Identification of V-Formations and Circular and Doughnut Formations in a Set of Moving Entities with Outliers

Abstract

Diverse movement patterns may be identified when we study a set of moving entities. One of these patterns is known as a V-formation for it is shaped like the letter V. Informally, a set of entities shows a V-formation if the entities are located on one of their two characteristic lines. These lines meet in a position where there is just one entity considered the leader of the formation. Another movement pattern is known as a circular formation for it is shaped like a circle. Informally, circular formations present a set of entities grouped around a center in which the distance from these entities to the center is less than a given threshold. In this paper we present a model to identify V-formations and circular formations with outliers. An outlier is an entity which is part of a formation but is away from it. We also present a model to identify doughnut formations, which are an extension of circular formations. We present formal rules for our models and an algorithm for detecting outliers. The model was validated with NetLogo, a programming and modeling environment for the simulation of natural and social phenomena.

1. Introduction

Diverse movement patterns may be identified when we study a set of moving entities, for example, a flock of birds [1] and a school of fish [2]. One of these patterns is known as a V-formation for it is shaped like the letter V; see Figure 1. Another movement pattern is known as a circular formation for it is shaped like a circle; see Figure 2.

Figure 1 

V-formation in birds. Source: [9].

Figure 2 

Circular formation in a fish bank. Source: [21].

Informally, a set of entities shows a V-formation if the entities are located on one of their two characteristic lines. The lines meet in a position where there is just one entity considered the leader of the formation [3]. Several authors have analyzed V-formations. In [4, 5], there is an attempt to explain from a physical point of view the reasons why certain species of birds, such as Canadian geese (Branta canadensis), red knots (Calidris canutus), and plovers (Calidris alpina), tend to fly this way.

Other authors try to simulate V-formations at a computational level. For instance, Nathan and Barbosa [6] propose a model based on rules that allows us to generate V-formations depending on specific parameters. The authors validated their model using NetLogo [7], a programming and modeling environment to simulate natural and social phenomena.

On the other hand, a circular formation is a set of entities grouped around a common center and where the entities’ distance to the center is less than a given threshold. Regarding related works with circular formations, we identified the following.

In [8] the authors experimented with a set of data referring to the movement of different animal species. It was found that despite being in different ecosystems, species follow similar behavioral patterns. The authors also tried to model general grouping behaviors of fish, birds, insects, and even people. One of these behaviors is circular formation in which they identified physical forces: attraction, repulsion, alignment, and frontal interaction.

On the other hand, researchers in the field of robotics and in control theory, inspired by social grouping phenomena and by the patterns of birds and fish, have developed applications to coordinate the movement of multivehicle systems. Among these patterns are circular [9] and V-formations; see Figure 3.

(a)

(b)

Figure 3 

Robots in circular formation and planes in V-formation. Sources: [11, 12].

On the other hand, regardless of the type of formation, Reynolds [2, 10] proposes a computer model of coordinated animal motion for bird flocks and fish schools. His model considers three simple rules that act upon individuals: (i) separation: steer to avoid crowding local flockmates, (ii) alignment: steer towards the average heading of local flockmates, and (iii) cohesion: steer to move toward the average position of local flockmates.

Although the previous works allow the simulation of a set of moving entities, they are not aimed at the explicit identification of V-formations and circular formations. The identification of these types of formations may be useful in fields as zoology, to analyze the movement of birds [3] and fish [11], and in the military and videogames where squadrons of combat planes, ships, and robots usually assume these types of formations [3, 12]. V-formations also usually appear in stock markets (stock prices) [13]. In this direction the model of Andersson et al. [14], although it is not aimed at the explicit identification of formations, identifies an entity leader in a set of moving entities. We use their model as the basis for identifying V-formations and circular formations, as we will show in our paper.

This paper is organized as follows. In Section 2, we present our models for V-formations and circular and doughnut formations, which are an extension of circular formations. In Section 3, we present support for outliers in our models. In Section 4, we present experiments. In Section 5, we conclude the paper and propose future works.

2. V-Formations and Circular and Doughnut Formations

2.1. Andersson’s Model

Next, we present the essential elements of Andersson’s model [14].

