1: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2: %% Trim Size: 9.75in x 6.5in
3: %% Text Area: 8in (include Runningheads) x 5in
4: %% ws-acs.tex : 4-3-2004
5: %% Tex file to use with ws-acs.cls written in Latex2E.
6: %% The content, structure, format and layout of this style file is the
7: %% property of World Scientific Publishing Co. Pte. Ltd.
8: %% Copyright 1995, 2002 by World Scientific Publishing Co.
9: %% All rights are reserved.
10: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
11: %
12: %%%%%%%%% FOR TEMPLATE OF TYPING OUT THE BIBLIOGRAPHY TEXT ONLY %%%%%%%%
13: \newcounter{myctr}
14: \def\myitem{\refstepcounter{myctr}\bibfont\parindent0pt\hangindent15pt[\themyctr]\enskip}
15: \def\myhead#1{\vskip10pt\noindent{\bibfont #1:}\vskip4pt}
16: %%%%%%%%% FOR TEMPLATE OF TYPING OUT THE BIBLIOGRAPHY TEXT ONLY %%%%%%%%
17:
18: \documentclass{ws-acs}
19: \usepackage{cite}
20: \usepackage{graphics}
21: \begin{document}
22:
23: \markboth{D. Helbing {\em et al.}}{Ant Behavior Under Crowded Conditions}
24:
25: %%%%%%%%%%%%%%%%%%%%% Publisher's Area please ignore %%%%%%%%%%%%%%%
26: %
27: \catchline{}{}{}{}{}
28: %
29: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
30:
31: \title{ANALYTICAL AND NUMERICAL INVESTIGATION OF ANT BEHAVIOR UNDER CROWDED
32: CONDITIONS}
33:
34: \author{\footnotesize Karsten Peters$^1$, Anders Johansson$^1$,
35: Audrey Dussutour,$^{2}$ and Dirk Helbing$^{1,3}$}
36: \address{$^1$ Institute for Transport \& Economics, Dresden University of Technology,\\
37: Andreas-Schubert-Str. 23, 01062 Dresden, Germany\\[1mm]
38: $^2$ Biology Department, Concordia University,\\
39: 7141 Sherbrooke St. W., Montreal, Queubec, H4B 1R6, Canada\\[1mm]
40: $^3$ Collegium Budapest~-- Institute for Advanced Study,\\
41: Szenth\'{a}roms\'{a}g u. 2, 1014 Budapest, Hungary}
42:
43:
44: \maketitle
45:
46: \begin{history}
47: %\received{(received date)}
48: %\revised{(revised date)}
49: %\accepted{(Day Month Year)}
50: %\comby{(xxxxxxxxxx)}
51: \end{history}
52:
53: \begin{abstract}
54:
55: Swarm intelligence is widely recognized as a powerful paradigm of
56: self-organized optimization, with numerous examples of successful applications in
57: distributed artificial intelligence. However, the role of physical interactions
58: in the organization of traffic flows in ants under crowded conditions has only
59: been studied very recently. The related results suggest new ways of congestion
60: control and simple algorithms for optimal resource usage based on local
61: interactions and, therefore, decentralized control concepts. Here, we present
62: a mathematical analysis of such a concept for an experiment with two alternative
63: ways with limited capacities between a food source and the nest of an ant colony. Moreover, we carry
64: out microscopic computer simulations for generalized setups, in which ants have
65: more alternatives or the alternative ways are of different lengths. In this way and
66: by variation of interaction parameters, we can get a better idea, how powerful
67: congestion control based on local repulsive interactions may be. Finally, we will
68: discuss potential applications of this design principle to routing in traffic or
69: data networks and machine usage in supply systems.
70: \end{abstract}
71:
72: \keywords{Ants; traffic; crowding; decentralized congestion control; multi-agent simulation;
73: ant algorithms; swarm intelligence}
74:
75: \section{Introduction}\label{sec:intro}
76:
77: A basic principle in ant trail formation is the indirect interaction and
78: communication of ants through the environment by deposited chemical pheromones
79: ({\it stigmergy}) \cite{Gra59,Wil62}. It enables ant colonies to make
80: adaptive choices based solely on local information. Together with travel
81: time experiences (e.g. in {\it Linepithema humile}
82: \cite{Gos89}) or effects of U-turns (e.g. in {\it Lasius niger} \cite{Bec92})
83: it leads individual ants to the use of the shortest among
84: several paths to transport food from distant places into their nest.
85: \par
86: The stigmergy concept in general allows for a complex collective
87: cooperative behavior of relatively simple agents \cite{Dor00a} and has
88: inspired a large number of new algorithms and applications especially in
89: the field of combinatorial optimization problems \cite{Dor04}. This
90: includes shortest path problems \cite{Dor97,Wed04}, routing in
91: communication networks \cite{Car98} or material flow scheduling on factory
92: floors \cite{Par00,Val01,Blu05}.
93: \par
94: However, efficient transportation on a system-wide level is not only concerned
95: with shortest paths, but also with traffic assignment and coping with
96: bottleneck situations. An insufficient regulation of traffic flow will
97: usually lead to a congestion of the apparently preferred path. This is
98: one of the most challenging problems in road traffic and
99: routing of data on the internet. Generally speaking,
100: the efficient distribution of limited resources by decentralized,
101: individual decisions is still an open problem in many networked systems.
102: \par
103: As will be shown in the following, the investigation of ant behavior can give some hints how to
104: deal with such problems. Therefore, Secs.~\ref{Sec2} and \ref{Sec3} will focus on ant traffic and
105: its load-dependent organization, including some analytical investigations. Section~\ref{Sec4}
106: will present some new numerical results for a microsimulation of ants based on the
107: social force concept. Finally, Sec.~\ref{Sec5} summarizes our results and discusses possible
108: areas of application.
109:
110: \section{Transport and Traffic in Ant Societies}\label{Sec2}
111:
112: Before we study the optimization of ant traffic under crowded conditions,
113: we shall briefly review some basic properties of the traffic organization in ants.
114: \par
115: Mass recruitment in ants has a vital reason: The
116: flow of ants should guarantee an efficient return of food to the nest.
117: This will favor flow maximizing trail systems and traffic organization.
118: However, another objective is the minimization or limitation
119: of round trip times $\tau$ (i.e. the time a
120: single ant needs to reach a food source from the nest and to return).
