A Distance Measuring Algorithm for Location Analysis

Ruilin Ouyang, Dinghao Ma, M. S. Morshed and Md. Noor-E-Alam*

Department of Mechanical and Industrial Engineering
Northeastern University
360 Huntington Ave, Boston, MA 02115, USA

Abstract:
Approximating distance is one of the key challenge in a facility location problem. Several
algorithms have been proposed, however, none of them focused on estimating distance between
two concave regions. In this work, we present an algorithm to estimate the distance between two
irregular regions of a facility location problem. The proposed algorithm can identify the distance
between concave shape regions. We also discuss some relevant properties of the proposed
algorithm. A distance-sensitive capacity location model is introduced to test the algorithm.
Moreover, several special geometric cases are discussed to show the advantages and insights of
the algorithm.

Key words: Distance measurement, facility location problem, 𝒍𝒑 distance, geometric distance.

1. Introduction:

Calculating distance between two irregular shape regions is known as a key challenge in the
context of facility location problem. It’s very easy to calculate the geometrical distance between
two points. However, when we consider the distance between two irregular regions, it’s difficult
to say that we have an accurate solution. Several algorithms have been developed to estimate the
distance between regions - some of them only works well under a specific condition.

P-norm distance or “𝑙𝑝 distance” is widely used in the facility location problem to calculate

*Corresponding author email: mnalam@neu.edu

1
distance between two points. The following equation for “𝑙𝑝 distance” was proposed by Francis
(1967).
𝑛 1⁄
𝑝

𝑙𝑝 (𝑞, 𝑟) = [∑|𝑞𝑖 − 𝑟𝑖 |𝑝 ] = ‖𝑞 − 𝑟‖𝑝 = ‖𝑟𝑞

⃗⃗⃗⃗ ‖𝑝
𝑖=1

; where 𝑞 = (𝑞1 , … , 𝑞𝑛 ) and 𝑟 = (𝑟1 , … , 𝑟𝑛 ) are two points in 𝑛-dimensional space (i.e.𝑞, 𝑟 ∈
ℝ𝑛 ). Love and Morris (1972) compared seven different distance functions including 𝑙𝑝 functions,
weighted 𝑙𝑝 functions and weighted Euclidean functions on two road distance samples to
determine the accuracy of various distance functions. They considered two criterions with respect
to the accuracy of estimation, which are:
• Sum of absolute deviations (the difference between actual road miles and estimated
road miles)
• Sum of squares (more sensitive and require more accuracy on shorter distance).

Love and Morris (1979) further compared different parameters for the “𝑙𝑝 distance” model to make
the distance calculation more accurate. They used statistical tests to compare the results on seven
samples and concluded that “the more parameters used in the model, the more accurate and
complexity will get”. Love and Dowling (1985) used “𝑙𝑘,𝑝 distance” model for floor layout
problems to find the appropriate k, p values compared to the “𝑙1 distance” (or rectangular distance)
model. In the literature, authors stated that the “𝑙1 distance” model was suitable for obtaining
optimal facility locations while “𝑙𝑘,𝑝 distance” model should be used when total cost function is
considered by adjusting k values rather than p values. Despite several generalizations of this
algorithm, none of these algorithms mentioned can successfully determine the exact location of
the potential point is in each region.

Batta et al. (1989) solved the location problem with convex forbidden regions using Manhattan
metric (p=1). Amiri-Aref et al. (2011) used P-norm distance model to find the new facility which
has maximum distance to other existing facilities with respect to the same weights and line barriers.
They emphasized the rectilinear distance metric (p=1) in the formulation and stated that further
research can be extended on different distance functions. Moreover, Amiri-Aref et al. (2013) also
presented the same distance metric (p=1) to solve the Weber location problem with respect to a

2
probabilistic polyhedral barrier. They transferred the barrier zone from a central point of a barrier
with extreme points to a probabilistic rectangular-shaped barrier. Later, Amiri-Aref et al. (2015)
utilized the “𝑙1 distance” model to deal with relocation problem with a probabilistic barrier. For
solving the restricted location-relocation problem with finite time horizon, some authors developed
three approaches which are well established in the literature. Dearing and Segars (2002) presented
that for solving single facility location problems with barriers using rectilinear distance model
(p=1), the result remains the same for original barriers and modified barriers.