Consider a set 𝐸 of 𝑛 entities {𝑒₁,𝑒₂,…,𝑒_𝑛} that move in a space, usually in a region during a time interval [𝑡₁,𝑡_𝑓]. This space is represented by the Euclidian plane. Time is represented continuously. 𝑇𝑝 denotes an infinite set of time-points: {𝑡∣𝑡∈[𝑡₁,𝑡_𝑓]}. On the other hand, 𝑇𝑠 denotes a finite set of discrete time-points {𝑡₁,𝑡₂,…,𝑡_𝑓}. Every 𝑡_𝑖∈𝑇𝑠 represents a time-step and corresponds to a time-point in which the position of a moving entity (and perhaps other data) was recorded. A unit-time-interval is an open interval between two consecutive time-steps, that is, (𝑡_𝑖−1,𝑡_𝑖), for all 𝑡_𝑖∈𝑇𝑠, 𝑖≠1.

The coordinates of an entity at a time-point are given by a pair of functions 𝑥𝑝𝑜𝑠 and 𝑦𝑝𝑜𝑠, both having the signature 𝐸×𝑇𝑝→ Real. Andersson assumes that between two consecutive time-steps an entity moves along a straight line with constant direction and constant velocity. This assumption has been adopted in other works (see [15–17] for a discussion). Therefore, in such a model in order to be practical enough, the recording frequency of the positions of moving objects should be adapted to the nature of each domain. For example, when detecting bird formations the elapsed time between two consecutive time-steps should not be longer than 30 seconds.

Consider an entity 𝑒_𝑖 at a time-point 𝑡, where 𝑡_𝑥−1<𝑡<𝑡_𝑥, 𝑡_𝑥∈𝑇𝑠, 𝑥≠1; then the direction of 𝑒_𝑖 at 𝑡, denoted by 𝑑(𝑒_𝑖,𝑡), is the angle specified by the straight line segment that goes from (𝑥𝑝𝑜𝑠(𝑒_𝑖,𝑡_𝑥−1), 𝑦𝑝𝑜𝑠(𝑒_𝑖,𝑡_𝑥−1)) to (𝑥𝑝𝑜𝑠(𝑒_𝑖,𝑡_𝑥),𝑦𝑝𝑜𝑠(𝑒_𝑖,𝑡_𝑥)); see Figure 4. The angle 𝑑(𝑒_𝑖,𝑡) is between [0,2𝜋). Andersson declares the direction of an entity at a time-step 𝑡_𝑥 to be undefined, because at a time-step an entity might change its direction (for more details see [14]).

Figure 4 

𝑑(𝑒𝑖,𝑡): angle of entity 𝑒𝑖 at a time-point 𝑡.

The front region of an entity 𝑒_𝑖 is a region associated with 𝑒_𝑖 at a time-point 𝑡, where 𝑡_𝑥−1<𝑡<𝑡_𝑥, 𝑡_𝑥∈𝑇𝑠, 𝑥≠1, which represents the perception region of an entity, for example, a visual or an auditive range. The front region of 𝑒_𝑖 is defined as follows: consider three straight line segments 𝑠₀, 𝑠₁, and, 𝑠₂, each of length 𝑟. Each segment has an endpoint at (𝑥𝑝𝑜𝑠(𝑒_𝑖,𝑡),𝑦𝑝𝑜𝑠(𝑒_𝑖,𝑡)). The direction of the segment 𝑠₀ is 𝑑(𝑒_𝑖,𝑡). Segments 𝑠₁ and 𝑠₂ form angles of 𝛼/2 and −𝛼/2(0≤𝛼<2𝜋) with regard to segment 𝑠₀, respectively. The circular sector (a wedge-shaped region) with a radius 𝑟(𝑟≥0), delimited by 𝑠₁ and 𝑠₂, makes up the front region of 𝑒_𝑖 at 𝑡 and it is denoted by front(e_i, t); see Figure 5.

Figure 5 

Front region of an entity 𝑒𝑖 at a time-point 𝑡.

