Fruit Foraging Flies: Lagrangian and Eulerian Descriptions of Insect Swarming
Joseph D. Majkut
Andrew Bernoff, Advisor
Chad Topaz, UCLA, Reader
May, 2006
Department of Mathematics
Copyright c 2006 Joseph D. Majkut. The author grants Harvey Mudd College the nonexclusive right to make this work available for noncommercial, educational purposes, provided that this copyright statement appears on the reproduced materials and notice is given that the copying is by permission of the author. To disseminate otherwise or to republish requires written permission from the author.
Abstract
In this work, I seek to model swarms of fruit flies, drosophila melanogaster, whose flights are characterized by straight flight segments interrupted by rapid turns called saccades. These flights are reminiscent of Levy-distributed random walks which are known to lead to efficient search behavior. I build two types of model for swarms of foraging fruit flies, whose behavior depends on swarm density and chemoattractant concentration, using rules inspired by experimentally observed flight patterns. First I will present a Lagrangian model where the path of each individual fly is tracked. I will also consider an Eulerian model where the fruit fly density evolves as a function of time and position in space. I will discuss the advantages and disadvantages of the two models and the relationship between them.
Contents
Abstract Acknowledgments 1 Introduction 1.1 Swarm Modeling . . . . . . . . . . . . . . . . . . . . . . . . . Biological Background 2.1 Collective Action . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Free Flight . . . . . . . . . . . . . . . . . . . . . . . . . . . . . L vy Flights and Searches e 3.1 The L vy Flight . . . . . . . . . . . . . . . . . . . . . . . . . . e 3.2 L vy Searches . . . . . . . . . . . . . . . . . . . . . . . . . . . e Lagrangian to Eulerian Formulations 4.1 A Simple Case . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Extension to Density-Dependent Behaviors . . . . . . . . . . A New Direction 5.1 Pursuit of a 1-D Lagrangian Model . . . . . 5.2 Search Efficiencies of the 1-D Schemes . . . 5.3 The One Dimensional Eulerian Framework 5.4 Lagrangian Modeling in Two Dimensions . Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii ix 1 2 5 5 7 11 11 12 17 17 19 21 21 30 34 38 41 43
2
3
4
5
6
Bibliography
List of Figures
2.1 2.2 2.3 3.1 5.1 5.2 5.3 5.4 5.5 5.6 5.7 5.8 5.9 5.10 5.11 5.12 Diagram of Two-phase Foraging . . . . . . . . . . . . . . . . A Characteristic Flight Path for a Fruit Fly . . . . . . . . . . . A Flight Path with an Odor Source . . . . . . . . . . . . . . . A L vy Flight in Two Dimensions . . . . . . . . . . . . . . . . e Moments for 1-D Model . . . . . . . . . . Position Histograms for 1-D Swarm . . . Moments for Foraging Model . . . . . . . Position Histograms for Foraging Model . Swarm Moments for Interacting Model . End of Simulation Positions . . . . . . . . Position Histograms for Interaction . . . . Search Efficiencies I . . . . . . . . . . . . . Search Efficiencies II . . . . . . . . . . . . Fluxes into and out of a Rectangle . . . . The 2-D Flight Path of a Foraging Fly . . . A Radial Histogram of Position . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 8 9 13 23 24 25 26 29 30 31 32 33 35 39 40
Acknowledgments
Professors Andrew Bernoff and Chad Topaz have been invaluable in the preparation of this work. Their ideas, insights and assistance carried it from its beginnings. Likewise, the guidance of Professors Lesley Ward and Darryl Yong was instructive as to how to work independently. Foremost, my fellow math majors deserve recognition for their talent and character, which have been the most inspiring elements in my undertaking of this project. The discussions and humor I shared with them were priceless. Claire Connelly, for the computer support and the HMC Math Thesis A X class, I thank you on behalf of all of my colleagues. L TE
Chapter 1
Introduction
The aggregation of fruit flies, Drosophila melanogaster, about a source of food is an everyday, albeit undesirable, phenomena. Biologically, fruit flies gather in groups to engage in a host of social behaviors. One of these is foraging for food sources in areas where food is not readily available to each individual, but rather must be found by the group. The idea that the group can search for food more effectively than the individuals can is fairly simple to grasp, but the behavior that has been observed leads to interesting mathematical questions. In the swarm, an individual's flight path is characterized by straight jaunts followed by quick and random changes in direction, known as saccades. The flight path of a fruit fly is essentially a random walk in the two horizontal directions, with vertical behavior much less interesting. When a fly forages, the length of the straight jaunts is controlled by the concentration of scent (chemoattractant) that the fly experiences (see Frye and Dickinson (2004)). Experiments indicate that the fly has no memory with regard to scent and cannot voluntarily move up a chemoattractant gradient, by biasing its saccade direction. In addition, Drosophila swarms are observed to act cooperatively in two distinct modes (see Tinette et al. (2004)). Certain flies in the group will travel away from the greater swarm, searching for and evaluating potential food sources, while the others stay in the group and follow the successful searchers. Interestingly, flies that participate in searching are biologically no different than the grouped flies and a single fly's mode of behavior will vary with time. With these factors in mind, a fruit fly swarm poses several questions: 1. If an actor is searching, how can it locate the source of an attractant without being able to consciously move toward the source?
2
Introduction 2. How do actors, with fly-like behavior, perform searches most efficiently and how does that relate to the observed biology? 3. How is it that biologically identical flies can demonstrate two modes of behavior, presumably based on an individual's preference, such that a swarm cooperates in foraging as opposed to competing or acting individualistically? These are the motivating questions that I have chosen to guide the work in this thesis. They turn out to be very rich mathematically, requiring concepts from several fields to be used to model, analytically and computationally, a typical fruit fly swarm. Many of the techniques I will share in the following pages come from previous studies of insect or animal aggregation which I will heuristically describe before delving into fruit fly behavior.
1.1
Swarm Modeling
Aggregation and collective movement abound in nature. A flock of birds headed South in November, a school of fish in the ocean, and a swarm of insects crowded onto a piece of fruit are all examples of aggregation and grouping behavior. The reasons for aggregation are many: protection from predators or the environment, mating, traveling, and foraging are all reasons for animals to form into groups and perform some collective action or simply interact socially amongst themselves. The kinds of behaviors that a swarm can exhibit can be as diverse as the organisms that form them. The structure, movement and lifetime of a swarm can vary because of the local environment, the constituent organisms, or the reasons for the social interaction. How these larger behaviors come from a set of organisms is an interesting question that has long been considered by biologists but has, in the last half century, peaked the interest of mathematicians. The mathematics of swarms can pose hard problems, with non-linearity and large parameter spaces, but swarming does lead to interesting findings regarding how local dynamics can affect global structure and how biological systems can display essentially optimized behavior. Swarms have been considered from a modeling perspective in two primary fashions. Lagrangian swarming models describe a swarm at the level of the individual, where an individuals' behavior is based on the local environment, and then conclusions can be drawn about the swarm through numerical simulations or estimates on the structure of the swarm after long times. The work by Mogilner et al. (2003) is an example of this approach.