Aneja and Parlar (1994) discussed two algorithms (Convex FR and Location with BTT) that can
solve Weber Location problems with forbidden regions for 1 < 𝑝 ≤ 2 when using 𝑙𝑝 distance
model to calculate the distance between demand points and facility location points. Bischoff and
Klamrot (2007) illustrated a solution method with respect to the problem for locating a new facility
while given a set of existing facilities in the presence of convex polyhedral forbidden areas. This
method at first decomposes the problem and then use genetic algorithm to find appropriate
intermediate points. In the literature, only p=2 or Euclidean distance has been considered and it
has been showed that this algorithm was very efficient. Altınel et al. (2009) considered the heuristic
approach for the multi-facility location problem with probabilistic customer locations. To calculate
the expected distance between probabilistic demand points and the facility locations, they used
Euclidean distance, squared Euclidean distance, rectilinear distance and weighted 𝑙1,2 -norm
distance models.

By virtue of network theory, we can find another way of calculating distances for location problem.
Blanquero et al. (2016) represented the competitive location problem on a network with customers
as the nodes in the network. They put the facilities on the edges and calculated the distances as the
lengths of the shortest path from each customer to the nearest facility. Peeters (1998) discussed a
new algorithm to solve the location problem by calculating distance as the sum of the lengths of
arcs between two vertexes and the length between a vertex and an arbitrary point on that arc. The
author used nine different models to test the new algorithms and showed much better performances
than the old algorithm which calculates the distance between each node first.

Very little effort was made in the prior literature to consider concave regions thoughtfully when

3
they estimated the distance. It may cause problems in real world application. In section 2, we will
talk about the problems that could be occurred in a real-world scenario. In section 3, we will
introduce our distance measurement algorithm which considers the concave regions. We will also
discuss some properties as well as one observation with proof about the proposed algorithm. In
section 4, we will illustrate a capacity location model which is very sensitive to the distance
between regions. In Section 5, we will show how our proposed algorithm could be used to find
optimal location using a test-case example. In section 6, we will consider several cases to present
advantages and insights of the proposed algorithm.

2. Problem Description:

In the prior literature, geometric-center was widely used for distance calculation. But using the
geometric-center of a region to represent itself may cause problems. Let us imagine a concave
region where the geometric-center is out of the region. Now, if we use geometric-center point to
represent the region then the distance between this region and another concave region might be
zero. For example, in the following Figure 1, we can see Shanghai city with its outline represented
as some collection concave regions. As we can see, the region for west district is a concave region;
its geometric-center may fall into east district. If we transform the region a little bit, then the two
geometric-centers may be the same point thus, the distance becomes zero. Therefore, in most cases,
the geometric-center can represent a region well but under special situation, it will cause fatal error
or inaccuracy in the final decision of a location problem. To prevent this situation, we will develop
an algorithm to measure the distance between two irregular regions. We will only consider the
continuous facility location problem which will allow us to set up a facility at any location. The
algorithm can prevent the potential point falls out of a region itself. It also has good property within
the region it represents. We will also show the relationship among the potential points we get with
geometric-centers.

4
 Figure 1: Map and Outline of center of Shanghai (adopted from Googlemap).

3 Proposed Algorithm:

3.1 Distance Measurement Algorithm:

Suppose, we have a region S, the point we use to represent 𝑆 is a point(𝑥, 𝑦) ∈ 𝑆, such that, the

value ∬(𝑥 ,𝑦 )∈𝑆 √(𝑥 − 𝑥𝑖 )2 + (𝑦 − 𝑦𝑖 )2 𝑑𝑥𝑑𝑦 has a minimum value. The distance between two
𝑖 𝑖

regions is estimated by the geometrical distance between two points which represent the regions.

It means (𝑥, 𝑦) is a point in S, and (𝑥, 𝑦) is the point to make the value of the function

∬(𝑥 ,𝑦 )∈𝑆 √(𝑥 − 𝑥𝑖 )2 + (𝑦 − 𝑦𝑖 )2 𝑑𝑥𝑑𝑦 gains the smallest value.