An entity 𝑒_𝑗 is in front of an entity 𝑒_𝑖, at a time-point 𝑡 (𝑡∈𝑇𝑝,𝑡∉𝑇𝑠), where 𝑡_𝑥−1<𝑡<𝑡_𝑥, 𝑡_𝑥∈𝑇𝑠, 𝑥≠1, if 𝑒_𝑗 is in the front region of 𝑒_𝑖, that is, if (𝑥𝑝𝑜𝑠(𝑒_𝑗,𝑡),𝑦𝑝𝑜𝑠(𝑒_𝑗,𝑡)) is inside front(e_i, t); this is denoted 𝑒_𝑗∈𝑓𝑟𝑜𝑛𝑡(𝑒_𝑖,𝑡) and we say that 𝑒_𝑖 follows 𝑒_𝑗 (𝑒_𝑖 is a follower of 𝑒_𝑗) at 𝑡. We consider an additional restriction to establish that one entity follows another: let 𝛽∈[0,𝜋]; then entity 𝑒_𝑖 follows entity 𝑒_𝑗 at 𝑡 if (i)  𝑒_𝑗∈𝑓𝑟𝑜𝑛𝑡(𝑒_𝑖,𝑡) and (ii) ‖𝑑(𝑒_𝑖,𝑡)−𝑑(𝑒_𝑗,𝑡)‖≤𝛽; see Figure 6. An entity 𝑒_𝑖 is said to follow an entity 𝑒_𝑗 during a time interval [𝑡_𝑎,𝑡_𝑏], where 𝑡_𝑎 and 𝑡_𝑏 are time-points in 𝑇𝑝, if and only if 𝑒_𝑖 follows 𝑒_𝑗 at 𝑡, for all 𝑡∈[𝑡_𝑎,𝑡_𝑏], 𝑡∉𝑇𝑠.

Figure 6 

Entity 𝑒𝑖 follows entity 𝑒𝑗 at a time-point 𝑡.

An entity 𝑒_𝑖 is said to be a leader of a formation during a time interval 𝐼=[𝑡_𝑎,𝑡_𝑏], where 𝑡_𝑎 and 𝑡_𝑏 are time-points in 𝑇𝑝, if (i) 𝑒_𝑖 does not follow any entity during 𝐼 and (ii) 𝑒_𝑖 is at least followed by 𝑚 entities at each time-point 𝑡∈𝐼, 𝑡∉𝑇𝑠. It is said that there is a leadership pattern if 𝑒_𝑖 is a leader of at least 𝑚 entities for at least 𝑘 unit-time-intervals.

2.2. V-Formations

Consider the formation in Figure 7 where entities displayed a V-formation at a time-point 𝑡, and 𝑒₄ is the leader entity. Note that if we applied Andersson’s model [14] to this formation, we would observe that no entity contains 𝑒₄ in its front region. Therefore, this example shows that Andersson’s model [14] is inadequate to identify leadership patterns in these types of formations, unless we consider a wide front region for each entity, as shown in Figure 8. Thus, the entities on the extremes, 𝑒₁ and 𝑒₇, will require a wide front region to be able to contain the leader entity 𝑒₄; this is not a realistic assumption in V-formations since an entity’s visual field is not usually that wide [4].

Figure 7 

An example of a V-formation at a time-point 𝑡.

Figure 8 

A V-formation: entities 𝑒1 and 𝑒7 would require a wide front region in order to contain the entity leader 𝑒4.

In the following section, we extend Andersson’s model to identify leadership patterns in V-formations.