121: Any delay incurred by an ant during its round-trip slows down
122: the amplification mechanism underlying the recruitment process. In situations
123: of competition for food sources even a small decrease in round-trip duration could
124: improve the fitness of a colony against a
125: competitor by rapidly building up a worker force at the food source. A shorter round trip
126: time for the same flow of food return
127: therefore increases the performance and flexibility and, hence, the
128: fitness of an ant colony.
129: \par
130: Thus, we expect that ant behavior supports an optimal traffic organization
131: in terms of the flow $\Phi$ and round-trip time $\tau$. It is no
132: surprise that these performance criteria also apply to routing algorithms
133: in information networks and that they are meaningful for the transport of goods as well.
134: \par
135: In some regards, ant traffic is comparable to
136: vehicular traffic or pedestrian behavior \cite{Hel01}. Laden, nestbound
137: ants experience a slower speed than unladen, outbound ants. Therefore, an
138: ant trail may be congested with large variations in individual speed and
139: agility. However, ants cooperatively driven by shared goals of the colony
140: are different from the rather selfish behavior and distinct individual
141: interests in human transport systems.
142: \par
143: In contrast to vehicular traffic,
144: which is organized in unidirectional lanes of traffic, a complete directional
145: separation for outbound and inbound ants is uncommon due to the chemical
146: (chemotaxis-based) trail attraction. Whereas in some ant species like the wood ant {\it Formica
147: rufa} \cite{Hol55}, some termites \cite{Miu98}, army ants \cite{Couz03} or
148: African driver ants \cite{Got95} forms of unidirectional flow organization
149: can be observed, in most species a separation of opposite flows does not occur.
150: In case of directional traffic, the separation follows certain principles, where
151: either returning or outbound ants use the middle of a trail and the
152: other fraction moves at the outer regions of a trail. In
153: black garden ants ({\it Lasius niger}) \cite{Bec92} or leaf cutting
154: ants \cite{Bur03} exploiting renewable or permanent resources, no
155: separation is observed at all. This may be explained by the large
156: variations in individual speeds \cite{Hel02}.
157: \par
158: Presumably, the absence of directional separation leads to
159: collisions between ants moving in opposite directions \cite{Bur03}. But in
160: contrast to larger organisms, ants can stop very abruptly and recover
161: their initial velocity quickly after a frontal collision with another
162: ant.
163: \par
164: Theoretically, ant traffic flows can be described in the framework
165: of self-driven many-particle systems \cite{Hel01}. For particle flows in
166: general and for ants on a trail in particular, we have the formula
167: \begin{equation}
168: \label{eq:flow-density} \Phi=w\rho V \, ,
169: \end{equation}
170: where $\Phi$ is the
171: flow, $\rho$ the density of ants and $V$ their average speed.
172: $w$ denotes the width of the trail. However, the relation between $V$ and
173: $\rho$, known as fundamental diagram, is specific to microscopic
174: properties of the driven particles and cannot be determined easily from
175: scratch. Whereas for vehicular traffic and pedestrians a rich body of
176: theoretical work and empirical measurements exists, not much knowledge has been
177: gathered on fundamental diagrams of foraging ants \cite{Chowd00,Chowd02, Nis03}.
178: For {\it Atta cephalotes} (a leaf cutting ant) Burd {\it et al.} found the
179: empirical speed-density relation \cite{Bur02}:
180: \begin{equation}\label{eq:flow-velocity}
181: \Phi=w\rho V_m[1-(w\rho/k_m)^n] \, ,
182: \end{equation}
183: where $\Phi$ is the flow, $V_m$ represents the maximum average velocity
184: reached under free flow conditions, and $k_m$, $n$ are empirical
185: parameters estimated in experiments. For {\it Atta} ants the values of
186: $V_m=4.04$ cm s$^{-1}$, $k_m=0.59$ cm$^{-2}$ and $n\approx 0.64$ were
187: confirmed \cite{Bur02,Bur03}. This gives an estimate for the velocity
188: corresponding to the maximum flow of $V=0.39 V_m$.
189:
190: \section{Load-Dependent Traffic Organization in Ants} \label{Sec3}
191:
192: The basic challenge for any self-organizing traffic optimization is
193: the ability to adapt the routes of partial flows to the overall flow situation in the network.
194: If one link is congested, a certain portion of the overall flow should use an alternative
195: path to reach its destination. The degree of utilization of alternative routes that can be
196: called optimal depends crucially on the capacity and weight of the available alternative
197: paths and the overall flow that has to be served.
198: \par\begin{figure}[htbp]
199: \begin{center}
200: (a)\epsfig{file=setup4.eps, width=8.0cm}\\[5mm]
201: (b)\epsfig{file=encounter3b.eps, width=2.2cm}
202: (c)\epsfig{file=encounter2b.eps, width=2.2cm}
203: (d)\epsfig{file=encounter1b.eps, width=2.2cm}
204: \caption[]{({a}) Experimental setup of a binary bridge experiment.
205: The setup mimics many natural situations, where the physical constraints
206: in the environment restrict the geometry of trails. Moreover, it resembles
207: the situation in technical and economic transport problems, where the geometry
208: of networks is usually predefined by the existing infrastructure. To explore different
209: capacities of the branches $i=1,2$, the width $w$ of the branches is varied.
210: ({b}--{d}) Typical encounter at one of the choice points $j=1,2$.
211: Two ants, coming from opposite directions collide on a branch near
212: the entry point (b). An ant is pushed backwards (c) and redirected towards the
213: other branch (d). (After \cite{Dus04}.)}
214: \label{fig:setup}
215: \end{center}
216: \end{figure}
217: To study the organization of ant traffic in a setup with alternative paths
218: and restricted capacity, the binary bridge experiment by Deneubourg {\em et al.}
219: \cite{Den90}, in which
220: a food source is separated from a nest by a bridge with two alternative, but
221: equally long branches (Fig.\ref{fig:setup}), can be generalized: According to Eq.
222: (\ref{eq:flow-density}), the capacity of a branch for ant traffic is
223: determined by its width $w$. Therefore, in a recent experiment
224: with the black garden ant ({\it Lasius niger}), the widths of both branches
225: has been varied from 10mm to 1.5mm to induce increasing levels of
226: crowding \cite{Dus04}.