𝑖 𝑖

We use the geometrical distance between two points we derived from the method above to
estimate the distance between two regions. For example, if we have region 𝑆1 and 𝑆2 , and we get
(𝑥𝑆1 , 𝑦𝑆1 ) and (𝑥𝑆2 , 𝑦𝑆2 ) to represent 𝑆1 and 𝑆2 , then the distance between 𝑆1 and 𝑆2 is 𝑑(𝑆1 , 𝑆2 ) =

2 2
√(𝑥𝑆1 − 𝑥𝑆2 ) + (𝑦𝑆1 − 𝑦𝑆2 ) .

3.2 Theorem:
First, it’s reasonable to say that the point represents the region because (𝑥, 𝑦) is the point that
makes the total travel distance needed to satisfy each demand point to gain the smallest value. The
point we derived from our algorithm also equals to the geometric-center of the region if the

5
geometric-center of the region falls into itself. We will proof the statement in the following
subsections.

3.3 Observation and Proof:

If a region’s geometric-center falls in the region itself, the point (𝑥, 𝑦) mentioned above is
exactly the geometric-center of the region.

Since we are looking for (𝑥, 𝑦) such that, the expression ∬ √(𝑥 − 𝑥𝑖 )2 + (𝑦 − 𝑦𝑖 )2 𝑑𝑥𝑑𝑦 gain
the minimum value, so we can remove the square root since 𝑓(𝑥) = √𝑥(𝑥 ≥ 0) is a strictly
monotonically increasing function. Therefore, we need:

(𝑥, 𝑦) 𝑠. 𝑡. ∬ √(𝑥 − 𝑥𝑖 )2 + (𝑦 − 𝑦𝑖 )2 𝑑𝑥𝑑𝑦 𝑔𝑎𝑖𝑛 𝑡ℎ𝑒 𝑚𝑖𝑛𝑖𝑛𝑢𝑚 𝑣𝑎𝑙𝑢𝑒

⟺ (𝑥, 𝑦) 𝑠. 𝑡. ∬[(𝑥 − 𝑥𝑖 )2 + (𝑦 − 𝑦𝑖 )2 ]𝑑𝑥𝑑𝑦 𝑔𝑎𝑖𝑛 𝑡ℎ𝑒 𝑚𝑖𝑛𝑖𝑛𝑢𝑚 𝑣𝑎𝑙𝑢𝑒

We can also separate the previous expression into two parts, one part only with 𝑥 and one part
only with 𝑦. For∬(𝑥 − 𝑥𝑖 )2 𝑑𝑥𝑑𝑦, since we are looking for a point (𝑥, 𝑦) such that it reaches the
minimum value, we can transfer it to∫(𝑥 − 𝑥𝑖 )2 𝑑𝑥. So, we have:
(𝑥, 𝑦) 𝑠. 𝑡. ∬[(𝑥 − 𝑥𝑖 )2 + (𝑦 − 𝑦𝑖 )2 ]𝑑𝑥𝑑𝑦 gain the minimum value
⟺ (𝑥, 𝑦) 𝑠. 𝑡. ∬(𝑥 − 𝑥𝑖 )2 𝑑𝑥𝑑𝑦 + ∬(𝑦 − 𝑦𝑖 )2 𝑑𝑥𝑑𝑦 gain the minimum value
⟺ (𝑥, 𝑦) 𝑠. 𝑡. ∫(𝑥 − 𝑥𝑖 )2 𝑑𝑥 + ∫(𝑦 − 𝑦𝑖 )2 𝑑𝑦 gain the minimum value

Then, we can write our function into Riemann sum:

𝑛

𝑓𝑜𝑟 𝑥: ∫(𝑥 − 𝑥𝑖 )2 𝑑𝑥 ⟺ lim ∑(𝑥 − 𝑥𝑖 )2 △ 𝑥 .