2.3. V-Formation Model

Let 𝐹={𝑒₁,𝑒₂,…,𝑒_𝑛} be a set of moving entities at a time-point 𝑡, card(𝐹)≥3. 𝐹 is a V-formation at 𝑡 if the following hold.(i)There exist nonempty subsets 𝐴 and 𝐵 of 𝐹 such that card(𝐴)≥2, card(𝐵)≥2, 𝐴∪𝐵=𝐹, and 𝐴∩𝐵={𝑒_𝑘}, where 𝑒_𝑘 is said to be the leader entity of the formation.(ii)Entities in 𝐴 tend to form a straight line 𝑙₁ at 𝑡.(iii)Entities in 𝐵 tend to form a straight line 𝑙₂ at 𝑡.(iv)Straight lines 𝑙₁ and 𝑙₂ meet at position (𝑥𝑝𝑜𝑠(𝑒_𝑘,𝑡), 𝑦𝑝𝑜𝑠(𝑒_𝑘,𝑡)).(v)∠𝑎𝑝_𝑡>0 (the smallest angle defined by straight lines 𝑙₁ and 𝑙₂ at 𝑡).Regarding conditions (ii) and (iii), to establish if a set of entities tend to form a straight line, we use Pearson’s correlation coefficient 𝑟 [18]. Thus, given a set of points {(𝑥₁,𝑦₁),(𝑥₂,𝑦₂),…,(𝑥_𝑛,𝑦_𝑛)}, 𝑟 indicates how much they adjust to a straight line (linearity degree). 𝑟∈(−1,1); if |𝑟|≈1, then points tend to form a straight line. 𝑟 is calculated as shown in (1). A threshold 𝜇_𝑝 may be specified by the user to indicate the linearity degree required for the lines of the formation; that is, |𝑟|≥𝜇_𝑝: ∑𝑟=𝑛∗^𝑛_𝑖=1𝑥_𝑖∗𝑦_𝑖−∑^𝑛_𝑖=1𝑥_𝑖∗∑^𝑛_𝑖=1𝑦_𝑖𝑛∗∑^𝑛_𝑖=1𝑥²_𝑖−∑^𝑛_𝑖=1𝑥_𝑖²∗𝑛∗∑^𝑛_𝑖=1𝑦²_𝑖−∑^𝑛_𝑖=1𝑦_𝑖². (1) To obtain the equation for each straight line (𝑦=𝑚𝑥+𝑏) characteristic of the formation (𝑙₁ and 𝑙₂) we may apply the equations that correspond to the straight line which most suits a set of points {(𝑥₁,𝑦₁),(𝑥₂,𝑦₂),…,(𝑥_𝑛,𝑦_𝑛)}; see (2). Table 1 presents the results for the formation in Figure 9: ∑𝑚=𝑛∗^𝑛_𝑖=1𝑥_𝑖∗𝑦_𝑖−∑^𝑛_𝑖=1𝑥_𝑖∗∑^𝑛_𝑖=1𝑦_𝑖𝑛∗∑^𝑛_𝑖=1𝑥²_𝑖−∑^𝑛_𝑖=1𝑥_𝑖²,∑𝑏=^𝑛𝑖=1𝑦_𝑖−𝑚∗∑^𝑛_𝑖=1𝑥_𝑖𝑛. (2) Regarding condition (v), ∠𝑎𝑝_𝑡 is calculated as follows: we get straight lines 𝑙₁ and 𝑙₂ from the formation and find the smallest angle between them as follows. 𝑎 is the positive angle in [0,2𝜋) that is specified by the line segment that goes from 𝑝_𝑘 to 𝑝_𝑖, where 𝑝_𝑘 is the position of 𝑒_𝑘 (leader entity) according to equation for 𝑙₁ and 𝑝_𝑖 is the position of 𝑒_𝑖∈𝐴, 𝑒_𝑖≠𝑒_𝑘 according to equation for 𝑙₁. 𝑏 is the positive angle in [0,2𝜋) that is specified by the line segment that goes from 𝑝_𝑘 to 𝑝_𝑗, where 𝑝_𝑘 is the position of 𝑒_𝑘 (leader entity) according to equation for 𝑙₂ and 𝑝_𝑗 is the position of 𝑒_𝑗∈𝐵, 𝑒_𝑗≠𝑒_𝑘, according to equation for 𝑙₂.Let 𝑤=|𝑎−𝑏|; then ∠𝑎𝑝_𝑡=𝑤 if 𝑤≤𝜋, and ∠𝑎𝑝_𝑡=2𝜋−𝑤, otherwise. For example, in the formation in Figure 9, 𝑎=0.7 rad (40.36°), 𝑏=5.14 rad (294.44°), and 𝑤=4.43 rad (254.08°); thus, ∠𝑎𝑝_𝑡=1.84 rad (105.91°).