227: \par
228: It was observed that the recruitment dynamics
229: was not significantly influenced by the branch width $w$, while the ants
230: flow organization changed significantly. For wide branches, most ants used
231: only one of the branches. This spontaneous symmetry breaking can be
232: explained by small initial fluctuations in the pheromone level. During the
233: recruitment process, these initial disturbances are amplified, as a
234: greater number of ants use the trail with the initially or randomly higher pheromone
235: concentration and reinforce it. Surprisingly, the resulting asymmetric trail usage was not
236: conserved for widths $w\le 6$mm. Instead, symmetric traffic using both branches
237: occurred.
238: \par
239: This symmetry restoring transition can be explained based
240: on head-on encounters of ants traveling in opposite directions \cite{Dus04}. If
241: the width of the bridges is reduced, the branches become crowded and
242: the increased number of encounters of ants reduces their average speed.
243: Therefore, the restriction to one branch occurs only upto
244: a certain critical level of crowding. Above it, head-on encounters at
245: the choice points of a narrow bridge (see
246: Fig.\ref{fig:setup}b-d) trigger the emergence of a symmetric usage of
247: both branches: When ants entering one of the branches
248: collide with ants from the opposite direction, they are pushed backwards
249: and, with a probability $\gamma$ of about 60\%, they use the other branch after
250: the pushing event at the choice point. While these pushing events were practically
251: not observed on bridges of 10mm width, they occurred frequently on narrow
252: bridges \cite{Dus04}. Loosely speaking, a high congestion of ants on one narrow
253: branch leads to a re-routing of ants through alternative branches.
254: \par
255: In the following, we will shortly present a mathematical analysis of this
256: result complementary to what has been published in Ref.~\cite{Dus04}. For this,
257: let $\phi_j =\phi_1$ denote the outbound flow of ants from the nest to the food source and
258: $\phi_j = \phi_2$ the nestbound flow. Moreover, let $F_{ij}$ denote
259: the probability to choose branch $i\in\{1,2\}$ at choice point $j\in\{1,2\}$ (see Fig. \ref{fig:setup}a).
260: For {\it Lasius niger}, $F_{ij}$ is given by \cite{Bec92}
261: \begin{equation} \label{choice}
262: F_{ij}=\frac{(k+C_{ij})^2}{(k+C_{1j})^2+(k+C_{2j})^2}=1-F_{i'j} \, ,
263: \end{equation}
264: where $F_{i'j}$ describes the probability to choose the
265: alternative branch $i'=3-i$ at point $j$. The concentration-dependent choice
266: begins to be effective if the pheromone
267: concentration $C_{ij}$ on branch $i$ behind the choice point $j$
268: exceeds a certain threshold $k$. If every ant leaves a fixed amount $q$ of
269: pheromone on its trail and the pheromone decays with a constant rate $\nu$,
270: we can determine the local time-dependent pheromone concentration via the equation
271: \begin{equation}
272: \frac{dC_{ij}}{dt} = q [ \Phi_{ij}(t) + \Phi_{ij'}(t-T)] - \nu C_{ij}(t) \, .
273: \end{equation}
274: Herein, $T$ is the average time for an ant to get from one choice point to the other
275: and $\Phi_{ij}$ the overall flow of ants in direction $j$ (behind choice point $j$)
276: using branch $i$. The formula
277: \begin{equation}
278: \Phi_{ij}(t) = \phi_j(t) F_{ij}(t) [1- \gamma a \Phi_{ij'}(t-T)/w]
279: + \phi_j(t) F_{i'j}(t) \gamma a \Phi_{i'j'}(t-T)/w
280: \end{equation}
281: takes into account the probability $F_{ij}$ with which branch $i$ is chosen,
282: but it also considers the effect of pushing events at choice point $j$
283: to the other branch. $a \Phi_{ij'}(t-T)/w$ is the proportion of ants decelerated
284: on a branch of width $w$, when the counterflow of ants at
285: choice point $j'=3-j$ and time $t$ is $\Phi_{ij'}(t)$. The parameter
286: $a$ is proportional to the interaction time period and the effective cross section (lateral width) of ants.
287: $\gamma$ denotes the probability of being pushed to the other branch in cases
288: of encounters, i.e. the re-routing probability. One interesting question to be addressed later
289: is, whether the empirically found value $\gamma=0.57$ is optimal due to evolutionary
290: pressure and whether the optimal value depends on the respective scenario.
291: \par
292: Due to the normalization $F_{ij}(t) + F_{i'j}(t) = 1$, we have the relation
293: $\Phi_{ij}(t) + \Phi_{i'j}(t) = \phi_j(t)$, i.e. $\phi_j(t)$ is the overall flow behind
294: choice point $j$, as demanded. Furthermore,
295: in the stationary case we find $dC_{ij}/dt = 0$,
296: $F_{ij}(t-T) = F_{ij}(t) = F_{ij}$, $\Phi_{ij}(t-T) = \Phi_{ij}(t) = \Phi_{ij}$,
297: and $\phi_j(t-T) = \phi_j(t) = \phi_{j'}(t) = \phi$
298: (as the nestbound flow and the outbound flow should
299: be equal). This implies
300: \begin{equation}
301: C_{ij} = \frac{q}{\nu} ( \Phi_{ij} + \Phi_{ij'} ) = C_{ij'} = \dots = \frac{q\phi}{\nu}
302: [ \gamma a (\Phi_{i'j}+\Phi_{i'j'})/w + 2F_{ij}(1-\gamma a \phi/w)]
303: \label{Ce}
304: \end{equation}
305: and
306: \begin{equation}
307: F_{ij} = F_{ij'} = 1 - F_{i'j} = 1 - F_{i'j'} \, ,
308: \end{equation}
309: because of $C_{ij} = C_{ij'}$ for all $i$ [cf. Eq. (\ref{choice})]. Additionally, with the definitions
310: \begin{equation}
311: \frac{C_{ij}+ C_{i'j}}{2} = \frac{C_{ij'}+ C_{i'j'}}{2} = C \quad \mbox{and} \quad
312: \frac{C_{ij}- C_{i'j}}{2} = \frac{C_{ij'}- C_{i'j'}}{2} = D
313: \end{equation}
314: one can show
315: \begin{equation}
316: C_{ij} = C_{ij'} = C + D \, , \qquad C_{i'j} = C_{i'j'} = C - D
317: \end{equation}
318: [see Eq. (\ref{Ce})] and
319: \begin{equation}
320: \Phi_{ij} + \Phi_{ij'} = \frac{\nu}{q} (C+D) \, , \qquad
321: \Phi_{i'j} + \Phi_{i'j'} = \frac{\nu}{q} (C - D) \, .