𝑛→∞
𝑖=1

We obtain the mininum value of the function when ( lim ∑𝑛𝑖=1(𝑥 − 𝑥𝑖 )2 △ 𝑥 ) ′ = 0 ⟺

𝑛→∞
∑𝑛
𝑖=1 𝑥𝑖
lim ∑𝑛𝑖=1 2(𝑥 − 𝑥𝑖 ) = 0 ⟺ 𝑥 = lim . Since the geometric-center falls into the region,
𝑛→∞ 𝑛→∞ 𝑛

here 𝑥 is exactly the coordinates for geometric-center. We can get the same result for 𝑦. So, we
proved that if a region’s geometric-center falls in the region itself, the point (𝑥, 𝑦) mentioned
above is exactly the geometric-center of the region.

6
4 Test Model:

In this section, we will introduce a model for testing the proposed algorithm. It’s a simple capacity
location problem; we have added a constraint to make it sensitive to the distance between regions.

4.1 Parameters

𝑎𝑥 demand of point 𝑥 at time period 𝑎

𝑐 fixed cost of building a facility
𝑑𝑥𝑦 distance between points 𝑥 and 𝑦
𝐿 maximum capacity of one facility
𝑀 large cost parameter

4.2 Variables

𝛾𝑥𝑦 if demand at point 𝑥 is satisfied by facility at point 𝑦, then 𝛾𝑥𝑦 = 1,

else 𝛾𝑥𝑦 = 0
𝜌𝑦 if a facility is built at y at time 𝑎, then 𝜌𝑦 = 1, else 𝜌𝑦 = 0

𝑀𝑖𝑛𝑖𝑚𝑖𝑧𝑒 ∑ 𝜌𝑦 𝑐 (1)
𝑦

𝛾𝑥𝑦 ≤ 𝜌𝑦 (2)
𝑑𝑥𝑚 𝛾𝑥𝑚 ≤ 𝑑𝑥𝑦 + 𝑀(1 − 𝜌𝑦 ) (3)

∑ 𝑎𝑥 𝛾𝑥𝑦 ≤ 𝐿 (4)
𝑥

∑ 𝛾𝑥𝑦 = 1 (5)
𝑦

The objective function (1) is a cost minimization function that aims to minimize the total cost of
building facilities. Constraint (2) indicates that location 𝑦 can satisfy the demand at location 𝑥 if
and only if there’s a facility at location 𝑦. Constraint (3) restricts the demand at location 𝑥 to be
met by its nearest facility. If there is a facility built at location 𝑥, then for any other point 𝑚, 𝛾𝑥𝑚 =
1 if and only if 𝑑𝑥𝑦 < 𝑑𝑥𝑚 . This constraint will make the model very sensitive to the distance

7
between regions. Constraint (4) shows that the total demand served by one facility can’t exceed its
capacity. Constraint (5) restricts that we can only set up one facility at one location.

5 Result Analysis:

We randomly generated a map with several concave regions shown in the following Figure 2. The
proposed algorithm for distance measurement is implemented with MATLAB R 2016a. We get
the geometric-center and the potential points by using our proposed algorithm. The size of the map
is 200×200 pixels and we use a coordinate system on the map. We fixed the origin of coordinate
at the lower left corner of the map and we also assumed that the coordinate of upper left corner is
(200, 0), the coordinate of lower right corner is (0,200). Therefore, all the coordinates we assumed
are integer numbers.

Figure 2: Test-case model map.

8
 Table 1: Coordinates and demand of representative point for each region.
Region Geometric Algorithm Deman Region Geometric Algorith Deman
Numbe -center Center d Numbe -center m Center d
r r
1 (189,25) (189,25) 9 16 (75,93) (64,112) 2
2 (186,66) (186,66) 10 17 (76,93) (76,93) 5
3 (184,112) (184,112) 2 18 (19,19) (19,19) 10
4 (186,149) (186,149) 10 19 (28,46) (28,46) 8
5 (169,162) (169,162) 7 20 (7,90) (7,90) 10
6 (148,183) (148,183) 1 21 (22,106) (22,106) 7
7 (73,183) (73,183) 3 22 (42,93) (37,92) 1
8 (121,136) (121,127) 6 23 (10,132) (10,132) 9
9 (151,98) (153,106) 10 24 (11,165) (11,165) 10
10 (163,38) (163,38) 10 25 (31,138) (31,138) 7
11 (157,15) (157,15) 2 26 (30,168) (30,168) 8
12 (108,30) (107,32) 10 27 (67,163) (67,163) 8
13 (115,23) (115,23) 10 28 (67,150) (67,150) 4
14 (133,83) (133,83) 5 29 (84,154) (84,154) 7
15 (72,27) (72,27) 9