Table 1 

Applying our model to the formation in Figure 9.

Figure 9 

A formation of entities at a time-point 𝑡.

In an analogous way to leadership patterns, we say that there is a V-formation pattern if the set 𝐹 of moving entities shows a V-formation for at least 𝑘 unit-time-intervals.

2.4. Circular and Doughnut Formations

In this section, we present a formal model to identify circular formations. We also present doughnut formations, which are an extension of circular formations.

2.4.1. Circular Formation Model

Let 𝐹={𝑒₁,𝑒₂,…,𝑒_𝑛} be a set of moving entities at a time-point 𝑡. Let 𝐶 be the minimum circle that encloses the entities in 𝐹 and let 𝐶_𝑟 be its radius and let (𝐶_𝑥,𝐶_𝑦) be its center (Welzl [19] and Megiddo [20] have showed that this circle can be found in linear time). 𝐹 is a circular formation at 𝑡 if(i)𝐶_𝑟≤𝑅 (a user-given radius with center at (𝐶_𝑥,𝐶_𝑦)),(ii)the minimum number of members in the formation is 𝑁_min.In Figure 10, we show a circular formation with 9 members.

Figure 10 

A circular formation.

2.4.2. Doughnut Formation Model

Let 𝐹={𝑒₁,𝑒₂,…,𝑒_𝑛} be a set of moving entities at a time-point 𝑡. Let 𝐶 be the minimum circle that encloses the entities in 𝐹 and let 𝐶_𝑟 be its radius and (𝐶_𝑥,𝐶_𝑦) its center. 𝐹 is a doughnutformation at 𝑡 if(i)𝐶_𝑟≤𝑅_ext (external radius, a user-given radius with center at (𝐶_𝑥,𝐶_𝑦)),(ii)for each entity of the formation its distance 𝑑 regarding (𝐶_𝑥,𝐶_𝑦) is greater than 𝑅_int (internal radius, a user-given radius with center at (𝐶_𝑥,𝐶_𝑦), 𝑅_int≤𝐶_𝑟),(iii)the minimum number of members in the formation is 𝑁_min.Note that a circular formation is a particular case of a doughnut formation, where 𝑅_int=0. In Figure 11, we show a doughnut formation with 8 members.

Figure 11 

A doughnut formation.

Circular and doughnut patterns are also defined in an analogous way to a V-formation pattern.

3. Outliers

Informally, for V-formations an outlier is an entity which is away from its characteristic lines, and for circular formations it is an entity found beyond the radius of the formation (𝑅).

3.1. Outliers in V-Formations

There are sets of entities which tend to display a V-formation; they may have at a time-point 𝑡 entities which are away from their characteristic lines and which, therefore, affect Pearson’s coefficient. These entities are called outliers [22, 23].

There are numerous methods to detect outliers in different domains [24]. Listing 1 presents an algorithm that receives an array of 𝑚 entities (lineMembers) which form a characteristic line of a formation at a time-point 𝑡. The algorithm determines if, after removing a maximum number of entities on the given array, the Pearson’s coefficient surpasses a given threshold 𝜇_𝑝. For example, if it is permitted to remove a maximum of two entities from a characteristic line of the formation, it is considered that the entities in Figure 12 display a V-formation having two outliers on each characteristic line. Then, the algorithm receives the minimum value of Pearson’s coefficient 𝜇_𝑝 which should be met and a maximum percentage of entities (percentageOutliers) which are permitted to be removed from the array of entities. This percentage is calculated with regard to the total number of entities 𝑚.

