.9999
85
81

8
0.8645
23
23
23
0.9426
36
32
38
0.9826
53
33
53
0.9981
73
55
68
0.9999
87
85

9
10
11
12
13
14
15
0.8753
0.8771
0.8821
0.8833
0.8884
0.8932
0.9012
31
23
25
27
29
33
34
31
23
25
27
29
33
34
24
25
26
27
28
29
30
0.9447
0.9451
0.9619
0.9623
0.9636
0.9684
0.9702
38
39
38
39
39
40
41
29
30
36
34
35
40
39
39
40
41
42
43
44
45
0.9828
0.9835
0.9844
0.9851
0.9860
0.9864
0.9905
54
46
50
48
54
55
56
39
46
42
44
43
41
37
54
55
56
57
58
59
60
0.9984
0.9988
0.9990
0.9990
0.9991
0.9991
0.9994
75
72
77
79
80
82
62
47
67
51
69
69
73
82
69
70
71
72
73
74
75
0.9999 0.999962 0.999977 0.999982 0.999994 0.999998 0.999999
88
90
94
119
121
125
143
85
89
105
110
113
121
120

After having applied the sweep cones proposal to the 75 Pareto front in the redundancy allocation optimization
problem, the preliminary results in Table 21 were obtained after a pruning process for several alpha threshold
( )
values which is equal to the cosine of the theta angle shown in Table 2: i. e.

Proceedings of the 2013 Industrial and Systems Engineering Research Conference

A. Krishnamurthy and W.K.V. Chan, eds.

Table 2: Pruned values with the sweeping cones approach

1,000,000 Iterations for each angle
Running time (secs)
Angle

34.7822 34.1318 34.2003 31.0877

33.627

33.4596 33.2643 33.2854 33.5499

42.0268 42.0979

41.582

41.409

41.926

41.755

41.236

a<G
Count

0.74<F

1.33%

2.66%

14

14

14

14

16

16

16

4%

15

15

15

15

38

38

5.33%

16

16

16

16

47

18

19

20

20

19

20

36

36

20

24

38

38

24

25

45

45

25

32

47

47

10

32

36

51

52

11

36

38

52

54

12

38

39

54

13

39

45

56

14

44

46

15

45

47

16

46

51

17

47

52

18

51

54

19

52

56

20

54

21

56

Threshold

14.66%
17.33%

22.66%
25.33%
28%

.
.
48
49
50
.
.
73

41.062

Pareto front subsets for 8 threshold values

40.9746

0.742<F 0.744<F 0.746<F 0.748<F 0.750<F 0.752<F 0.754<F 0.755<F

9

60
50
40
Series1
30

Series2

20

Series3
Series4

10

Series5
0
Series7

Series6
Series7

Series5

Series8

Series3

Series1
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

74
75

Figure 5: Pareto front series

4. Conclusions
The method was successfully applied in the RAP problem. The results obtained by the algorithm provided a
threshold value reduces the optimal Pareto set in such way that gives the decision maker collections of values to
select simplifying his or her election task. This methods is not affected by the non-convexity of the Pareto front
because of the optimal values selection depends only on the angle formed by the and the weight (
)
cone axis vector.

References
1.
2.
3.

4.

Jurgen Branke Kalyanmoy Deb,Kaisa Miettinen Roman Sowinski in: Multiobjective Optimization,
Interactive and Evolutionary Approaches, Springer-Verlag Berlin Heidelberg 2008
Zitzler, E. and Thiele, L., (1999), Multi-objective Evolutionary Algorithms: A Comparative Case Study
and the Strength Pareto Approach, IEEE Transactions on Evolutionary Computation, 3(4) 257-271
Taboada,Baheranwala,Coit,Wattanapongsakorn(2007),"Practical
solutions
for
multi-objective
optimization:An application to system reliability design problems",Reliability Engineering and Systems
Safety .92,314-322
Halme,Korhonen (2000). Restricting weights in value efficiency Analysis,European Journal of Operation
Research,126,175-188

Proceedings of the 2013 Industrial and Systems Engineering Research Conference

A. Krishnamurthy and W.K.V. Chan, eds.

5.
6.
7.
8.
9.

Venkat, V., Jacobson, S., Stori, J. 2004. "A Post-Optimality Analysis Algorithm for Multi-Objective
Optimization". Computational Optimization and Applications. 28, 357-372.
Padhye, N., Kalia, S. and Deb, K. 2009. Multi-Objective Optimization and Multi-criteria Decision Making
For FDM Using Evolutionary Approaches
Carrillo, Aguirre, Taboada , (2011)"Applications and performance of the non-numerical ranking
preferences method for post-Pareto optimality", Procedia Computer Science,6,243-248
Carrillo, Taboada . (2012)" A Post-Pareto Approach for Multi-Objective Decision Making Using a NonUniform Weight Generator Method". Complex Adaptive Systems, Publication 2. Washington D.C.
Carrillo, Taboada . (2012)" General Iterative procedure for the Non-numerical ranking preferences method
in Multi-objective Optimization". Complex Adaptive Systems, Publication 2. Washington D.C.

10. Taboada, et al D. (2007)."Practical solutions for multi-objective optimization: An application to systems

reliability design problems". Reliability Engineering and System Safety 92 (2007) 314322
11. Taboada, H. & Coit, D. (2007). Data Clustering of Solutions for Multiple Objective System Reliability
Optimization Problems. Quality Technology & Quantitative Management Journal, 4(2):35-54
12. Runckler, T. 2009. "Fuzzy cMeans" Herding,IFSA-EUSFLAT 2009
13. Li,Liao,Coit,(2009)."A two-stage,apprach for multi-objective decision making with applications to system
reliability optimization",Reliability Engineering and System Safety,94,1585-1592.
14. Oswaldo Aguirre MSa,*, Heidi Taboada PhDb ."A Clustering Method Based on Dynamic Self organizing
Trees for Post-Pareto Optimaliyy Designs".Complex Adaptive Systems, Volume 1 Cihan H. Dagli, Editor
in Chief Conference Organized by Missouri University of Science and Technology 2011- Chicago, IL