Swarm Modeling 3 They investigate the social forcings commonly used to model swarms and look for conditions on swarm stability. The other tool used to describe swarms is known as the Eulerian framework. Swarms, in Eulerian models, are thought of in a continuum sense - where equations similar to those found in fluid mechanics describe some attribute of the swarm, typically the density of individuals. These governing equations, are typically nonlinear partial differential equations and hence must be solved numerically. Topaz and Bertozzi (2004) provide an excellent example of this approach. Some studies start with a Lagrangian framework and then use it to build an Eulerian model using limiting cases and simplify assumptions. This is an interesting approach and one which I would like to use in my work on fruit flies. The behavior that I will seek to model for Drosophila is similar to the behavior exhibited by the bacteria, Escherichia coli, which is highly studied. Like the flies' jaunts and saccades, bacteria's trajectories are characterized by runs and tumbles. A run is a path swam in a straight line and a tumble is a random change in direction. Berg (1983) gives a very nice mechanical description of the mechanics of the behavior. It is essentially a random walk in three dimensions, where the time between tumbles is a Poisson random variable, governed by the direction that the bacteria is traveling. Bearon and Pedley (2000) argue that mean run time increases when a bacterium is moving up a chemoattractant gradient, a behavior known as chemotaxis. While chemotaxis behavior is not directly observed in flies, the modeling studies performed by Bearon and Pedley (2000), Othmer and Stevens (1997), and Grunbaum (2000) use techniques that I will emulate in my study of fruit fly foraging.
Chapter 2
Biological Background
An investigation into the foraging behavior of a swarm of fruit flies first requires a strong understanding of two elements of fruit fly behavior: social interaction and free flight characteristics. In this chapter, I will explain some of the Biology necessary to build a meaningful model of Drosophila swarming.
2.1
Collective Action
In the wild, it is easy to observe a group of fruit flies on a food source. The question is, however, is there much of a biological justification for the consideration of collective behavior? This question is taken up in Tinette et al. (2004), who observed collective action in swarms of Drosophila. Flies in their experiments did not individually assess all available food sources, but rather, food sources were located and tested by a few members of the swarm, and then the group aggregated on a favorable source. The experiments in Tinette et al. (2004) were carried out in a clear box, where the location, strength, and quality of food sources could be changed as well as the amount and species composition of the flies used in the test. The food sources were contained in small glass containers, such that the experimenters could count the number of flies accessing a food source in any given trial. They found differences in the aggregation behaviors of certain mutant strains, but we need not be concerned with those differences in this project. Instead, I will focus on the general observations of the fruit fly.
6
Biological Background
2.1.1
Search-Aggregation Cycles
Tinette et al. (2004) observed two phases of fly behavior, the so-called primer and follower modes. These behaviors are disparate, and can be defined as follows: Primer: A 'primer' fly goes in search of favorable food sources, testing various possibilities. Eventually, a primer fly will settle on a preferable source. Follower: A following fly does not search for or evaluate possible food sources, but is visually attracted to other flies, particularly aggregations. Flies Exhibiting the two Foraging Behaviors
Potential Food Sources
"primers" "followers"
Figure 2.1: This figure shows my interpretation of the behavior discussed by Tinette et al. (2004). Biology suggests that these behaviors rely on different sensory systems for information. Tinette et al. (2004) indicate that the 'follower' behavior is vision-based, as flies demonstrate a powerful visual responses while in free flight. While searching behavior is partially dependent on visual cues for orientation purposes (see Frye et al. (2003)) it is primarily based on scents, known as chemoattractants.
Free Flight 7
2.1.2
Conditions for Collective Action
Fruit flies require certain conditions to participate in collective action as was observed by Tinette et al. (2004). Two that are interesting from the perspective of modeling are the dependence on the number of flies in the swarm and the saturation of the domain to be searched with scents from food sources. Tinnete et al. observed that the flies did not utilize the searchaggregation cycle when the number of flies was extremely low; with four flies in the test chamber there was no distinguishing between the flies' behavior. The search-aggregation cycle, however, was clearly observed with 50 flies in the chamber, and most experiments were carried out for groups of 200 flies. The other condition of interest for the initiation of searchaggregation cycles was the presence of a scent gradient in the environment. When the environment was heavily saturated with chemoattractant, the flies did not participate in search-aggregation, but were inactive. The search-aggregation cycles were seen in environments where point sources created gradients in chemoattractant. The search-aggregation method of foraging for food would presumably be advantageous because it requires less energy output from the swarm, as a few flies do the majority of the searching. In this thesis, I hope to gain a firm understanding of this behavior and some quantitative idea of its efficiency with regard to the energy used by the swarm during foraging. The modeling of a swarm, however, begins with the free flight characteristics of an individual.
2.2
Free Flight
Free flight by Drosophila has been the focus of quite a bit of experimental work of late, as researchers have tried to analyze the aerodynamics, energetics, control, and behavioral characteristics of fruit fly locomotion. These studies have motivated the creation of novel experimental techniques and interesting analysis. They have been particularly successful in recording the behavioral characteristics of Drosophila in free flight. Fruit fly flight is characterized by a series of straight segments interrupted by rapid changes in horizontal heading known as saccades. A fly completes a saccade in 50 - 100 milli-seconds, making the corresponding angular velocity is in excess of 1000 degrees per second (see Frye and Dickinson (2004)). A representative flight path is shown in Fig. 2.2. A fruit fly will generally turn through approximately 90 degrees during a saccade.
8
Biological Background The direction that a fly turns (right or left) is seemingly at random, unless it is performing a collision avoiding maneuver. Flies tend to turn away from potential collisions. For instance, if a fly is approaching a wall on the left side, it will favor a saccade to the right. The length of time between saccades and correspondingly the length of the straight flight segments is also random, but can be influenced by environmental factors. A Saccading Flight Path
90
Figure 2.2: A sample of a flight of varying segment lengths interrupted by saccades. Note that turns are nearly 90 degrees at every saccade and that segment lengths are not constant. Frye and Dickinson (2004) observed that saccade angle in flies was normally distributed about 90 and -90 degrees when the fly was not performing a collision avoidance turn.
2.2.1
Odor Localization
Flies, like many animals, utilize scent to locate food sources in their vicinity. Odor localization by a fruit fly in free flight is important to understand, as it is the behavior that a model will have to consider for a fly in 'primer' mode. The flight path of a fruit fly participating in odor localization is still characterized by saccades of 90 degrees, but in the vicinity of a chemoat-
Free Flight 9 A Flight Path Near a Food Source
Figure 2.3: A free flight path similar to those exhibited by flies in the vicinity of an odor source (red dot). Note that saccades tend to occur much more often in the vicinity of the odor source. This behavior is well observed in Drosophila. The end result is that a fruit fly in the vicinity of a food source will spend much more time close to that source than far away. Eventually, the source will enter the fly's radius of detection and be found. tractant saccades occur more often and with smaller distances in between successive redirections. This was observed Frye et al. (2003), using the same techniques that were developed to track normal free flight and introducing a hidden source of attractive scents in the floor of the test chamber. Fig 2.2.1 is an example of a free flight with saccading influenced by proximity to a food source, note that the fly spends much more time in the vicinity of the food source than in the far-field. I will suppose that flies exhibit similar behavior when in trying to swarm and in areas of high density.