Table 1 shows the comparison between coordinates of representative location using geometric-
center and our proposed algorithm for the point of each region. We can see that for region 8, 9, 12,
16 and 22 the representative points are different, and they are all concave regions. Now, we will
use this test-case instance to find optimal location using the test case model.

The model is solved on a computer with Intel(R) Xeon(R) E3-1231 v3 CPU running at 3.4 GHz
with 16 GB memory and a 64-bit operating system. For implementation, AMPL programming
language is used, and CPLEX 12.6.3 solver is utilized to solve. In this instance, other parameters
are assumed as follows: 𝑐 = 200, 𝐿 = 50, 𝑀 = 10000. For 𝑎𝑥 , values were randomly generated
ranging from 1 to 10. The number in each region represents the demand at each region. Optimal

9
facility assignments are shown in Figure 3 and Figure 4 using geometric-center and algorithm
center, respectively.

Figure 3: Optimal facility assignment using geometric-center with demand.

Figure 4: Optimal facility assignment using the proposed algorithm with demand.

As we can see from Table 2, two algorithms provide us the same optimal solution. However, our
proposed algorithm requires less number of simplex iterations and branch-and-bound nodes. While
the purpose of our algorithm is to make the distance estimation more accurate, the difference of
objective function value or travel distance cannot provide better insight about it. Therefore, in the
next section, we will present three test-cases to analyze the advantages of our proposed algorithm.

10
Comment: 1. The difference of number of simplex iterations and branch & bound nodes does not
show which algorithm is better or not. It’s just a coincidence. 2. It’s hard to tell the difference or
the advantage of our algorithm from a real case or general case.

6 Special Case Analysis:

In this section, we present three special cases to show further the usefulness and performance of
the proposed algorithm.

Case1:
We can see from Figure 5, region 1 is a concave region and its geometric-center falls into region
2, and the actual distance between the two geometric-centers are 0. If we use this result to estimate
the distances in a location problem, then there’s no difference for the location of region 1 and 2. If
there’s another region available then the demand in that region can be either served by region 1 or
region 2. Clearly, we don’t want this situation to occur in real life problems. In Figure 6, we can
see the representative points for region 1 falls on the boundary of region 1, which will show the
difference between the region 1 and 2. If there’s another region, the demand of that region can be
satisfied by preference.

Indicates geometric center of part I

* Indicates for geometric center of

part II

Figure 5: Geometric center for case 1.

11
 Indicates algorithmic center of part
I

* Indicates algorithmic center of part

II

Figure 6: Algorithmic center for case 1.

Case 2:
In this case, we have 3 regions. From Figure 7, we can see that the distance between region 1 and
region 2 are smaller than the distance between region 2 and region 3. This means that the demand
at region 2 are more likely to be served by region 1. But in Figure 8, by our algorithm, the distance
between region 2 and region 3 are smaller than the distance between region 1 and region 2. This
means the demand at region 2 are more likely to be served by region 1. From the graph, we can
see that it’s better to let the region 2 served by region 3 since it looks closer to each other.

Indicates geometric center of part I

* Indicates geometric center of part II

Indicates geometric center of part III

Figure 7: Geometric center for case 2.

12
 Indicates algorithmic center of part I

* Indicates algorithmic center of part II

Indicates algorithmic center of part III

Figure 8: Algorithmic center for case 2.

Case 3:
In this case, we can see from Figure 9 and Figure 10, by both algorithms, the representative points
fall together. This is a very special case for which our algorithm cannot identity the location of two
regions.

Indicates geometric center of part I

* Indicates geometric center of part II

Figure 9: Geometric center for case 3.

13
 Indicates algorithmic center of part I

* Indicates algorithmic center of part II

Figure 10: Algorithmic center for case 3.