ALGORITHM: Outlier detection on a characteristic line INPUT:  lineMembers = [𝑒1,𝑒2,…,𝑒𝑚] //Array of 𝑚 entities      𝜇𝑝 //Threshold for Pearson’s coefficient      percentageOutliers //Maximum percentage of outliers permitted on lineMembers OUTPUT:  outliers = [] // Array of outliers BEGIN  (1) Obtain the equation of the line 𝑙 which most suits the positions (𝑥,𝑦) of the entities in lineMembers, see (2).  (2) Get the maximum number of outliers permitted on lineMembers: nbrOutliers = 𝑚∗𝑝𝑒𝑟𝑐𝑒𝑛𝑡𝑎𝑔𝑒𝑂𝑢𝑡𝑙𝑖𝑒𝑟𝑠/100.  (3) Find the nbrOutliers entities in lineMembers which have the maximum distance to 𝑙.  (4) Remove from lineMembers the entities found in Step  3.  (5) Calculate Pearson’s coefficient using positions (𝑥,𝑦) of each entity in lineMembers:     Pearson = PearsonCoefficient(lineMembers).  (6) If Pearson ≥𝜇𝑝 then return in outliers the entities found in Step  3. END

Listing 1 

Outlier detection algorithm.

Figure 12 

A V-formation with two outliers.

Example 1. Consider the set of entities in Figure 13 which would form a straight line if entities 𝑒₂ and 𝑒₄ were not considered; that is, 𝑒₂ and 𝑒₃ are outliers.

Figure 13 

Characteristic line of a V-formation with two outliers.

Input algorithm parameters:

𝑙𝑖𝑛𝑒𝑀𝑒𝑚𝑏𝑒𝑟𝑠=[𝑒₁,𝑒₂,…,𝑒₈], 𝜇_𝑝=0.99 and percentageOutliers = 20%.

Step 1. Equation of 𝑙: 𝑦=−0.754𝑥+2.01.

Step 2. nbrOutliers = ⌈8∗0.2⌉=2.

Step 3. 𝑒₂ and 𝑒₄.

Step 4. lineMembers =[𝑒₁,𝑒₃,𝑒₅,𝑒₆,𝑒₇,𝑒₈].

Step 5. Pearson = PearsonCoefficient (lineMembers) = 1.

Step 6. 1≥0.99; then return outliers = [𝑒₂,𝑒₄].

The algorithm determines that entities 𝑒₂ and 𝑒₄ are outliers. Then, we conclude that entities 𝑒₁,𝑒₃,𝑒₅,𝑒₆,𝑒₇, and 𝑒₈ tend to form a straight line with Pearson’s coefficient greater than 0.99.

3.2. Outliers in Circular Formations

Let 𝐹={𝑒₁,𝑒₂,…,𝑒_𝑛} be a set of moving entities at a time-point 𝑡. Consider an entity 𝑒_𝑖∈𝐹 such that 𝐹-{𝑒_𝑖} is a circular formation and the distance from 𝑒_𝑖 to (𝐶_𝑥,𝐶_𝑦) (center of the minimum circle that encloses the entities in 𝐹-{𝑒_𝑖}) is greater than 𝑅 at 𝑡. This entity may be considered as a member of the formation which is temporally away from it; that is, 𝑒_𝑖 is an outlier. To identify this type of entities, we introduce a parameter 𝑅_{𝑚𝑎𝑥𝑂𝑢𝑡𝑙𝑖𝑒𝑟}, where 𝑅_{𝑚𝑎𝑥𝑂𝑢𝑡𝑙𝑖𝑒𝑟}>𝑅. An entity is considered an outlier at a time-point 𝑡 if its distance 𝑑 to (𝐶_𝑥, 𝐶_𝑦) is greater than 𝑅 and less than 𝑅_{𝑚𝑎𝑥𝑂𝑢𝑡𝑙𝑖𝑒𝑟}; see Figure 14.

Figure 14 

A circular formation with two outliers.

Since the distancing of the outlier entity from the formation is temporal, an analyst may introduce a second parameter 𝑇_{𝑚𝑎𝑥𝑇𝑖𝑚𝑒𝑆𝑒𝑝𝑎𝑟𝑎𝑡𝑖𝑜𝑛} to control the maximum continuous time of permitted distancing. That is, if an entity separates from a circular formation at a time-point 𝑡, then, to be considered an outlier, it will have to reincorporate (𝑑≤𝑅) to the formation before 𝑡+𝑇_{𝑚𝑎𝑥𝑇𝑖𝑚𝑒𝑆𝑒𝑝𝑎𝑟𝑎𝑡𝑖𝑜𝑛}. This same aspect can also be considered for outliers in V-formations.