322: \end{equation}
323: A detailed calculation gives
324: \begin{equation}
325: C = \frac{C_{ij}+ C_{i'j}}{2} = \frac{q}{2\nu} (\Phi_{ij} + \Phi_{ij'} + \Phi_{i'j} + \Phi_{i'j'}) = \frac{q\phi}{\nu}
326: \end{equation}
327: and, using relation (\ref{Ce}),
328: \begin{eqnarray}
329: D &=& \frac{C_{ij}- C_{i'j}}{2}
330: \nonumber \\
331: &=& \frac{q\phi}{2\nu} [ \gamma a (\Phi_{i'j} + \Phi_{i'j'} - \Phi_{ij} - \Phi_{ij'})/w
332: + 2F_{ij}(1-\gamma a\phi /w) -2F_{i'j}(1-\gamma a \phi /w)] \nonumber \\
333: &=& - \gamma a \phi D/w + \frac{q\phi}{\nu} (1-\gamma a \phi / w)
334: \frac{(k+C_{ij})^2 - (k+C_{i'j})^2}{(k+C_{ij})^2 + (k+C_{i'j})^2} \nonumber \\
335: &=& C (1-\gamma a \phi / w) \frac{(2k + C_{ij} + C_{i'j})(C_{ij} - C_{i'j})}{(k+C+D)^2 + (k+C-D)^2}
336: - \gamma a \phi D / w
337: \nonumber \\
338: &=& \frac{2C(1-\gamma a \phi / w)(k+C)D}{(k+C)^2 + D^2} - \gamma a \phi D / w \, .
339: \end{eqnarray}
340: This allows one to calculate all values of $C_{ij}$, $\Phi_{ij}$, and $F_{ij}$ with the above formulas.
341: We find
342: \begin{equation}
343: \label{eq:Csolution}
344: C_{ij}=C_{ij'}=\frac{q \phi}{\nu}+D \quad \mbox{and} \quad C_{i'j}=C_{i'j'}=\frac{q \phi}{\nu}-D,
345: \end{equation}
346: with
347: \begin{equation}\label{eq:D=0}
348: D=0
349: \end{equation}
350: or
351: \begin{eqnarray}
352: D^2 &=& \frac{2C(1-\gamma a \phi / w)(k+C)}{1+\gamma a \phi / w} - (k+C)^2 \nonumber \\
353: &=& C^2 - \frac{k^2 + \gamma a \phi k^2 / w + 4\gamma a \phi k C / w
354: + 4\gamma a \phi C^2 / w}{1+\gamma a \phi / w} \nonumber \\
355: &=& \frac{q^2\phi^2}{\nu^2} - \frac{k^2+\gamma a \phi(k+2q\phi/\nu)^2/w}
356: {1+\gamma a \phi / w}
357: \, . \label{eq:Dsquare}
358: \end{eqnarray}
359: As long as $D^2>0$, the stable stationary solution is characterized by
360: an asymmetric usage of the alternative branches with $C_{ij}=\frac{q \phi}{\nu}\pm
361: \sqrt{D^2}$ and $C_{i'j}=\frac{q \phi}{\nu}\mp \sqrt{D^2}$.
362: If $\gamma=0$, $D^2>0$ is
363: fulfilled for $q\phi/\nu>k$. For $q\phi/\nu<k$, the traffic flow is too low to
364: generate a certain preference, as the pheromone concentration is not high enough to be
365: noticed. Then, the analytical solution collapses to $D=0$, resulting in a symmetric usage of
366: both branches, as ants choose between both branches at random.
367: \par\begin{figure}[htb!]
368: \begin{center}
369: \epsfig{file=Fig2_analytical.eps, width=12.0cm}
370: \end{center}
371: \caption[]{Analytical results according to Eq. (\ref{eq:Csolution}) for
372: the parameter values $q=1$, $k=6$, and $\nu=1/40$min$^{-1}$.
373: ({a}) Fraction of the overall flow on each branch
374: in the stationary state, when both branches have a width of 3mm and the value of
375: the re-routing probability $\gamma$ is 0.57 as
376: observed for {\it Lasius niger}. The dashed
377: curve shows the solution in the absence of pushing ($\gamma=0$), where the
378: model predicts asymmetrical traffic for {\em any} flow above a very low value of the overall
379: flow $\phi$. (After \cite{Dus04}.)
380: ({b}) Threshold of the symmetry-restoring transition as a function of the overall
381: flow $\phi$ of ants and the re-routing probability $\gamma$ for branch widths of
382: $w=1.5$mm, 3mm, and 6mm. The value
383: $\gamma=0.57$ found for {\it Lasius niger} is indicated by the dashed line. }
384: \label{fig:bifu}
385: \end{figure}
386: While the asymmetric solution is maintained
387: also for high flows, if no pushing is considered (Fig.\ref{fig:bifu}a), the
388: situation changes considerably, when a finite re-routing probability $\gamma>0$
389: is taken into account:
390: Above a certain critical threshold $\phi_{\rm S}(w,\gamma)$
391: of the flow, where
392: $D^2(\phi_{\rm S})$ becomes zero again, a symmetric stationary
393: assignment of traffic to both branches is observed, which is caused by pushing events.
394: The critical flow $\phi_{\rm S}$, at which this symmetry-restoring transition takes place, is a
395: function of the width $w$ of the branches and the re-routing probability $\gamma$
396: (Fig.\ref{fig:bifu}b).
397:
398: \section{Multi-Agent Simulation of Ant Traffic}\label{Sec4}
399:
400: As the pushing mechanism is based on random encounters in the area of the
401: decision points only (the nodes of the trail network), it is not clear at all that these would be representative for
402: the congestion level on the subsequent part of the respective branch, i.e. the whole
403: link. Therefore, it
404: is hard to imagine that a mechanism as simple as pushing, based on local interactions,
405: would lead to an optimal traffic organization. Nevertheless, it could be shown that
406: the value $\gamma = 0.57$ of the re-routing probability
407: caused the usage of an additional branch shortly before
408: the capacity of a single branch was reached \cite{Dus04}. Was this just accidental, or
409: is the value of $\gamma$ carefully chosen, e.g. by evolutionary optimization?
410: \par
411: In order to find out whether optimality would also be found for other,
412: more complicated scenarios, we have carried out computer simulations
413: (see Fig. \ref{fig:simulation}). Ants were
414: simulated assuming a distribution of velocities and times at which
415: a network with alternative paths (branches)
416: was entered. Their motion was specified by
417: a social force model \cite{socforce1,socforce2}, according to which ants had desired directions
418: (given by the endpoints of network links) and repulsive interactions.