10
Biological Background
2.2.2
Activity in Drosophila
The distributions of segment lengths in between saccades is not well studied. There is, however, some biological backing that says fruit fly activity demonstrates scalability in time. This means that descriptions of fruit fly activity look very similar at all scales. The experiment that led to the scalability conclusion is described in Cole (1995). Scalable activity can lead to flight lengths which are not distributed normally, but rather follow a power law distribution, so-called L vy Flights, the properties of which will be exe amined in the following chapter.
Chapter 3
L vy Flights and Searches e
L vy walks are a class of random walk, which are often found in biological e behavior and are prevalent in foraging. I will use L vy Flights to model the e foraging behavior of fruit flies, as there is a significant biological justification to so and the behavior will correspond with the qualitative behavior descriptions from the past chapter.
3.1
The L vy Flight e
L vy Flights, in this context, will be continuous time random walks in one e dimension, where segment length between direction reversals, x, is a random variable with the probability distribution function: p( x ) = 0 cx - if x < xmin if x xmin , (3.1)
where 1 < 3 and xmin > 0. The lower bound on and the minimum segment length, xmin , make the distribution normalizable, and hence a valid power-law probability density function (pdf). The parameter c is the normalizing constant and can be found from integration. For values of > 3, the Central Limit Theorem says that over many iterations, the distribution of flight lengths will be normal. For 1 < 3, however, the Central Limit Theorem (CLT) is not valid and the distribution is upset by rare but extraordinarily long segment lengths. One of the requirements of the Central Limit Theorem is that the random variables have a finite second moment, according to Feller (1968). The Central Limit Theorem does not apply to random variables with power-law distributions with 3
12
L vy Flights and Searches e because they have infinite second moment, as follows: E[ x2 ] =
xmax xmin
x2 p( x )dx cx (2-) dx if > 3 if < 3.
= =
(3- ) c 3- xmin
The infinite second moment of L vy flights is the source of much of their e interesting behavior. The L vy flight formulation is not hard to extend to e two dimensions, where coefficients and the limits of exponents are different, but the behavior's properties remain. A characteristic L vy flight in e two dimensions is shown in Fig. 3.1. L vy flights give rise to counter-intuitive behavior, coming from the e breakdown of the CLT. For instance, while processes which meet the hypothesis of the CLT lead to variance that scales linearly with time, variances in processes that display L vy flights have other time-scaling behaviors, e [Weeks et al. (1996)]: t2 if 1 < < 2, 2 (4-) if 2 < < 3, (t) t t if > 3. These scaling laws are in the regime of anomalous diffusion, where variance does not scale with time linearly, but with some power of time. In the case of L vy flights, the transport is super-diffusive. The occasional e very long flights undertaken by the walker cause it to move further from the origin, or other beings in a swarm, faster than a walker undergoing Brownian motion. This makes L vy flights an ideal governing behavior for e foraging for scarce resources.
3.2
L vy Searches e
Random walks are a common way to conceptualize searching behaviors in biology and other fields and L vy flights are included in the classes of e random searches that have been investigated. In fact, L vy flights can be e shown to be an ideal searching strategy for domains where resources are scarcely distributed.
L vy Searches e
13
Figure 3.1: A L vy flight in two dimensions. Note that the flight path has a e scalable nature, where the flights of similar length are clustered with very long flights separating them. If the image was viewed at a scale much further out, the entire path currently visible would be a single cluster within a much larger flight path. Given the scalability of Drosophila behavior, it is believed that fruit fly flight paths are actually L vy in nature, as opposed e to being distributed normally.
14
L vy Flights and Searches e
3.2.1
Unbiased Searches
Viswanathan et al. (1999) show that a random search governed by an inverse power law is optimal for values of which lead to L vy flights. The e argument is easy to see in one dimension. They let the segment lengths in a random walk be defined as in eq. 3.1, then a searcher behaves with the following rules: 1. The searcher begins a search by selecting a path direction randomly and a segment length from eq. 3.1. 2. If there is ever a target within a distance rv > 0, then the searcher proceeds directly to the 'found' target. This is the radius of detection about the searcher. 3. If there is no target within rv of the searcher, then the searcher proceeds. 4. If the searcher travels the segment length selected in (1) without encountering a target, then it selects a new direction of travel and a new segment length randomly. Viswanathan et al. (1999) then define the search efficiency () by () = 1 . <l>N
There, < l > is the mean segment length and N is the mean number of flights taken by the searcher from one target to the next. They then distinguish two kinds of searches, destructive and nondestructive. Targets in destructive searches are destroyed after being visited by the searcher. Destructive searches are characterized by a larger mean number of inter-target steps, Nd (/rv )-1 , where non-destructive searches are characterized by, Nn (/rv )
-1 2
.
The parameter is the mean free path of the searcher between targets, but can be thought of as the mean distance between target sites. Nd > Nn because a searcher in a non-destructive search can turn back to a previously visited site, whereas a destructive searcher leaves a site empty, decreasing
L vy Searches e the number of viable targets in the area. Viswanathan et al. (1999) show that for non-destructive searches () is maximized by choosing, = 2- 1 . [ln (/rv )]2 (3.2)
15
Thus when the search is non-destructive, and rv but not known, = 2 is the optimal choice for a random search. Likewise, they show that for a destructive search, (t) has no maximum for allowable values of , but efficiency increases as 1. The reasons that L vy searches are optimal e when resources are scarce can be understood by the fact that the probability of returning to a previously visited site is lower for a L vy searcher than e a Gaussian one, and thus the L vy searcher is more likely to visit new areas e and find more targets. Similar results can be found for searchers who obey L vy flights, but bias them either with a priori information about the dise tribution of walkers, or some environmental factor like a chemoattractant gradient.
3.2.2
Biased Searches - Length
Marthaler et al. (2004) investigate the effectiveness of L vy searches where e segment length is biased on a priori information. This is similar to the behavior that we see in fruit flies, where the animal preforms chemokinesis. In Marthaler et al., they analyzed a non-destructive search in two ways, where 1 is calculated from the ideal search exponent defined by eq. 3.2 and where 2 is calculated according to, 2 = H ( ( x ) - r v ) 1 + when, ( x ) = max( P( x )) + 1 ( x ) + 1
+ 3H (rv - ( x )),
-
| x - z| P(z)dz.
In the above, P( x ) is the probability of a local target at the point x, thus ( x ) is an effective upper bound on the distance from the point x to the nearest target. See that as increases decreases and the forager takes generally longer steps as the density of targets decreases. The argument here is in one dimension, but if x is thought of as a vector then the two-dimensional case is an easy extension. Marthaler et al. (2004) evaluated the efficiencies of the two choices for numerically and found that 2 was much more efficient with respect to . With one thousand targets in a domain 200 units
16
L vy Flights and Searches e long and rv = 1.5, 1 7 and 2 33. The second method is over three times as efficient. This seems odd, as 1 was supposed to be optimal, but with increasing target density (falling ) 1 decreases and leads larger flight lengths, allowing the forager to 'skip' over targets.