An analyst can also specify a maximum permitted number of outliers maxNumberOutliers in the formation. This value can be calculated from a percentage (percentageOutliers) with regard to the total number of entities in the formation.

3.3. Outliers in Doughnut Formations

Let 𝐹={𝑒₁,𝑒₂,…,𝑒_𝑛} be a set of moving entities at a time-point 𝑡. Consider an entity 𝑒_𝑖∈𝐹 such that 𝐹-{𝑒_𝑖} is a doughnut formation. For doughnut formations we consider two types of outliers: external and internal. An entity 𝑒_𝑖 is considered an external outlier at 𝑡 if its distance 𝑑 to (𝐶_𝑥,𝐶_𝑦) (center of the minimum circle that encloses the entities in 𝐹-{𝑒_𝑖}) is greater than 𝑅_ext and less than 𝑅_{𝑚𝑎𝑥𝑂𝑢𝑡𝑙𝑖𝑒𝑟}, where 𝑅_{𝑚𝑎𝑥𝑂𝑢𝑡𝑙𝑖𝑒𝑟}>𝑅_ext. On the other hand, 𝑒_𝑖 is considered an internal outlier at 𝑡 if 𝑑 is less than 𝑅_int; see Figure 15.

Figure 15 

Outliers in doughnut formations.

Analogously to circular formations, 𝑇_{𝑚𝑎𝑥𝑇𝑖𝑚𝑒𝑆𝑒𝑝𝑎𝑟𝑎𝑡𝑖𝑜𝑛} and maxNumberOutliers values may be specified, not only for external outliers but also for internal outliers.

4. Experiments and Discussion

4.1. Experiment for V-Formations

For our experiment, we used Nathan’s model [6] implemented in NetLogo in order to generate V-formations. NetLogo is a programmable modeling environment for simulating natural and social phenomena that have been used in previous works [14, 25].

We worked with a population of 15 entities. The dimensions of the Euclidian plane where the entities move were 𝑥∈[−35,35] and 𝑦∈[−35,35], where the unit of measure for distance in NetLogo is a patch. To generate V-formations, we used the parameters in Table 2. These parameters allow us to define aspects related to vision and movement of the entities in Nathan’s model.

Table 2 

Parameters to generate V-formations in NetLogo using Nathan’s model.

For outlier detection, we considered for our model the following parameters: 𝜇_𝑝=0.90 and percentageOutliers = 20%. Our experiment began from tick 40199 and ended at tick 40400, because in Nathan’s model the generation of V-formations requires a stabilization time (tick 40199 in our experiment, a tick is the unit of measure for time in NetLogo). Time-steps were taken in intervals of 20 ticks, for if we consider a smaller step, a change in the position and direction of an entity would be practically imperceptible. Thus, the total number of steps was 10. Figure 16 shows the position of entities at ticks 40200, 40220, 40380, and 40400 and Table 3 summarizes our results.

Table 3 

V-formations: results of the experiment in NetLogo.

(a)

(b)

(c)

(d)

Figure 16 

V-formations. Position of entities at ticks (a) 40200, (b) 40220, (c) 40380, and (d) 40400.

4.1.1. Analysis of V-Formation Results

According to the results shown in Table 3, our model classified several sets of entities as V-formations. These results agree with what was expected since Nathan’s model had already stabilized and V-formations were being generated effectively for those times. Moreover, it was necessary to apply the outlier detection algorithm just in two ticks (at ticks 40380 and 40400, see entities in brackets in Table 3, second column). On the other hand, we found that formations showed a high linearity degree since Pearson’s coefficients in all cases were greater than 0.92.

We also applied Andersson’s model. We considered 𝑚=2 and for the front region of each entity the parameters in Table 2; that is, 𝑠₀=9 patches (vision distance) and 𝛼=1.79 rad (103°) (vision cone). Andersson’s model only identified formations of two and three entities. This is reasonable as explained in Section 2.2 (see Figure 8). In addition, his model does not give information about the shape of the formation that it identifies; even if we considered a wider front region, his model would not indicate that a formation is, for example, V-shaped.