419: The encounters of individual ants were determined,
420: and the probability of re-routing due to a
421: collision near the choice points was given by the parameter $\gamma$.
422: Alternatively, with a probability of $1-\gamma^{dt}$,
423: where $dt$ denotes the time discretization of the numerical integration method,
424: we allowed ants to climb over each other.
425: \par
426: Furthermore, in our model we assumed that, once an ant had decided to walk in one
427: direction on a specific link (branch), it would continue in this direction upto the endpoint of this
428: link. That is, we neglected U-turns due to interactions.
429: When an ant reached a node (choice point), the next destination was chosen among all connected links
430: that did not turn more than 90 degrees compared to the ant's previous direction. Stronger turns
431: (apart from U-turns) were only considered, if there were no other links available. The choice
432: probabilities were assumed to be proportional to the squares of the pheromone
433: concentrations on the respective links. These were increased by 1 each time an ant passed. However,
434: before, the pheromone concentrations were multiplied by a certain decay factor in each time step $dt$.
435: \par
436: Microsimulations based on this simple multi-agent model (and similar ones) can reproduce the
437: analytical results in Sec.~\ref{Sec3} quite well (see, e.g., Ref.~\cite{Dus04}).
438: Moreover, they allow one to study scenarios in which
439: \begin{itemize}
440: \item the branches have different width and/or lengths,
441: \item there are more than two alternative branches, or
442: \item parameters such as the re-routing probability are chosen
443: different from the empirical value.
444: \end{itemize}
445: This facilitates to study the system performance with and without the pushing mechanism
446: and to relate it to the empirically observed value of $\gamma$.
447: \par\begin{figure}[htbp]
448: \begin{center}
449: \epsfig{file=Frame.ps, width=13.0cm}
450: \caption[]{Simulation setup for a binary bridge scenario. To explore the effect of different
451: capacities of the two alternative branches $i=1,2$, the width $w_i$ of the branches is varied.
452: A video film of this scenario clearly shows that
453: the ants mainly use the shorter of two wide branches, while they use both narrow branches
454: (see {\tt http://www.trafficforum.org/ants}).}
455: \label{fig:simulation}
456: \end{center} \end{figure}
457: It has been found that the collision-based adaptive routing mechanism can be called
458: optimal. It seems not only to optimize the flow, but also the round-trip time
459: by limiting the density-related speed reduction.
460: Figure~\ref{fig:results_max_flow} shows the averaged round trip time $\tau$ of ants
461: vs. the reached overall flow $\phi$ as characteristic performance measures
462: for the quality of a routing algorithm. The binary bridge is assumed to be symmetrical,
463: and the parameters applied in the simulation correspond to the findings for
464: {\it Lasius niger}. The region where the capacity of one
465: branch is reached, is characterized by a sharp increase in the round trip
466: time. Our simulations indicate that, after the symmetry-restoring
467: transition to the usage of both branches, which occurs
468: for heavy traffic with $\gamma>0$, the available capacity for returning food
469: is doubled compared to a hypothetical
470: scenario without direct interactions and without pushing ($\gamma = 0$).
471: \par\begin{figure}[htb!] \begin{center}
472: \epsfig{file=Capacity.ps, width=10.0cm}
473: \end{center}
474: \caption[]{Results of Monte Carlo simulations depicting the round trip time vs.
475: the overall flow for the binary bridge experiment, taking into account pushing (+) or not (x).
476: The curves are fitted as guide for the eyes, while the dashed curve
477: represents the case $\gamma=0$, the solid line corresponds to
478: simulations considering pushing.\\
479: A symmetric binary bridge with $w=3$mm of both branches was simulated for
480: different flows $\phi$. For each data point (representing the result of one Monte
481: Carlo simulation), the values of the round trip time $\tau$ were averaged over 4 minutes after the
482: steady state was reached. We used the same parameters as in Fig.\ref{fig:bifu}. The empirically observed
483: value of $\gamma=0.57$ was applied to consider the effect of pushing events. Note that
484: pushing allows to reach higher flow values, as the use of additional branches provides additional
485: capacity.}
486: \label{fig:results_max_flow}
487: \end{figure}
488: Our simulations also substantiate
489: positive effects of pushing ($\gamma>0$) in terms of reduced round trip times. For the same
490: round-trip time, the interaction-based
491: sensitivity to the level of crowding allows a much higher overall flow.
492: In other words, the observed traffic organization guarantees shorter round
493: trip times for the same flow of food return compared to hypothetical scenarios without
494: pushing.
495: \par
496: To guarantee the optimality of traffic assignment for a wide
497: range of flows, the value of $\gamma$ is crucial. It governs the transition
498: point and balances the influence of the pheromone-driven behavior and
499: the congestion-related re-organization of flows. For low traffic volumes,
500: the beneficial properties of pheromone-guided trail-following behavior
501: should dominate. Especially for paths of different path lengths, the
502: majority of ants should be engaged in the shortest path.
503: \par
504: We expect that there is an optimal value of $\gamma$ between 0 and 1: On the one hand,
505: a high value of $\gamma$ would result in re-routings to longer paths, which is not
506: necessary if the capacity of the shortest path is not exhausted. However,
507: in very crowded situations, ants should use additional paths, even if they are longer.
508: The transition to this behavior should be driven by the utilization of the available capacity
509: of the preferred branch. On the other hand, if the value of $\gamma$ is too small,
510: a transition due to pushing never takes place (see Fig.~\ref{fig:bifu}b)
511: or only, after a undesireable level of congestion has been produced.
512: From these considerations we conclude that the optimal value of
513: $\gamma$ would trigger the symmetry-restoring transition just before the
514: capacity of a branch is reached, as has been actually observed \cite{Dus04}.
515: \par\begin{figure}[htb!]
516: \begin{center}
517: (a)\epsfig{file=Fig3a.eps,
518: width=10.0cm}\\ (b)\epsfig{file=gamma_tau.ps,
519: height=4.2cm}(c)\epsfig{file=gamma_v.ps, height=4.4cm}
520: \end{center}
521: \caption[]{Results of Monte Carlo simulations for binary bridges with different
522: length of the alternative paths. An overall flow $\phi=15\mbox{min}^{-1}$ was applied.