7. Conclusion:

The proposed algorithm can identify the difference between regions when there are concave
regions in the map. We have shown that under several conditions, the algorithm performs better
than the geometric-center approach. When the geometric-center falls out of the region itself, our
algorithm has a significant advantage to make the measurement more accurate. When the
geometric-center falls into the region itself, our algorithm will keep the original feature of the
geometric-center, and the algorithm center and geometric-center fall at the same point. The
algorithm is also easy to implement. However, there are limitations of this algorithm; the most
noticeable one is that we can only implement it on a continuous facility location problem.

References:

[1] Altınel, I. K., Durmaz, E., Aras, N., & Özkısacık, K. C. (2009). A location–allocation
heuristic for the capacitated multi-facility Weber problem with probabilistic customer
locations. European Journal of Operational Research 198 (3): 790-799.
doi:10.1016/j.ejor.2008.10.014
[2] Amiri-Aref, M., Javadian, N., Tavakkoli-Moghaddam, R., & Baboli, A. (2013). A new
mathematical model for the Weber location problem with a probabilistic polyhedral
barrier. International Journal of Production Research 51 (20): 6110-6128.
doi:10.1080/00207543.2013.796422
[3] Amiri-Aref, M., Farahani, R. Z., Javadian, N., & Klibi, W. (2015). A rectilinear distance
location–relocation problem with a probabilistic restriction: mathematical modelling and

14
solution approaches. International Journal of Production Research 54 (3): 629-646.
doi:10.1080/00207543.2015.1013642
[4] Amiri-Aref, M., Javadian, N., Tavakkoli-Moghaddam, R., & Aryanezhad, M. (2011). The
center location problem with equal weights in the presence of a probabilistic line
barrier. International Journal of Industrial Engineering Computations 2 (4): 793-800.
doi:10.5267/j.ijiec.2011.06.002
[5] Aneja, Y. P., & Parlar, M. (1994). Algorithms for Weber Facility Location in the Presence
of Forbidden Regions and/or Barriers to Travel. Transportation Science 28 (1): 70.
[6] Batta, R., Ghose, A., & Palekar, U. S. (1989). Locating Facilities on the Manhattan Metric
with Arbitrarily Shaped Barriers and Convex Forbidden Regions. Transportation Science 23 (1):
26-36. doi:10.1287/trsc.23.1.26
[7] Bischoff, M., & Klamroth, K. (2007). An efficient solution method for Weber problems with
barriers based on genetic algorithms. European Journal of Operational Research 177 (1): 22-
41. doi:10.1016/j.ejor.2005.10.061
[8] Blanquero, R., Carrizosa, E., G.-Tóth, B., & Nogales-Gómez, A. (2016). P-facility Huff
location problem on networks. European Journal of Operational Research 255 (1): 34-42.
doi:10.1016/j.ejor.2016.04.039
[9] Dearing, P. M., & Segars Jr., R. (2002). An Equivalence Result for Single Facility Planar
Location Problems with Rectilinear Distance and Barriers. Annals of Operations Research 111
(1-4): 89-110. doi:10.1023/A:1020945501716
[10] Francis, R. L. (1967). Letter to the Editor—Some Aspects of a Minimax Location
Problem. Operations Research 15 (6): 1163-1169. doi:10.1287/opre.15.6.1163
[11] Love, R. F., & Dowling, P. D. (1985). Optimal Weighted lp Norm Parameters for Facilities
Layout Distance Characterizations. Management Science 31 (2): 200-206.
doi:10.1287/mnsc.31.2.200
[12] Love, R. F., & Morris, J. G. (1979). Mathematical Models of Road Travel
Distances. Management Science 25 (2): 130-139. doi:10.1287/mnsc.25.2.130
[13] Love, R. F., & Morris, J. G. (1972). Modelling Inter-City Road Distances by Mathematical
Functions. Operational Research Quarterly (1970-1977) 23 (1): 61. doi:10.2307/3008514
[14] Peeters, P. H. (1998). Some new algorithms for location problems on networks. European
Journal of Operational Research 104 (2): 299-309. doi:10.1016/s0377-2217(97)00185-9

15