3.2.3
Biased Searches - Direction
In the same paper Marthaler et al. (2004), Marthaler et al. also analyze searches where direction is biased by the distribution of sources, similar to an animal which would tend to turn up gradients. This behavior is not observed in fruit flies Frye et al. (2003), but is seen in some kinds of chemotaxis. They performed the same numerical analysis on a searcher with = 2 and the probability of turning left or right dependent on the derivative Px of P( x ), where, .75 ifPx < 0 pl = .50 ifPx = 0 .25 ifPx > 0 and, pr = 1 - p l . They found that the searched with biased direction was much more efficient, with 55.
Chapter 4
Lagrangian to Eulerian Formulations
The flight path of a single fly participating in foraging behavior is not difficult to model (see Chapter 5). The questions then becomes, how can an individual-based model be altered such that it could be applied to an Nbody swarm which will display behaviors similar to those seen experimentally? One route is to numerically simulate the swarm using a Lagrangian framework, but this is computationally intensive and I also want to build an Eulerian description of a swarm. By approaching the problem from both directions, I hope to better understand how the guiding questions for the thesis can be answered and see some interesting mathematics. In this chapter, I present several simple examples of ways by which an N-body system with certain rules for interaction can be modeled at the individual level, tracking position and velocity, and then described using a Eulerian PDE, focused on density and velocity fields.
4.1
A Simple Case
The process of relating a local model of a random walk for every member of a swarm to a PDE for the entire swarm is described by Othmer et al. (1988). They let a particle move along the x-axis with constant speed v, and suppose that the particle randomly changes direction at intervals governed by a Poisson process with rate . If pr ( x, t) and pl ( x, t) are the probability density of particles that are at ( x, t) and are moving to the right and left,
18
Lagrangian to Eulerian Formulations respectively. These satisfy the system, pr pr +s = -pr + pl t x pl p - s l = pr - pl t x and their sum is the probability that a particle is at ( x, t). They then define the probability flux as, J s( pl - pr ). Then, by taking the proper derivatives, p J + =0 t x p J + 2J = -s2 . t x Where p( x, 0) and J ( x, 0) are defined by pr ( x, 0) and pl ( x, 0). By differentiation and substitution, these two can be combined to the hyperbolic system, 2 p p 2 p + 2 = s2 2 t2 t x with the initial conditions, p j0 ( x, 0) = - ( x ). t x Eq. 4.1 is the Telegrapher's equation. The mean squared displacement for the system can be found by assuming that p( x, t) and its first two derivatives approach zero at | x | . Then with p( x, 0) = p0 ( x ) and (4.1)
< x2 >=
it holds that
-
x2 p( x )dx,
d2 < x 2 > d < x2 > + 2 = 2s2 , dt2 dt with homogeneous initial conditions. Thus,
< x2 (t) >=
Thus,
s2
t-
1 (1 - e-2 ) . 2 for small t for large t.
< x2 (t) >
s2 t2 s2 t
In this example a behavior, which is simple at the individual level, leads to a PDE for the density of the swarm. The solution to that PDE looks like solutions to the diffusion equation for large values of t, and the swarm disperses over the real line.
Extension to Density-Dependent Behaviors 19
4.2
Extension to Density-Dependent Behaviors
The model presented for insect swarming in the above example is overly simplistic to well-describe a swarm. Most notably, it lacks any sort of interaction between swarm constituents. Grunbaum (1994) approaches interaction, translating a Lagrangian model into an Eulerian one. He assumes that at any given observation, individuals are distributed as a set of Poisson points according to the density distribution. In any spatial interval, under this assumption, the number of individuals is a Poisson distributed random variable.
4.2.1
Lagrangian Framework
Consider a one-dimensional swarm on the x-axis, where individual behavior is governed by two parameters, and , the sensing distance and target density. Thus during the social behavior of an animal, the intervals of length on either side of the animal are tested for other individuals, and individuals move toward areas of density . If i is the number of individuals in a spatial interval di (length ) then the local density for an animal at the point x is, 1 + 1 + 2 ( x ) = . 2 Grubaum then creates a time dependent social forcing which is evaluated at random intervals at rate s , governed by, f 1 (1 , 2 ) = sgn(2 - 1 - 1 - 2 ) f 2 (2 , 2 ) = sgn(1 - 2 ) f s (1 , 2 ) = f 1 (1 , 2 ) f 2 (1 , 2 ). In addition, individuals are subject to a random Gaussian force, f r , with mean 0 and variance 2 at random intervals with rate, r . Individuals respond to the forcing, F (t) = f s + f r according to Newton's Law, m d2 x dx + d ( - U ) = F ( t ), 2 dt dt (4.3) (4.2)
where d is the drag co-efficient and U is advection due to an external flow field.
20
Lagrangian to Eulerian Formulations
4.2.2
Eulerian Translation
Grunbaum (1994) defines a velocity autocorrelation and a power spectrum to derive the form of the Eulerian model. The velocity autocorrelation, Rv ( ) ( is the non-dimensional time interval), and power spectrum, Sv ( ) are related by the Fourier transform. By transforming eq. 4.3 and applying various substitutions using the parameters of the swarm, one comes to the limiting expression, Rv ( ) = De-| | + (u + ( x, t))2 , (4.4)
where D is the diffusion coefficient for the expected density flux due to the random forcing, is the characteristic velocity of the swarm, and ( x, t) is the expected value of eq. 4.2, which takes the form, ( x, t) =
1 =0 2 =0
c(1 , 2 ) f ( x, t, 1 , 2 ).
(4.5)
Here, c(1 , 2 ) is the joint distribution for (1 , 2 ) about x at time t. The assumption that the distribution of individuals in the swarm is a distribution of Poisson points says that 1 and 2 are independent, such that, c(1 , 2 ) = c1 (1 )c2 (2 ), (4.6)
where cm (k ) is the probability that there will be k individuals in the interval dm with respect to x, k c m ( k ) = e -m m , (4.7) k! with, m ( x, t) = ( x , t)dx . (4.8)
dm
Then, by eq. 4.4, the PDE that determines the density flux of individuals is 2 = D 2 - [(U + ( x, t))], t x x (4.9)
which becomes a nonlinear partial integro-differential equation upon substitution of eqs. 4.5, 4.6, and 4.7 4.8. Grunbaum (1994) evaluates solutions to eq. 4.9 numerically, and finds that the solutions to the Eulerian formulation positively compare with simulations of the swarm from the Lagrangian perspective, which serves as his verification of the Poisson point assumption.
Chapter 5
A New Direction
5.1 Pursuit of a 1-D Lagrangian Model
A fairly simple Lagrangian model of fruit flies foraging in one spatial dimension can tell us quite a bit about the swarming of fruit flies and the search efficiency of a swarm based on its behavior. I have built such a model, which I will describe in the following sections. The model that I have implemented describes the positions of N 'bugs' by their positions on the line, X, and a count-down clock for each, denoted by, . In this model, a bug at position, xi , will saccade if and only if i = 0. Upon saccade, i will then bee reset according to some function, which incorporates the interaction with other flies and attraction to chemoattractants into the model. The condition that flies only saccade according to their internal timer means that the bugs only make decisions while saccading, which is not entirely physical, but is computationally easier than evaluating the fitness of the environment for each fly at each time-step. I will show later that this approach of evaluating the environment only during saccades results in similar results to more physical, integrative, models and biological experiements. In this model, the flies move with a constant speed, c = 1, and the time steps are normalized, such that the entries in X and are always integer valued. Let xin+1 denote the position of bug i at time, t = n + 1, and similarly represent the internal clock, in+1 , then the movement of the flies can be expressed by, xin+1 = xin + ci , in+1 = in + Fin . In the above system, ci can take on the values ci = {-1, 1, 0}, which corre-
22
A New Direction spond to moving left, right and staying in place, it's value is a random variable set during the last saccade. The random variable, Fi , is used to change the internal clock of bug i and has a large affect on individuals' behavior. Fi is a random variable whose value also depends on the last saccade by the conditions, P{ Fin = -1|in > 0} = P{ Fin 1 f i ( X, 0 ), (5.1) (5.2)
=
0 |in
= 0} =
where 0 > 0. The probability density function, f i ( X, 0 ), is the piece of the model by which it is easiest to change the behavior of the swarm, as it, along with ci , determine the swarm constiuents' response to their environment and each other. In this section, I will develop a one dimensional model of fruit fly behavior, beginning with a group of L vy walkers. e
5.1.1
Base Case : L vy Walks e
As a validation of the model, we can look at the behavior of the swarm when f i ( X, 0 ) takes the form of a L vy distributed random variable, e f i (0 ) = c0 .
-
(5.3)
In this case we would expect to see the behavior discussed in Chapter 3. Figure 5.1 shows the mean, variance, skew, and kurtosis for three swarm simulations. These reflect the behavior we would expect to see from a swarm of non-interacting L vy walkers in one dimension. The mean and e skew are both near zero for each of the runs, which indicates no directional bias in saccade direction or length. The variance scales with a power of t. According to Weeks et al. (1996), one would expect the variances to scale like t1.68 , where the proportionality constant would be 2.5. Lastly, the kurtosis lines for the different simulations all indicate that the long term distribution of the swarm is leptokurtic (kurtosis is greater than 3), which essentially means that the distribution has a high point and long tails. Figure 5.2 shows the distribution at 4 time steps for the simulation in blue in Fig. 5.1.
5.1.2
The Next Step : L vy Walkers and Attractors e
Now consider the flies to have a run-and-tumble like behavior, where they are more likely to assume lower values of 0 upon saccade while in some
Pursuit of a 1-D Lagrangian Model 23 Swarm Moments for 3 L vy Walking Swarms e
Mean Position
30 20 10 0 10 20 30 0 1000 2000 3000 4000 2 1 0 5 4 3 x 10
5
Variance
0
1000
2000
3000
4000
Time Skew
1 0.5 0
Time Kurtosis
10 8 6 4
0.5 1 0
2 0
1000
2000
3000
4000
0
1000
2000
3000
4000
Time
Time
Figure 5.1: This figure shows the [left to right ; top to bottom] sample mean, variance, skew and kurtosis for three iterations of the model, where N = 1000 and = 2.32. In these simulations, the initial condition was delta-like, with all of the bugs at the origin, and the simulations were run for 4000 time steps. vicinity, rv , of a food source. The Fi will now depend on both 0 and xsink , the distance to the nearest food source, by, Fi =
xsink rv 0 ,
if xsink rv
0 , otherwise,
where the p.d.f for 0 remains as it was in eq. 5.3. I chose this function for Fi because it allows for the flies to become more 'excited' about environments closer to an attractant and to not saccade once they reach the food source.
24
A New Direction
Histograms of a L vy Walking Swarm e
T = 1000
80 60 40 40 20 10 0 -1000 -500 0 500 1000 0 -2000 -1000 0 1000 2000 30 20 60 50
T = 2000
x
T = 3000
80 60 40 20 0 80 60 40 20 0 -4000
x
T = 4000
-2000 -1000
x
0
1000
2000
-2000
x
0
2000
4000
Figure 5.2: This figure shows 100-bin histograms of bug position for the blue simulation in Fig. 5.1 for 4 time steps. The delta-like initial condition would appear as a bar of height 1000 at the origin and is not pictured. Note both the heavy-tailed nature of the distribution and the height of the enter in the T = 4000 frame
Pursuit of a 1-D Lagrangian Model 25 FIgure 5.3 shows the moments for three swarm simulations with attractants, along with a reference case from the last section. In these cases, the overall swarm behavior is similar to that from the first case. The mean poSwarm Moments for L vy Walkers with an Attractor e
Mean Position
30 20 10 0 -10 -20 -30 0 1000 2000 3000 4000 2 1 0 5 4 3 x 10
5
Variance
0
1000
2000
3000
4000
Time Skew
1 0.5 8 0 -0.5 2 -1 0 1000 2000 3000 4000 0 0 1000 6 4 12 10
Time Kurtosis
2000
3000
4000
Time
Time
Figure 5.3: This figure shows the swarm moments for four simulations. The black line represents the histogrammed simulation from the last section. The three colored lines, blue, red, and green correspond to simulations with sinks at x = {-300, 100, 1000}, respectively. In those simulations, N = 1000, = 2.32, and rv = 50. Again, the initial condition was X = 0 and time was stepped 4000 times. sition of the swarm, and the skewness of their distribution, are both close to zero. The variances for the swarms is scaling appropriately for L vy walke ers and despite the fact that saccades made within rv of the food sources are characterized by decreased length the swarms' behavior remains superdiffusive. Lastly, the kurtosis of each simulation is what would be expected.
26
A New Direction Long-Time Swarm Histograms for 2 Simulations with Sinks
140 40 35 30 25 80 20 60 15 40 10 5 0 120
100
20
0 -4000
-3000
-2000
-1000
x
0
1000
2000
3000
4000
5
4
3
2
1
x
0
1
2
3
4 x 10
4
5
Figure 5.4: This figure shows histograms for simulations where the sink was loacated at x = 100 (left) and x = 1000 (right). The distribution on the left was taken at T = 4000, the final time step shown in Fig. 5.3. It clearly shows a large population of bugs near the attractant. Likewise, the distribution at the left shows a high concentration of bugs about the sink. This distribution was binned at T = 5 x 10-4 . Note that the kurtosis for the red simulation (sink at x = 100) is higher, probably because of a large number of bugs near the sink. Since that simulation has the sink closest to the origin, it is expected that more bugs will fall into the attractor over a similar amount of time. A high concentration of near x = 100 also explains the slightly lower variance seen in this second case. Figure 5.4 validates that there is a large grouping about the sink, when x = 100, at later times. It also shows that a similar grouping will occur when the sink is far from the origin, x = 1000, but that it takes much more time to develop. The time lag makes the x = 1000 and - 300 cases look similar to the regular random walk with respect to the overall swarm characteristics. Interaction terms, however, will change the largescale swarm dynamics more significantly.
5.1.3
An Appropriate Model : Directionally Interacting L vy Walke ers with Sinks
The last model had flies gather in some number about food sources, but excluded what we saw in Chapter 2 regarding the interaction between flies and cooperative foraging. We seek a model that will combine the attraction that we implemented in the last section with a way for flies to choose the direction of their saccades based on the positions of the other members of the swarm and their path length based on the local swarm density. In order to
Pursuit of a 1-D Lagrangian Model 27 do this, I calculate a version of the Morse potential, usually seen in potential energies of diatomic molecules, as a measure of the force the swarm has on an individual. My version of the Morse potential is directional meaning I calculate it on the left and right, by Ml = Mr =
e-|x -x |
i j
j
e-|x -x | .
i k
k
In the above equations, j is the set of indexes for bugs to the left of the bug whose potential is being calulated and k indexes the bugs to the right. The total Morse potential is calculated by, M0 = Ml + Mr . The use of exponential forces, such as the Morse potential that I have defined above, is a common technique in swarm modeling. Mogilner et al. (2003) use an exponential forcing and investigate its effect on the swarm shape and cohesiveness. Usually, the pairwise interaction is calculated with another term, with a different sign in the exponent, so that a pair can have a "comfortable" distance for which their pairwise potential energy is zero. I do not use this technique here, because it would force a non-physical interaction between bugs which are close on the line. This interaction is usually imposed to model collision avoidance, but since the bugs in this model must collide, we ignore the repulsive term and calculate the attractive one. This is the first model in this chapter that uses the model state to influence the direction that flies turn while in the swarm. Thus there are three behaviors that the flies in this model must exhibit while saccading, beyond those of the standard random walker. 1. When a fly is within rv > 0 of a food source, then the length of flight segments will depend linearly on the distance between that fly and the food source. 2. A fly should choose the direction which exhibits the higher Morse Potential most of the time. 3. A fly will limit it's flight length when it's total Morse Potential is higher than some minimum value, to encourage 'packets' of flies. The model's parameter, Fi can be written as follows, Fi = g( X )h( xsink )0 , for = 0.
28
A New Direction In this formulation, g( X ) gives the flight length dependence on the total morse potential by a step function g( X ) = where, Mi = 1/10, if Mi > M0 1, otherwise
e-|x -x | .
i j
j =i
Then I use the same function for h as we did in the last section, such that. h( xsink ) =
xsink rv ,
if xsink rv
1, otherwise
where xsink is the distance to the nearest food source. 0 is still distributed with the inverse power law that defines a L vy walk. Lastly, ci depends on e X by the following: P{ci = 1| Mi,+ > Mi,- } = 0.9 P{ci = 1| Mi,+ = Mi,- } = 0.5 P{ci = 1| Mi,+ < Mi,- } = 0.1
and P { c i = -1} = 1 - P { c i = 1}. This formulation allows the two dependencies, to interact and keep flies in highly desirable areas. I decided to make the dependence on the left and right Morse potentials ( Mi,+ and Mi,- ) non-deterministic to perhaps better reflect the behavior one might actually see in a swarm. Fig. 5.5 shows the moments for simulations with various values of M0 . In this figure, there are several interesting differences in the swarm moments that were not observed in the previous cases. Chiefly, in this set of simulations, we see a large difference in the swarm variance for different values of M0 . For low values of M0 , the swarm remains cohesive through time, and has half the variance of both the large M0 simulations and the previous cases. This cohesiveness indicates that the swarm spends much of the time near the critical values of M0 and a large proportion of the flight lengths are smaller. Figure 5.6 shows the positions of the bugs at the final time for the M0 = 5 and 200 simulations. The histograms of position and the swarm
Pursuit of a 1-D Lagrangian Model 29
Swarm Moments for the Interacting Swarm
Mean Position
60 40 20 0 -20 0 8 6 4 2 0 x 10
5
Variance
1000
2000
3000
4000
0
1000
2000
3000
4000
Time Skew
1.5 1 0.5 10 0 -0.5 -1 0 5 0 20 15
Time Kurtosis
1000
2000
3000
4000
0
1000
2000
3000
4000
Time
Time
Figure 5.5: The swarm properties for the directionally interacting swarm are shown in this figure. In those simulations, a sink was located at x = 100 and M0 = {5, 8, 25, 200}, corresponding to the blue, red, green and black lines, respectively. As before, N = 1000 and time was recorded for 4000 steps.
30
A New Direction
M 0= 5
500 450 400 350 300 250 200 150 100 50 0 -4000
M 0= 200
500 450 400 350 300 250 200 150 100 50 0 -4000
-3000
-2000
-1000
0
1000
2000
3000
4000
-3000
-2000
-1000
0
1000
2000
3000
4000
x
x
Figure 5.6: Historgrams of the final positions of the bugs for simulations where M0 = 5 (left) and 200 right. Note that the simulation with the smaller critical value appears to be much more effective, with a larger grouping right about the attractant. moments show that there is interesting behavior going on within these swarms. Figure 5.7 shows the evolution of a swarm operating under the full set of behavioral rules, where the bin with the highest number of flies moves from around the origin to the area around the sink. The time series for simulations with different values of M0 look similar to the M0 = 25 case shown above. To better understand how these various modeling techniques work, we can look at their various search efficiencies.
5.2
Search Efficiencies of the 1-D Schemes
In a Lagrangian setting, it is easy to look at search efficiency as a function of the model parameters. I have placed bugs with the same behavioral characteristics I introduced in the last section in a finite domain with periodic boundary conditions. When a fly reaches xin = L, then xin+1 = L. The finite domain is more physical, in that flies can only move so far before starving or perhaps entering another foraging swarm, and makes the problem easier computationally. Since flies cannot move off to infinity, they must eventually interact with the attractant. Additionally, when flies saccade within rs > 0 of the sink, they stop moving and "stick" in that position, having found the sink. Figures 5.8 and 5.9 shows the search efficiency of the 3 models. These examples show that a strategy were each actor is searching are much faster 2x than flies on a random walk, without any environmental or interaction effects. Likewise, the directionally-interacting models are faster than the model featuring only the searching flies. The
Search Efficiencies of the 1-D Schemes 31
T = 1000
500 400 300 200 100 0 -1000 250 200 150 100 50 0 -2000
T = 2000
-500
x
T = 3000
0
500
1000
-1000
x
T = 4000
0
1000
2000
200 150 100
250 200 150 100
50 0 -4000
50 0 -4000
-2000
x
0
2000
4000
-2000
x
0
2000
4000
Figure 5.7: This figure shows the group of flies on a food source grow with time. It is the histogram time series for the green simulation from Fig. 5.5. The growth of the bin above x = 100 shows the effectiveness of the search strategy, as the point of highest density migrates from the origin, where the the bugs were initialized. mean times to having great than 75% of the swarm located for the simulations illustrated above are, Random 22582, Attractant 11686, M0 = 200 M0 = 25
7051, 6972.
From those limited samples, it appears that the search efficiency is practically the same for the two full models with the spatial parameters im-
32
A New Direction Search Efficiency for 2 Simulations - I
Random
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.5
1
1.5
2
2.5 x 10
4
Sink Only
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
2000
4000
6000
8000
10000
12000
14000
Figure 5.8: This figure shows the percentage of bugs that have found and landed on the food source vs time step, for 4 versions of the 1-D model and 5 runs of the simulation each. For each simulation, the attractor was located at x = 100, L = 500, rs = 5, and X (0) = 0. In the top, the bugs are simply L vy walkers. The simulation shown on the bottom is for the bugs with an e attractor. Note the differing scales in time, which show the efficiency of the swarms.
Search Efficiencies of the 1-D Schemes 33 Search Efficiency for 2 Simulations - II
M = 25
0
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
1000
2000
3000
4000
5000
6000
7000
8000
M0 = 200
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
1000
2000
3000
4000
5000
6000
7000
8000
Figure 5.9: This figure shows the search efficiencies for 5 directionallyinteracting swarms with attractants. The M0 = 25 and 200 simulations are ordered top to bottom. The horizontal scales are the same for these two plots, meaning that their efficiencies are similar, with M0 = 25 swarms seeming to be a little quicker.
34
A New Direction plemented with that set of spatial parameters (values of L and rs ). The M0 = 25 runs, however, display more variance than the M0 = 200 simulations. I believe that an interesting extension of this work would be to investigate the efficiency of the full model over a parameter space with ranging values of M0 and rs /L.
5.2.1
Interpretation and Possible Extensions
The full model of a directionally-interacting swarm seems to reflect the dynamics of actual swarms of searchers fairly well. The governing rules do not create a bias in direction for the swarm in the case where food sources are far away from the swarm or non-existant, the walkers still diffuse as would be expected from a L vy Walk, and the full model is able to complete e searches faster than if each bug were to find the food source on its own. This model may also be exposing some interesting behavior, which might warrant further investigation. First, it seems like bugs in the full model who end up far from the origin, and food sources, tend to be grouped relatively closely with respect to their distance from their initial position. This is evident in the long-time histograms, where there will be a bin with several flies in it far from the origin, see Figs. 5.6 and 5.7. It would be interesting to see whether these smaller swarms, or packets, form early and then diffuse out, aggregate in the far field, or some combination of both. These groups could take the 'primer' and 'follower' behavior to another scale, where they actually move the entire swarm out of its immediate area. In a sense, these packets might be why a swarm moves from the kitchen to the back porch.
5.3
The One Dimensional Eulerian Framework
The Lagrangian formulation in the last section is numerically convenient, being both fast in computation and easily traceable, but as I showed in Chapter 4, it can be helpful to have an Eulerian formulation of the model to provide model validation and analytical results. In this section, I will derive the equations of motion for the continuum case of the model presented in the last section of the chapter. The continuum version of my one dimensional model closely reflects the models built by Grunbaum (2000), where the continuum approximation of bugs moving in one dimension is also governed by an internal state, such as satisfaction with the environment. In this case, however, the internal state is the countdown clock, , which actually measures a fly's satisfaction
The One Dimensional Eulerian Framework 35 with its environment at the time and position of the last saccade. This will lead to some interesting governing equations, which will take the form of partial integro-differential equations. To begin, let ( x, t, ) be the density of bugs at position x and time t with time left until a saccade. Then consider both the left- and right-moving densities, r ( x, t, ) and l ( x, t, ), which move with speed c and satisfy, ( x, t, ) = r ( x, t, ) + l ( x, t, ), and, ( x, t, )d = ( x, t). Now consider a rectangle of size x by , placed somewhere in the upper half plane created when x R and > 0, as pictured in Figure 5.10. Due Fluxes in the Continuum Case
Right Traveling
}
Left Traveling
X
}
X
Redistributed Density
Figure 5.10: The diagram shows the fluxes in and out of the rectangle with time. The characteristics move down in and right or left, depending on which population is being tracked. In addition, density which reaches the = 0 line at ( x0 , t0 ) is redistributed to r ( x0 , t0 , ) and l ( x0 , t0 , ) with some probabilities.
36
A New Direction to the complicated fluxes for a stationary rectangle it is best to consider a rectangle which travels with the characteristics, and exploit the material derivative to say that, for the right-moving density, Dr = F (, x, t, ), Dt where F (, x, t, ) is the redistribution function, which carries mass from the x-axis back into the domain. Then, by the definition of the total derivative, Dr = Dt = Thus for r ( x, t, ) and l ( x, t, ), r r r +c - = t x Fr (, x, t, )
r t
+
r t
r x x t
+
r t r .
r + c x -
l l l -c - = Fl (, x, t, ). t x The theory on random walks governed by jumps in an internal state presented by Grunbaum (2000) predicts this system of equations. He starts with a system of the form, r r + c - ( f r r ) = ( )(l - r ) - ( )r + t x l l + c - ( f l l ) = ( )(r - l ) - ( )l + t x
0 0
( )r ( ) T (, )d ( )l ( ) T (, )d .
In the above, is the turning rate for bugs at , is the rate at which jumps in occur, f r and f l are operators which alter the rate of change of in time, and T (, ) is the transition probability from to , given that a transition occurs, where,
0
T (, )d = 1.
In the model that I have built, Grunbaum's model is reduced by, ( ) = 0 for > 0, ( ) = 0 for > 0,
The One Dimensional Eulerian Framework 37 when > 0 and f l = f r = 0. Then since ( ) = , the integral simply samples the transition probability, to yield the system seen in 5.4, where, Fr ( x, t, ) = [r ( x, t, 0) + l ( x, t, 0)] Tr (, 0), and similarly for Fl . Hence, Fr and Fl can be interpreted as the transition probabilities for the total density at the x-axis as it is distributed into the right- and left-traveling densities through . In the simplest case, where the bugs in the Lagrangian simulation would be on L vy Walks, e Fr ( x, t, ) = Fl ( x, t, ) = c - , 2
when > min . In the cases where the behavior rules are more complicated, the function F takes on some interesting forms. For the model of foraging L vy Walkers, that do not interact but are attracted to a sink as I described e in Sec. 5.1.2, c - , for | x - xsink | > rv 2 - Fr ( x, t, ) = Fl ( x, t, ) = c rv for | x - xsink | < rv , 2 |x-x |
sink
where rv > 0 is the vision radius of the bugs and xsink is the position of the nearest food source. This function redistributes density on the = 0 boundary to lower values of for points within the viewing radius of a food source. Lastly, we can look at the full model for a directionally interacting swarm, where for the right-traveling case, Fr ( x, t, ) = P SM where, 0.9, if Mr > Ml P = 0.5, if Mr = Ml 0.1, otherwise, and Ml = Mr =
x (-| x - x |) d dx - 0 ( x , t, ) e ( x , t, )e(-| x- x |) d dx . x 0
,
S reflects the dynamics of being near a sink, and takes a simple form, S= 1, if | x - xsink | > rv rv , otherwise. |x-x |
sink
38
A New Direction Lastly, the function M gives the dependency on the total Morse potential by, 1, if ( Ml + Mr ) < M0 M= 10, otherwise. The function M forces the distributive function, F, to spread more of the saccading density at low values of when the total Morse potential exceeds its critical value. The model, a combination of the equations listed over the last page, is now a partial integro-differential equation. Further study in this area will have to develop solution techniques for this type of equation, either analytical or numerical approximations, to validate the Lagrangian model.
5.4
Lagrangian Modeling in Two Dimensions
The extension of the Lagrangian model in one dimension to two spatial dimensions is easy to make and biologically relevant. The interesting foraging behavior exhibited by fruit flies occurs in the horizontal plane, as saccading is usually confined to the plane. Thus the two-dimensional version of my model will allow for comparisons between my numerical results to the data presented by Frye et al. (2003) and for easier interpretation of the numerical output. The model remains essentially the same as the version shown earlier in this chapter, only scalar position and velocity are replaced with vector quantities. The transition to two dimensions adds complexity to the turns that accompany saccades. In the one-dimensional model, a saccading fly either did not change direction or turned in the opposite direction. In the two-dimensional model, turns are sampled from identical normal distributions about 90 degrees. A sample flight path is shown in Fig. 5.11. In that simulation, the fly is constrained to the arena outlined in red and will saccade away from the wall when it gets close, mimicking biological behavior. I designed these simulations to reflect, as closely as possible, the arena tests performed by Frye et al. (2003), to provide at least a heuristic validation of the model for individual free flight. Frye et al. (2003) develop an integrative model for a fly in an arena, where the probability of saccade is evaluated at each time step, and it favorably compared to the biological behavior. Thus, I have created Figure 5.12 to show the similarities between the two modeling techniques, and the similarity between my model and the biology. Further study of this two- dimensional model will have to include investigations on the role of interaction in two dimensions
Lagrangian Modeling in Two Dimensions 39 A 2-D Foraging Flight Path
200
150
100
50
0
-50
-100
-150
-200 -200
-150
-100
-50
0
50
100
150
200
Figure 5.11: This figure shows the flight path of a foraging fly attracted to the food source at the blue circle. Note the number of saccades near the food source as opposed to other parts of the arena. on swarm cohesiveness and search efficiency.
40
A New Direction
Position Histogram for a Foraging Fly
100 x 10 18
!3
80 16
60 14
40 12 20 10 0
8 !20
6 !40
!60
4
!80
2
!100 !100
!80
!60
!40
!20
0
20
40
60
80
100
Figure 5.12: This figure shows the percentage of time spent in each cell over a simulation of length 40,000. In this simulation, one bug foraged through the flight arena for a food source located at xsink = (-30, 20). The bug moved with speed c = 30, its initial position was assigned randomly, = 2.32, and rv = 50. This image compares favorably with similar figures from Frye et al. (2003).
Chapter 6
Conclusions
With this thesis, I investigated Lagrangian and Eulerian descriptions of insect swarms. The individual-based Lagrangian techniques are fit primarily for numerical, stochastic, simulation. Alternatively, the continuumreminiscent Eulerian descriptions are difficult to solve, either analytically or numerically. These descriptions do, however, provide information that is much more general to specific swarm behaviors than the Lagrangian techniques. In the course of my investigation, I focused my energy on the swarming behavior of the fruit fly, drosophila melanogaster because it involved the combination of two interesting behaviors, saccading flight patterns and collective searching. Behavioral studies of drosophila demonstrate these two behaviors and I attempted to build a reasonable modeling technique to both reflect and combine these behaviors. I found that the flight of fruit flies, either individual or in a group, could be modeled as a random walk process governed by an internal clock, which counts down to the next saccade. From the biological information available, these models appear to be valid. Increasing saccade frequency near food sources, and thus decreasing the length of inter-saccade flights, results in a fly spending much more time in the vicinity of a food source and also allows a swarm of flies to find a food source in less time than it takes a swarm of flies which do not decrease their flight lengths. Likewise, decreasing inter-saccade flight length in areas of high Morse potential simulates the interaction that leads to swarm formation. Turning which is influenced by Morse potential also contributes to the swarming behavior. The increase in search efficiency seen when flies interact directionally and with food sources reflects the biological results of Tinette et al. (2004). This project has raised more questions than it has answered. In building a
42
Conclusions successful model, I have prepared a technique by which we can now study how the different parameters in my models, such as M0 , rv , , and N the critical Morse potential, flies' vision radius, the L vy parameter, and the e swarm size affect the behavior of the swarm both in search efficiency and shape. I believe there are four directions that future research on this project will proceed along, 1. A full parameter space investigation of the models introduced above. 2. Development of an approximation method for the continuum model which will serve to validate the Lagrangian results. 3. Numerical experiments with multiple food sources of varying qualities in the domain of a swarm, looking for the behaviors described by Tinette et al. (2004). 4. Investigations into the behavior of flies undertaking long walks from the swarm, either individually or in small packets, and what conditions lead to these flights.
Bibliography
Bearon, R. N. and Pedley, T. J. (2000). Modelling run-and-tumble chemotaxis in a shear flow. Bulletin Of Mathematical Biology, 62:775791. Berg, H. C. (1983). Random walks in biology. Princeton University Press, Princeton, NJ. Cole, B. J. (1995). Fractal time in animal behavior - the movement activity of drosophila. Animal Behaviour, 50:13171324. Feller, W. (1968). An introduction to probability theory and its applications. Vol. I. Third edition. John Wiley & Sons Inc., New York. Frye, M. A. and Dickinson, M. H. (2004). Closing the loop between neurobiology and flight behavior in drosophila. Current Opinion In Neurobiology, 14:729736. Frye, M. A., Tarsitano, M., and Dickinson, M. H. (2003). Odor localization requires visual feedback during free flight in drosophila melanogaster. Journal Of Experimental Biology, 206:843855. Grunbaum, D. (1994). Translating stochastic density-dependent individual behavior with sensory constraints to an eulerian model of animal swarming. Journal Of Mathematical Biology, 33:139161. Grunbaum, D. (2000). Advection-diffusion equations for internal statemediated random walks. Siam Journal On Applied Mathematics, 61:4373. Marthaler, D., Bertozzi, A., and Schwartz, I. (2004). L vy searches based on e a priori information: The biased l vy walk. UCLA CAM Report, 04-50. e Mogilner, A., Edelstein-Keshet, L., Bent, L., and Spiros, A. (2003). Mutual interactions, potentials, and individual distance in a social aggregation. J. Math. Biol., 47(4):353389.
44
Bibliography Othmer, H. G., Dunbar, S. R., and Alt, W. (1988). Models of dispersal in biological-systems. Journal Of Mathematical Biology, 26:263298. Othmer, H. G. and Stevens, A. (1997). Aggregation, blowup, and collapse: The abc's of taxis in reinforced random walks. Siam Journal On Applied Mathematics, 57:10441081. Tinette, S., Zhang, L., and Robichon, A. (2004). Cooperation between drosophila flies in searching behavior. Genes Brain And Behavior, 3:39 50. Topaz, C. M. and Bertozzi, A. L. (2004). Swarming patterns in a twodimensional kinematic model for biological groups. Siam Journal On Applied Mathematics, 65:152174. Viswanathan, G. M., Buldyrev, S. V., Havlin, S., da Luz, M. G. E., Raposo, E. P., and Stanley, H. E. (1999). Optimizing the success of random searches. Nature, 401:911914. Weeks, E. R., Urbach, J. S., and Swinney, H. L. (1996). Anomalous diffusion in asymmetric random walks with a quasi-geostrophic flow example. Physica D, 97:291310.