523: The length of branch 1 was fixed ($l_1=10$cm), while the length $l_2$ of branch 2
524: was varied. The simulations were carried out for different values of the re-routing
525: probability $\gamma$. Each point represents an average over 10 minutes of simulated ant traffic.
526: ({a}) Fraction $\Phi_{2}$
527: of the overall flow using branch 2 vs. $l_1/l_2$ for
528: $\gamma=1$ (every encounter results in a re-routing), $\gamma=0.57$
529: (the value observed in {\it Lasius niger}) and without pushing ($\gamma=0$).
530: For all values of $\gamma$, the inequality $\Phi_{2}<\Phi_{1}$
531: for $l_2>\l_1$ reflects the pheromone-driven preference for the shortest path. The curves are
532: fitted as guide for the eyes: While the dased curve corresponds to $\gamma=0.0$, the solid
533: curve represents $\gamma=0.57$ and the dashed-dotted curve $\gamma=1.0$.
534: ({b}) Averaged round-trip time $\tau$ vs. the re-routing probability $\gamma$ for $\l_1=10$cm,
535: $l_2=20$cm and $\phi=\Phi_{11}+\Phi_{21}=15$min$^{-1}$. In this situation,
536: the experimentally observed value of $\gamma=0.57$ is the optimal value to
537: obtain a minimal round-trip time. A second order polynomial was fitted to the
538: simulation results as guide for the eyes.
539: ({c}) Average velocity $V$
540: vs. $\gamma$. The same parameters as in (b)
541: were applied. This plot shows that the re-routing probability $\gamma$ does
542: not optimize the velocity, which is maximal if nearly equal flows are assigned
543: to both branches.}
544: \label{fig:results_gamma_length}
545: \end{figure}
546: For asymmetric binary bridges with branches of the same width but different
547: length, the interrelation between the optimal traffic distribution and the
548: parameters of the local feedback through physical interactions near the choice
549: points can be discussed also for flows below the capacity of a single
550: branch (Fig.\ref{fig:results_gamma_length}):
551: Due to the pheromone-guided behavior, the shorter branch is used more frequently, but with
552: increasing $\gamma$, a larger and larger portion of flow utilizes the
553: longer branch. This effect is observed even if the critical flow for
554: symmetry restoring is not reached. For geometrically asymmetric binary
555: bridges, the symmetry restoring transition is characterized by a
556: continuous function
557: \begin{equation}
558: \Phi_{1j}/\Phi_{2j} \sim l_1/l_2 \, .
559: \end{equation}
560: Thus, the flows beyond
561: the transition point $\phi_S$ are directly proportional to the lengths of the
562: alternative paths, but below the transition point, a jump
563: in the assignment occurs for the symmetric binary bridge. In Fig.
564: \ref{fig:results_gamma_length}a, we have $\phi<\phi_S(\gamma)$ for
565: $\gamma=0$ and $\gamma=0.57$, whereas $\phi>\phi_S(\gamma)$ if
566: $\gamma=1$.
567: \par
568: The optimal assignment depends on the relative length of
569: both branches and the fundamental diagram of traffic. In certain
570: situations we have found numerically that a re-routing probability of
571: $\gamma=0.57$ provides the optimal behavior in terms of flows and
572: round-trip times (Fig. \ref{fig:results_gamma_length}b) even for
573: sub-critical flows ($\phi<\phi_{\rm S}$). However, for other geometries of the
574: path network, we have found that other values of the re-routing probability
575: ($0.5<\gamma<0.8$) can improve the traffic assignment.
576: \par
577: In the same way
578: as for binary bridges with asymmetric length, traffic organization is
579: also sensitive to capacity. If two branches $i$ with different widths $w_i$ are
580: simulated, the flows in the heavy traffic case (after symmetry restoring)
581: adapt to the capacity of the available branches according to
582: \begin{equation}
583: \Phi_{1j}/\Phi_{2j} \sim w_{2}/w_1 \, .
584: \end{equation}
585: In general, the results obtained in
586: simulations with different capacities are similar to the traffic organization in
587: the case of two branches with equal capacities, but different lengths.
588: \par
589: \begin{figure}[htb!]
590: \begin{center}
591: \epsfig{file=antfignew.eps, width=13cm}
592: \end{center}
593: \caption[]{Simulation results for a completely symmetrical bridge with two large
594: branches that are again subdivided into two small branches each. The overall width of the
595: bridge (i.e. all branches) is the same everywhere. (a) Fraction of ants
596: on the two large branches as a function of the overall ant flow $\phi$. (b) Fraction of
597: ants on the four small branches. The lower figures give a schematic representation of
598: the distribution of ant flows over the large and small branches (c) at very low flows,
599: (d) at medium flows and (e) at large flows.}\label{fourbridge}
600: \end{figure}
601: Our microscopic simulations support the conjecture that the discovered
602: collision-based traffic organization principle in ants generates optimized traffic for a
603: wide range of conditions. It optimizes the utilization of available
604: capacities and minimizes round-trip times. Simulation results for a symmetrical bridge with
605: four branches are shown in Fig.~\ref{fourbridge}. It can be seen that the majority of ants
606: uses one of the large branches and, within both branches, the majority of ants uses one
607: of the small branches. Which of the branches is preferred basically depends on random fluctuations
608: or the initial conditions. When the overall ant flow is increased, the ant flows on the branches
609: become more equally distributed. Above a certain overall flow, the usage
610: of the bridge is completely symmetrical, as for a bridge with two branches.
611: On a logarithmic flow scale,
612: it appears that we do not have a {\em discontinuous} transition from the use of one small branch
613: to the use of two small branches and another transition to the use of three or four branches.
614: However, the sharpness of the transition is certainly a matter of the choice of parameters.
615:
616: \section{Summary and Outlook}\label{Sec5}
617:
618: In this article, we have discussed a mechanism of traffic organization in ants
619: that, on the one hand, establishes the shortest connection based on pheromone attraction,
620: but on the other hand, guarantees the usage of additional capacities by collision-based
621: re-routing. In the future, such mechanisms are likely to have interesting applications in the tradition
622: of ant algorithms and swarm intelligence: For example, if we have a number of
623: machines with comparable production functions, it will be most profitable to
624: use one machine only when the production load is small, and to add further machines
625: when the load exceeds certain thresholds. Note that the optimal thresholds, just as
626: observed for ants, are usually somewhat below the production capacity
627: of a machine, as it will in many cases not be advisable to go to machine limits for reasons
628: of costs, reliability, maintenance, etc. Machine parameters such as load-dependent
629: cycle times play a role similar to the lengths of alternative paths in
630: ant systems.