4.2. Experiments for Circular and Doughnut Formations

For experiments with circular and doughnut formations, we also worked in NetLogo and used Wilensky’s model [26] (a model inspired in Reynolds’ work [10]), which generates randomized formations of entities in NetLogo.

We worked with a population of 102 entities. The dimensions of the Euclidian plane where the entities move were 𝑥∈[−35,35] and 𝑦∈[−35,35]. To generate formations, we used the parameters in Table 4.

Table 4 

Parameters to generate formations in NetLogo using Wilensky’s model.

The model was executed on 1200 consecutive ticks (one run) and we conducted an analysis of circular formations and doughnut formations every 400 ticks. We considered that a set of entities showed a circular/doughnut pattern if the set showed a circular/doughnut formation during all the run (1200 ticks). A total of 100 runs were conducted. The parameters used to detect circular and doughnut formations are shown in Tables 5 and 6. Table 7 summarizes our results for the 100 runs. Figure 17 shows the results for one of the runs.

Table 5 

Parameters for circular formation detection in NetLogo.

Table 6 

Parameters for doughnut formation detection in NetLogo.

Table 7 

Circular and doughnut formations: results of the experiments in NetLogo.

(a)

(b)

(c)

(d)

Figure 17 

Circular and doughnut formations. Position of entities at ticks (a) 0, (b) 400, (c) 800, and (d) 1200.

4.2.1. Analysis of Experiments for Circular and Doughnut Formations

According to the results shown in Table 7, the model classified several sets of entities as circular formations and around three remained throughout the 1200 ticks of a run. Regarding doughnut formations, the model detected around two formations that remained during the 1200 ticks of a run. These results agree with what was expected since Wilensky’s model had already stabilized in NetLogo.

With respect to Andersson’s model, we considered 𝑚=3 and the parameters in Table 2. His model identified 752 formations during the 100 runs, a value that is greater than the total number of circular and doughnutformations identified during the 100 runs. This is reasonable because with 𝑚=3 (note that the average number of entities in his formations was four) it is expected that any set of at least three entities that are close to each other (considering 𝑠₀=9 patches) will be identified as a formation. However, as our previous experiments with V-formations, his model is unable to indicate the shape of a formation that it identifies.

5. Conclusions and Future Work

In this paper, we propose two models:(I)a model to identify V-formations with outliers. The model considers the location of entities to determine if they form this type of formation during a time interval;(II)a model to identify circular formations with outliers. The model considers the location of entities to determine if they form this type of formation during a time interval. In addition, we proposed an extension to identify doughnut formations with outliers.The rules for model (I) are flexible since they allow V-formations which are not necessarily perfectly aligned, in accordance with the real world. Furthermore, we consider outliers in V-, circular, and doughnut formations, that is, members of the formation that could be temporarily away from it. Our experimental results in NetLogo showed that our models identified these types of formations in such simulation environment where they were generated. We also showed that when we applied Andersson’s model, his model did not give information about the shape of the formation that it identified; that is, his model suffers from lack of semantic information about shape formations.

Regarding future work, we plan to conduct a series of experiments in the stock market where V-formations usually appear [13]. Moreover, we plan to extend our models to identify isolated entities, that is, entities that even if they are considered members of a formation, they follow their own path and do not influence the path of other entities [27]. We also plan to propose models to identify patterns such as convergence, that is, a set of entities that converge or approach a place; divergence, that is, a set of entities that disperse or move away from a place [28], and self-organization, that is, a set of entities that move as a formation without there being a leader or an entity guiding the rest [29] or that this leader is unknown to the members [30–32].

Conflict of Interests

The authors declare that there is no conflict of interests regarding the publication of this paper.

Acknowledgment

This paper presents preliminary results of the project “Apoyo al Grupo de Sistemas Inteligentes Web-SINTELWEB” with Quipú code 205010011129, developed at the Universidad Nacional de Colombia, Sede Medellín.