631: \par
632: Altogether, cost-efficient load-balancing of machines or transportation
633: capacities can be a difficult issue, and interaction-based ant algorithms seem to be a
634: promising approach, particularly as they do not require central control with the potential
635: of information losses and delays or decision conflicts. Instead, they are based on simple
636: local decision rules.
637: \par
638: Applications to the routing of traffic are also an interesting subject. Clearly, congestion
639: can only be avoided as long as there are enough alternative paths over which traffic
640: can be distributed. If this capacity is not available, the transferability of our ant
641: model is limited. However, it appears that congestion in data and vehicle traffic could
642: be reduced by better routing algorithms. The interaction-based approach by ants would
643: be promising here as well, although generalizations will be needed due to
644: multi-origin-multi-destination flows that are far from being stationary. One
645: interesting research direction would be the investigation of decision functions different
646: from Eq.~(\ref{choice}). Another focus could be the extension of the interaction concept, replacing
647: collisions by passing messages between oppositely moving flow directions. In this way, information
648: about the traffic volume on different (directed) links could be passed by local interactions
649: similarly to ad-hoc networks based on transversal hopping, which have been suggested
650: for inter-vehicle communication \cite{IVC}. Thereby, it should be possible to establish
651: a decentralized algorithm for system-wide travel time or congestion information
652: (cf. Ref.~\cite{control}), which could be the basis for optimal (re-)routing decisions.
653:
654: \subsection*{Acknowledgements}
655:
656: D.H. and A.D. acknowledge the contributions of
657: Vincent Fourcassie and Jean-Louis Deneubourg to
658: the derivation of the analytical results regarding the collisions-based
659: symmetry-restoring mechanism proposed in Ref.~\cite{Dus04}.
660: K.P. and D.H. thank for partial financial support by the
661: EU-Project MMCOMNET.
662:
663: \begin{thebibliography}{99}
664:
665: \bibitem[1]{Wil62} E.O. Wilson, Chemical communication among workers of
666: the fire ant {\it Solenopsis saevissima}. 1. The Organization of mass-foraging. {\it Anim. Behav.} {\bf 10}, 134-147 (1962).
667:
668: \bibitem[2]{Gos89} S.~Goss, S.~Aron, J.L.~Deneubourg and J.M.~Pasteels,
669: Self organized shortcuts in the Argentine ant. {Naturwissenschaften} {\bf
670: 76}, 579-581 (1989).
671:
672: \bibitem[3]{Bec92} R. Beckers, J.L.~Deneubourg and S. Gross, Trails and
673: U-turns in the selection of a path by the ant {\it Lasius niger}. {\it J.
674: Theor. Biol.} {\bf 159}, 397-415 (1992).
675:
676: \bibitem[4]{Dus04} A.~Dussutour, V.~Fourcassie, D.~Helbing and
677: J.-L.~Deneubourg, Optimal Traffic Organization in Ants under Crowded
678: Conditions. {\it Nature} {\bf 248}, 70-73 (2004).
679:
680: \bibitem[5]{Dus05} A.~Dussutour, J.-L.~Deneubourg and V.~Fourcassie,
681: Temporal organization of bi-directional traffic in the ant {\it Lasius
682: niger}. {\it Journal of Experimental Biology} {\bf 208}, 2903-2912 (2005).
683:
684: \bibitem[6]{Gra59} P.-P.~Grass\'e, La Reconstruction du nid et les
685: Coordinations Inter-Individuelles chez Bellicositermes Natalensis et
686: Cubitermes sp. La theorie de la Stigmergie: Essai d'interpretation du
687: Comportement des Termites Construteurs. {\it Insectes Sociaux} {\bf 6},
688: 41-48 (1959).
689:
690: \bibitem[7]{Dor00a} M.~Dorigo, E.~Bonabeau and G.~Theraulaz, Ant
691: algorithms and stigmergy. {\it Future Generation Computer Systems} {\bf
692: 16}(8), 851-871 (2000).
693:
694: \bibitem[8]{Dor04} M.~Dorigo and T.~Stuetzle, Ant colony optimization.
695: MIT Press, Cambridge MA.(2004).
696:
697: \bibitem[9]{Dor97} M.~Dorigo and L.M. Gamberdella, Ant colony system: A
698: cooperative learning approach to the travelling salesman problem. {\it
699: IEEE Trans. Evol. Comput.} {\bf 1}(1), 53-66 (1997).
700:
701: \bibitem[10] {Car98} G.~Di Caro and M.~Dorigo: AntNet, Distributed
702: stigmergetic control for communications networks. {\it Journal of
703: Artificial Intelligence Research} {\bf 9}, 317-365 (1998).
704:
705: \bibitem[11]{Wed04}H.F.~Wedde, M.~Farooq, and Y.~Zhang, BeeHive: An
706: Efficient Fault-Tolerant Routing Algorithm Inspired by Honey Bee
707: Behaviour. {\it Proceedings of ANTS 2004 - 4th International Workshop on
708: Ant Colony Optimization and Swarm Intelligence}, M.~Dorigo et al.({\it
709: ed.}),vol. 3172 of LNCS, pp. 83-94, Springer Verlag, Berlin (2004).
710:
711: \bibitem[12]{Par00}
712: H.V.D.~Parunak: Industrial and Practical Applications of DAI. in, Multiagent
713: Systems, G.~Weiss ({\it ed.}), pp. 377-421, MIT Press, Cambridge MA. (2000).
714:
715: \bibitem[13]{Val01}
716: P.~Valckenaers, H.~Van Brussel, M.~Kollingbaum and O.~Bochmann, Multi-Agent
717: Coordination and Control Using Stigmergy Applied to Manufacturing Control.
718: Lecture Notes in Artifical Intelligence 2086, pp. 317-334, Springer Verlag, Berlin (2001).
719:
720: \bibitem[14]{Blu05}C.~Blum, Beam-ACO---Hybridizing ant colony optimization
721: with beam search: An application to open shop scheduling. {\it Computers
722: \& Operations Research} {\bf32}(6), 1565-1591 (2005).
723:
724: \bibitem[15]{Che99}S.~Chen and K.~Nahrstedt, Distributed
725: quality-of-service routing in ad hoc networks. {\it IEEE Journal on
726: Selected Areas in Communcations} {\bf 17}(8) (1999).
727:
728: \bibitem[16]{Cou03}S.J.~Douglas, D.~De Couto, B.A.~Aguayo and
729: R.M.~Chambers, Performance of multihop wireless networks: shortest path is
730: not enough. {\it ACM SIGCOMM Computer Communication Rev. Arch.} {\bf
731: 33}(1), 83 - 88 (2003).
732:
733: \bibitem[17]{Car05} G.~Di Caro, F.~Ducatelle and L.M.~Gambardella,
734: AntHocNet: An adaptive nature inspired algorithm for routing in mobile ad
735: hoc networks. {\it European Transactions on Telecommunications}
736: {\bf16}(2)(2005).
737:
738: \bibitem[18]{Hel01}D.~Helbing, Traffic and related self-driven
739: many-particle systems. {\it Rev. Mod. Phys.} {\bf 73}, 1067-1141 (2001).
740:
741: \bibitem[19]{Hol55}S.J.~Holt: On the foraging ecology of the wood ant.
742: {\it Journal of Animal Ecology} {\bf 24}, 21-34 (1955).
743:
744: \bibitem[20]{Miu98} T.~ Miura and T.~ Matsumoto, Open-air litter foraging
745: in the nasute termite {\it Longipetermes longipes}. {\it J. Insect Behav.}
746: {\bf 11}, 179-189 (1998).
747:
748: \bibitem[21]{Couz03} I.D.~Couzin and N.R.~Franks, Self -organized lane
749: formation and optimized traffic flow in army ants. {\it Proc. R. Soc.
750: Lond. B} {\bf 270}, 139-146 (2003).
751:
752: \bibitem[22]{Got95}W.H.~Gotwald: Army Ants, the biology of social
753: predation. Cornell Univ. Press, Ithaca, N.Y. (1995).
754:
755: \bibitem[23]{Bur03} M.~Burd and N.~ Aranwela, Head on encounter rates and
756: walking speed of foragers in leaf-cutting ant traffic. {\it Insectes Soc.}
757: {\bf 50}, 3-8 (2003).
758:
759: \bibitem[24]{Hel02} D.~Helbing and T.~Platkowski, Drift or fluctuation
760: induced ordering and self-organization in driven many-particle systems.
761: {\it Europhys. Lett.} {\bf 60} 227-233 (2002).
762:
763: \bibitem[25]{Chowd00} D.~Chowdhury, L.~Santen and A.~Schadschneider,
764: Statistical physics of vehicular traffic and some related systems. {\it
765: Phys. Reports} {\bf 329}, 199-329 (2000).
766:
767: \bibitem[26]{Chowd02}D.~Chowhury, V.~Guttal, K.~Nishinari amd A.~Schadschneider,
768: A cellular-automata model of flow in ant trails: non-monotonic variation of speed with density.
769: {\it J. Phys. A: Math. Gen} {\bf 35}, L573-L577 (2002).
770:
771: \bibitem[27]{Nis03}K.~Nishinari, D.~Chowdhury and A.~Schadschneider,
772: Cluster formation and anomalous fundamental diagram in an ant-trail model.
773: {\it Phys. Rev. E} {\bf 67}, 036120 (2003).
774:
775: \bibitem[28]{Bur02} M.~Burd, D.~Archer, N.~Aranwela and D.J.~Stradling,
776: Traffic dynamics of the leaf-cutting ant {Atta cephalotes}. {\it Am. Nat.}
777: {\bf 159}(3) 283-293 (2002).
778:
779: \bibitem[29]{Den90} J.-L.~Deneubourg, S.~Aron, S.~Gross and J.-M.~Pasteels,
780: The self-organizing exploratory pattern of the Argentine ant. {J. Insect
781: Behav.}{\bf 3} 159-168 (1990).
782:
783: \bibitem[30]{schreck}
784: J.~Wahle, A.~L.~C.~Bazzan,F.~ Kl\"ugl and M.~Schreckenberg,
785: Decision dynamics in a traffic scenario,
786: {\em Physica A} {\bf 287}, 669-681 (2000).
787:
788: \bibitem[31]{Wardrop}
789: J.~G.~Wardrop,
790: Some theoretical aspects of road traffic research,
791: in \textit{Proceedings of the Institution of Civil Engineers II}, Vol. 1 (1952), pp. 325-378.
792:
793: \bibitem[32]{Wolpert}
794: D.~H.~Wolpert and K.~Tumer,
795: Collective intelligence, data routing and Braess' paradox,
796: \textit{Journal of Artificial Intelligence Research} \textbf{16}, 359-387 (2002).
797:
798: \bibitem[33]{Yama}
799: T.~Yamashita,K.~Izumi and K.~Kurumatani,
800: Effect of using route information sharing to reduce traffic congestion,
801: {\em Lecture Notes in Computer Science} {\bf 3012}, 86-104 (2004).
802:
803: \bibitem[34]{m7}
804: B.~Yuan and K.~Chen,
805: Evolutionary dynamics and the phase structure of the minority game,
806: {\em Phys. Rev. E} {\bf 69}, 067106 (2004).
807:
808: \bibitem[35]{socforce1}
809: D.~Helbing and P.~Moln\'{a}r,
810: Social force model of pedestrian dynamics,
811: {\em Phys. Rev. E} {\bf 51}, 4282-4286 (1995).
812:
813: \bibitem[36]{socforce2}
814: D.~Helbing, I.~J.~Farkas and T.~Vicsek,
815: Simulating dynamical features of escape panic,
816: {\em Nature} {\bf 407}, 487-490 (2000).
817:
818: \bibitem[37]{IVC} M.~Sch{\"o}nhof, M.~Treiber, and A.~Kesting,
819: Information propagation in an ad-hoc car-to-car network enhancing
820: road traffic performance,
821: in: J.~Trinckauf (ed.),
822: Proceedings of the 20th {D}resden Conference on
823: Traffic and Transportation Sciences
824: ({D}resden {U}niversity of {T}echnology, 2005).
825:
826: \bibitem[38]{control}
827: D.~Helbing,S.~L\"ammer and J.-P.~Lebacque,
828: Self-organized control of irregular or perturbed network traffic,
829: in C. Deissenberg and R. F. Hartl (eds.)
830: {\em Optimal Control and Dynamic Games}, pp. 239-274, Springer, Dordrecht (2005).
831:
832: \end{thebibliography}
833: \end{document}
834:
835:
836: