Helping ants for adaptive network routing

ARTICLE IN PRESS

Journal of the Franklin Institute 343 (2006) 389–403

0016-0032/$3

doi:10.1016/j

�CorrespoE-mail ad

1Also curr

Berkeley.

www.elsevier.com/locate/jfranklin

Helping ants for adaptive network routing

Azadeh Soltani, M.-R. Akbarzadeh-T�,1, M. Naghibzadeh

Ferdowsi University of Mashhad, Mashhad, Iran

Received 5 February 2006; accepted 6 February 2006

Abstract

Appropriate routing in data transfer is a challenging problem that can lead to improved

performance of networks in terms of lower delay in delivery of packets and higher throughput.

Considering the highly distributed nature of networks, several multi-agent based algorithms, and in

particular ant colony based algorithms, have been suggested in recent years. However, considering

the need for quick optimization and adaptation to network changes, improving the relative slow

convergence of these algorithms remains an elusive challenge. Our goal here is to reduce the time

needed for convergence and to accelerate the routing algorithm’s response to network failures and/or

changes by imitating pheromone propagation in natural ant colonies. More specifically, information

exchange among neighboring nodes is facilitated by proposing a new type of ant (helping ants) to the

AntNet algorithm. The resulting algorithm, the ‘‘modified AntNet,’’ is then simulated via NS2 on

NSF network topology. The network performance is evaluated under various node-failure and node-

added conditions. Statistical analysis of results confirms that the new method can significantly reduce

the average packet delivery time and rate of convergence to the optimal route when compared with

standard AntNet.

r 2006 The Franklin Institute. Published by Elsevier Ltd. All rights reserved.

Keywords: Ant colony; Network routing; AntNet algorithm; Multi-agent systems; NSFNet

1. Introduction

Routing algorithms, at the core of network control systems, play an important role inthe exponentially growing communication systems worldwide. If appropriately configured,

0.00 r 2006 The Franklin Institute. Published by Elsevier Ltd. All rights reserved.

.jfranklin.2006.02.007

nding author.

dresses: [email protected] (A. Soltani), [email protected] (M.-R. Akbarzadeh-T).

ently a visiting scholar with Berkeley Initiative on soft computing (BISC) at University of California,

www.elsevier.com/locate/jfranklin

ARTICLE IN PRESSA. Soltani et al. / Journal of the Franklin Institute 343 (2006) 389–403390

they can provide faster, smoother and more reliable data packet routing, and in particulargreatly influence several measures of network performance such as end-to-end delay andthroughput. Conventional routing algorithms depend on global exchange of informationamong nodes and hence become impractical as network size increases. These algorithmsare based on finding shortest path such as in distance-vector algorithms (RIP1) [1] andlink-state algorithms (OSPF2) [2]. However the rapid growth of networks and change oftechnology show that we need to reconsider our approach to routing in order to respond tothe new system requirements such as rapid convergence, quick response to networkchanges and quality of service as well as having good performance.Communication networks are distributed platforms with distributed information, and

therefore provide good environment for multi agent systems and distributed decisionmaking. Consequently, in the last decade, new types of routing algorithms based on multiagent systems have been introduced. Amin A. and Mayes, J.T. (1998) introduced a newagent-based distance vector routing (ADVR) algorithm based on multi agent systems [3].Traditional DVR algorithms cause large network overhead due to the large number ofmessages generated through the router update process. Number of these messages is anexponential function of the number of nodes in the network. In comparison, since ADVRis an agent based solution, the number of its messages is bounded by the number of agentsin the network.Most of the other multiagent-based algorithms take their inspiration from ants’

behavior. Real ants are able to find shortest path between their nest and food source byfollowing pheromone trail of other ants. For example Schonderwoerd R. (1997)implemented ant-like agents for routing [4]. In his algorithm, each source node s sendsan ant toward destination d at regular intervals, where d is selected in a random scheme.When ants reach node i, select their next hop n to their destination according to routingtable of node i, then update node i’s routing table. They increase the probability ofchoosing n as a next hop (increasing the pheromone) while selection probability of otherneighbors is decreased for destination d. Di Caro and Dorigo (1997) also introduced a newalgorithm based on ant behavior (AntNet) [5]. In their method two types of ants exist,forward and backward ants. Similar to the Schonderwoerd’s algorithm, each node sends aforward ant to different destinations at regular intervals. But in this algorithm the forwardant does not update the routing tables of nodes that it visits, its only duty is to find a pathto destination d and to collect information. When a forward ant arrives at its destination d,it generates a backward ant and dies. The backward ant then goes back in the same path asthe forward ant that created it and updates the routing tables for intermediate anddestination nodes.The above AntNet has received considerable attention by various researchers. For

example, B. Baran and R. Sosa improved AntNet by proposing an intelligent initializationof routing tables, an intelligent update after network resource failures, and a noisy decisionmaking against undesirable networks ‘‘freezing’’ their routing probabilities in dynamicenvironment [6]. Later, in 2002, Kassabalidis and El-Sharkawi, M.A. showed that for largenetworks good routing solutions can be achieved by combined use of network clustering,autonomous systems and ant colony [7]. In other research, AntNet is used in routing for adhoc wireless networks. For example, Marwaha, S., et al. introduced a novel method for

1Routing Information Protocol.2Open Shortest Path First.

ARTICLE IN PRESSA. Soltani et al. / Journal of the Franklin Institute 343 (2006) 389–403 391

routing based on AntNet and AODV3 [8]. AntNet is also used in QoS routing by Subingand Zemin [9] among others.

As several of the above research have pointed out, while strong in regards to distributedrouting, Standard AntNet still has a weakness in terms of speed of convergence andresponse to network changes/failures. While the gradual process of pheromoneaggregation and evaporation allows AntNet to reach its globally optimal performancewith only distributed information, the inability of nodes to directly share knowledgeamong each other may cause the algorithm to respond too slowly to the rapid networkchanges. In other words, AntNet’s strength in a totally indirect and distributedinformation sharing may be its own fallacy as well. This is while, in addition toaggregation and evaporation, natural ant colonies do have a method of pheromonepropagation as a method of sharing information among neighboring nodes. Some of therecent research by authors in [10,11] as well as by Di Caro in [12,13] have attended to thisaspect of natural ant colonies.

Specifically, authors in [10,11] introduced a new type of helping ants to increasecooperation among neighboring nodes, thereby reducing AntNet algorithm’s convergence.This concept of helping ants, from a naturalistic perspective, is inspired by the fact thatinsect coordination via pheromones relies on at least three dynamics [14]: aggregation ofsuccessive deposits by different ants, evaporation of pheromone to discard obsoleteinformation, and propagation of pheromone from one location to other nearby locationsto share information. ACO uses aggregation and evaporation, but does not account forpheromone propagation. In the proposed algorithm, this natural process of pheromonepropagation is considered by introducing a novel type of helping ants. From a practicalengineering perspective, the present paper will further consider and compare the addednumber of generated ants as a measure of generated overhead for this new algorithm.Statistical analysis reveals that the algorithm provides superior performance not because ofthe amount of added information (overhead) but as a result of an enhanced method ofdistributing information.

This paper is organized as follows. After a brief overview of the Standard AntNetalgorithm, the proposed helping ants and the Modified AntNet algorithm are discussed inSection 2. In Section 3, the experimental settings and the metrics of performanceevaluation are discussed. Results of six different experiments are illustrated in Section 4.Statistical analysis of these experiments, obtained through NS2 simulation environment, inSection 5 shows the performance comparison under various network test conditions asapplied to NSF network.

2. The proposed helping ants for AntNet (Modified AntNet)

Ants have a natural ability to find shortest path between their colony and sources offood, while communicating indirectly (stigmergy) and using only pheromones dropped byother ants. Since this path finding method is based only on local information, it is deemeda promising paradigm for distributed network routing as well. AntNet [15] is a distributedmulti-agent algorithm that is inspired by this routing mechanism as observed in naturalcolonies of ants.

3Ad hoc on demand distance vector.


In AntNet, each node works independently by sending many ants through the networkin order to find best paths for data packets that it initiates. Each two neighbors aretherefore independent by design and cannot use each other’s experience. While this methodmay provide the benefit of lower overhead and prevent propagation of inaccurateinformation, it also causes an undesirably late convergence. In contrast, earlierconventional algorithms seem to promote global exchange of network information forquick and optimal path selection. In such conventional techniques, each node gets thisinformation from its neighbors. However, exchange of such enormous amount ofinformation increases overhead. Also, every neighbor’s action depends on other nodes,which means that their response to changes and faults of the network may also not bevery rapid.The proposed Modified AntNet algorithm is a hybrid approach that tries to take

advantage of strong features of the above two methodologies while aiming to avoidtheir weaknesses. To achieve this goal, we introduce a new type of helping ant to imitatethe mechanism of information sharing among nearby ants by pheromone propagation intheir natural colonies. Modified AntNet has therefore three kinds of agents: forward,backward and helping ants. Similar to AntNet, each forward ant explores the networkand collects information while moving towards a certain destination. Upon arriving at itsdestination, this forward ant generates a backward ant, before dying. Also similarly toAntNet, the backward ant travels back, on the same path as the forward ant thatcreated it, and updates the routing tables on its way. In addition to these typicalforward and backward ant types in AntNet, Modified AntNet also has helping antswhose duty is to distribute routing information to source’s neighboring nodes. Whenthe backward ant reaches a source node s, it creates helping ants and sends them to theneighbors of s to inform them about the new path ðs! dÞ. Helping ants updatethe neighboring nodes’ tables n (similar to backward ant) and die (will not propagatefurther). As will be shown here, adding this type of ant helps nodes utilize theirneighbor’s experience, as nodes have more information to make better decisions.Moreover in the Modified AntNet, helping ants are intelligent ants that, withoutpaying attention to the previous routing table, bring information to the neighboringnodes.Two data structures reside at each network node k (the network is considered to have N

nodes and each node k has Nk neighbors), as follows:

(i)
An array Mk ¼ ½ðmj ;s2j ;Tbestj Þ�; j ¼ 1 . . .N, where mj is the mean, s2j is the variance and
Tbestj is the best trip time to destination j, from node k.
(ii) A routing table Pk which has N rows and Nk columns. It stores for each pair ðd; nÞ the
probability pdn, which is the goodness of choosing node n as the next hop for thedestination d. These probabilities are normalized such that,

Xn2Nk

pdn ¼ 1; d 2 ½1 . . .N�; n 2 ½1 . . .Nk�. (1)

Modified AntNet Steps:

�
At regular intervals each network node s sends a new forward ant Fs!d to destination d.d is selected according to the data traffic pattern generated by local workload, with the


following probability,

probabilityd ¼f sdPN

d 0¼1 f sd 0, (2)

where f sd is the local traffic sent from s to d (in bits or number of packets).
� The identifier of every visited node k and the time Ts!k elapsed from its launch timeuntil it arrives at node k are pushed onto a memory stack Ss!dðkÞ that is carried by theforward ant F s!d . � At node k, a forward ant Fs!d chooses its next node from among the node k’s neighborsthat it has not visited already. If all of them have been previously visited, the selectionwill be based on the routing table and the size of queues, with the following probability,
p0nd ¼pnd þ aln

1þ aðNk � 1Þ, (3)

where ln is the availability factor, as calculated below

ln ¼ 1�qnPNk

n0¼1 qn0

, (4)

where qn is the length of the queue of messages to be forwarded from node k to itsneighbor node n, and 0:2 � a � 0:5. This modified probability helps choose the nextnode, according to the routing table as well as network traffic state.
� When the forward ant F s!d reaches its destination node d, it generates a backward ant
Bd!s, transfers all of its memory to the backward ant, and dies.
� The backward ant Bd!s goes back through the same path as the forward ant F s!d . Ateach node k along the path it pops its stack Ss!dðkÞ to determine the next hop node. � When a backward ant reaches node k (coming from a neighbor node f), updates Mk and
Pk for d, as well as all destinations between k and d in Ss!dðkÞ, if the elapsed trip time isstatistically ‘‘good’’.� The sample mean and variance of the model Mk are updated with the trip time ok!d

stored in the stack memory. Di Caro used arithmetic, exponential and windowedstrategies to compute the statistics. He reported better result with the method usedhere, however he also reported that changing strategy does not affect performancesignificantly.

md ¼ md þ Zðok!d � mdÞ, (5)

s2d ¼ s2d þ Zððok!d � md Þ2� s2dÞ, (6)

Z ¼ cð5=WmaxÞ; co1.

The factor Z is the learning rate and weighs the number of most recent samples thatwill affect average, and Wmax is the maximum allowed size of the observationwindow.� In the routing table, pfd is increased while probabilities of other neighbors aredecreased.

pfd ¼ pfd þ rð1� pfd Þ, (7)

pnd ¼ pnd � rpnd ; n 2 Nk; naf . (8)


Factor r that is used to change the probabilities is obtained according to thefollowing formula:

r ¼ c1Tbest d

T

� �þ c2

T sup � T inf

ðT sup � T inf Þ þ ðT � T inf Þ

� �. (9)

Several functions using linear, quadratic and hyperbolic combinations of Mk and Pk

are tested and above formula is selected [14]. The empirical factors c1 and c2 arefound to be optimum for the values 0.3 and 0.7, respectively. T is elapsed time,T inf ¼ Tbest d , T sup ¼ md þ zðsd=

ffiffiffiffiffiffiwdpÞ and wd is the moving observation window,

after each new sample wd is incremented modulus wmax, z ¼ 1=ffiffiffiffiffiffiffiffiffiffiffi1� gp

, and g is theconfidence interval.

�
When the backward ant reaches s, updates Ps and Ms before dying. � If the new trip time DTsd ¼ td � ts is good, i.e. DTsdomd , s generates helping ants, puts itsidentifier, DTsd and ts (current time at s) in their stacks, and sends them to its neighbors ðniÞ. � When the helping ant reaches neighboring node i, it computes DTis ¼ ti � ts. Therefore,assuming the queuing time is negligible, the new trip time ðDTidÞ through the neighbor s
can be estimated as follows,

DTid ¼ DTis þ DTsd . (10)

�
If the trip time is good, i.e. DTidomd , the helping ant updates the data structure inneighbor node i with the new trip time is DTid for next hop s.
psd ¼ psd þ rð1� psd Þ, (11)

pnd ¼ pnd � rpnd ; n 2 Ni; nas, (12)

r ¼ c1Tbest

DTid

� �þ c2

T sup � T inf

ðT sup � T inf Þ þ ðDTid � T inf Þ

� �. (13)

3. Experimental setting

3.1. Topology

To evaluate the proposed method, we simulated 6 scenarios. In these experiments weused NSFNet network topology, similar to that of [14]. NSFNet is the old USA T1backbone (1987). It is a WAN with 14 nodes and 21 bi-directional links. The linkbandwidth is 1.5Mb/s and link delays range from 4 to 20ms. This is a well balancednetwork as shown in Fig. 1.

3.2. Metrics of performance evaluation

In order to compare the results of this approach with that of Di Caro and Dorigo (1998),the following metrics are chosen:

End-to-end delay: The length of time required to move a packet from source to destination.Clearly, if a network converges to its best routes quicker, average delay will be reduced.

ARTICLE IN PRESS

1

0

13

23

1211

109

4

5

6

7

8

Fig. 1. NSFNet.

A. Soltani et al. / Journal of the Franklin Institute 343 (2006) 389–403 395

Jitter: Variance of delay. This metric shows faster converging routines with smallerjitter. In other words, after convergence, packet usually will be sent through the goodroutes with higher consistency, i.e. smaller change in delay.

Loss rate: The percent of packets which are lost. This metric distinguishes routines thatsend packets through invalid paths, or paths which have loops or faulty links.

Overhead: The number of ant packets sent in a given network (in this paper).As mentioned before, convergence has a strong effect on above metrics. So for

evaluating the rapidness of convergence we compare these metrics in our experiments.

4. Results

Due to statistical nature of the test benchmark, each of the below experiments wereperformed 10 times, results were averaged, and compared statistically. In this section weobtain several hypotheses based on average of results. Mann–Whitney test is then appliedin Section 5 to analyze these hypotheses.

Experiment 1: CBR4 traffic is transmitted from node 7 to node 3 in NSFNet where thereare several good alternative paths for packet routing. AntNet can show its ability to findalternative routes better when traffic rate is higher than link bandwidth. Therefore, weconsider our network to be under similar conditions, i.e. packet size is 256 bytes and meanpacket inter-arrival time is 0.001 s and so transfer rate is approximately 2.05Mb/s(256*8*1000). The simulation runs for 120 s with the initial 20 s for priming the network.Immediately after the first 45 s of simulation is past, link between nodes 5 and 7 is broken,and in the 85th second of the simulation it is reconnected. Ant generation rate is equal forboth algorithms. This experiment compares the behavior of the two algorithms to thechanging states of the network (Fig. 2).

The right and left tables of Table 1 illustrate the network performance with the ModifiedAntNet and Standard AntNet, respectively. Furthermore, the columns in each table showthe behavior of the network before the link failure (20–45 s) during the failure (45–85 s) andafter the link comes back up (85–120 s). As can be seen, Modified AntNet shows a betterperformance over the Standard AntNet in this experiment. Specifically, the ModifiedAntNet reduces mean delay by 6.2%, 4.2% and 3% in above three states, respectively.

4Constant Bit Rate.

ARTICLE IN PRESS

Average of End to End Delay (Modified AntNet) Average of End to End Delay (Standard AntNet)0.18

0.16

0.14

0.12

0.1

0.08

0.06

0.04

0.02

Del

ay (

sec)

20 40 60 80 100 120Simulation Time (sec)

0.18

0.16

0.14

0.12

0.1

0.08

0.06

0.04

0.02

Del

ay (

sec)

20 40 60 80 100 120

Simulation Time (sec)

Fig. 2. Average of End-to-End Delay in each instant (Experiment 1).

Table 1

Average of results (Experiment 1)

20–45 45–85 85–120

Modified AntNet

Mean of delay 0.095485 0.100900 0.107837

Mean of jitter 0.000366 0.000443 0.000495

#Received packets 18485.5 28714.4 25901.1

Loss rate 25.84% 25.96% 26.00%

#Ant packets 9265.4 7542.0 6802.0

Standard AntNet

Mean of delay 0.101824 0.105405 0.111093

Mean of jitter 0.000272 0.001354 0.000477


Loss rate 26.47% 27.76% 26.77%

#Ant packets 7983.0 6520.8 5635.2

A. Soltani et al. / Journal of the Franklin Institute 343 (2006) 389–403396

It also has lower jitter. During the failure (45–85 s), jitter is reduced by 67%. This is whileincreased overhead (# Ant Packets) may appear to be a drawback of the proposedModified AntNet.

Experiment 2: To investigate whether the improved performance is due to a betterdistribution of information in the proposed algorithm or merely as a result of increasedamount of information distribution, i.e. increased overhead, the previous experiment isrepeated while the forward ant generation rate is reduced by a factor of 50% in theModified AntNet as compared to that of Standard AntNet (Fig. 3).According to Table 2, the behavior of both algorithms is approximately equal, while the

overhead in the Modified AntNet is significantly lower.Experiment 3: In this experiment, routing performance is tested on a different part of the

network where there are not many good alternative paths for packet routing. CBR traffic istransmitted here from node 12 to node 9 in NSFNet. Packet size is again 256 bytes@ 2.05Mb/s. The simulation is run for 120 s with initial 20 s for priming the network.Right after 45 s of the simulation is passed, the link 9–10 is intentionally broken, and in the

ARTICLE IN PRESS

Average of End to End Delay (Modified AntNet) Average of End to End Delay (Standard AntNet)0.18

0.16

0.14

0.12

0.1

0.08

0.06

0.04

0.02

Del

ay (

sec)


0.18

0.16

0.14

0.12

0.1

0.08

0.06

0.04

0.02

Del

ay (

sec)



Table 2


20–45 45–85 85–120

Modified AntNet

Mean of delay 0.094334 0.105038 0.114216

Mean of jitter 0.000182 0.000915 0.000526


Loss rate 26.52% 26.67% 26.23%

#Ant packet 4662.5 4027.2 3287.1

Standard AntNet

Mean of delay 0.101824 0.105405 0.111093

Mean of jitter 0.000272 0.001354 0.000477


Loss rate 26.47% 27.76% 26.77%

#Ant packet 7983.0 6520.8 5635.2


85th second of the simulation, it is reconnected. This experiment is similar to experiment 1,but occurs between nodes 12 and 9 (Fig. 4).

Table 3 shows that, during the failure (45–85 s), in the Modified AntNet the mean delayand jitter is improved by 12% and 26%, respectively. Also, after the failure (85–120 s)Modified AntNet reduces delay and jitter by about 4% and 37%, respectively. Moreover,packet loss rate in the Modified AntNet is less than the Standard AntNet (by about 8.3%and 9.5%).

Experiment 4: Experiment 3 is performed again and ant generation rate in the ModifiedAntNet rate is considered to be 1

2of ant generation rate in the Standard AntNet (Fig. 5).

Table 4 shows that, during the failure (45–85 s), even with reduced ant generation rate,Modified AntNet still reduces mean delay and jitter about 16% and 45%, respectively.Also after failure (85–120 s) Modified AntNet has better behavior. Moreover, packet lossrate and overhead in Modified AntNet are less than Standard AntNet.

Above experiments show that the performance of the Modified AntNet is furtherimproved when paths are longer or when alternative paths are fewer. In other words, for

ARTICLE IN PRESS

Table 3


20–45 45–85 85–120

Modified AntNet

Mean of delay 0.096525 0.115156 0.098921

Mean of jitter 0.000480 0.002921 0.000424


Loss rate 25.64% 34.60% 24.78%

#Ant packets 9080.8 8691.5 6081.5

Standard AntNet

Mean of delay 0.095014 0.130984 0.103118

Mean of jitter 0.000963 0.003984 0.000674

Received packets 18613.5 24983.2 25517.6

Loss rate 25.35% 37.49% 27.14%

#Ant packets 7808.2 7627.4 4795.8

Average of End to End Delay (Modified AntNet)Average of End to End Delay (Standard AntNet)

0.180.200.22

0.160.140.120.1

0.080.060.040.02

Del

ay (

sec)

20 40 60 80 100 120


0.180.2

0.22

0.160.140.120.1

0.080.060.040.02

Del

ay (

sec)

20 40 60 80 100 120



Average of End to End Delay (Modified AntNet) Average of End to End Delay (Standard AntNet)

0.180.200.22

0.160.140.12

0.10.080.060.040.02

Del

ay (

sec)


0.180.2

0.22

0.160.140.12

0.10.080.060.040.02

Del

ay (

sec)

20 40 60 80 100 120




ARTICLE IN PRESS

Table 4


20–45 45–85 85–120

Modified AntNet

Mean of delay 0.092162 0.109695 0.100692

Mean of jitter 0.000243 0.002164 0.000803


Loss rate 25.71% 36.76% 25.88%

#Ant packets 4644.1 4782.0 3017.8

Standard AntNet

Mean of delay 0.095014 0.130984 0.103118

Mean of jitter 0.000963 0.003984 0.000674


Loss rate 25.35% 37.49% 27.14%

#Ant packets 7808.2 7627.4 4795.8

Average of End to End Delay (Modified AntNet) Average of End to End Delay (Standard AntNet)

0.18

0.20

0.16

0.14

0.12

0.1

0.08

0.06

0.04

0.02

Del

ay (

sec)

20 40 60 80 100 120


0.18

0.20

0.16

0.14

0.12

0.1

0.08

0.06

0.04

0.02

Del

ay (

sec)




longer paths, nodes need more time for convergence and helping ants can be more effectivein reducing this time.

Experiment 5: The goal of this experiment is to evaluate the speed of finding a new nodein the NSFNet. CBR traffic is transmitted from node 7 to node 3. Packet size is again 256bytes @ 2.05Mb/s. The simulation was run for 120 s with initial 20 s for priming thenetwork. At first, the link 5–7 does not exist in the network and in the 50th second of thesimulation it is added to the network (ant generation rate in the Modified AntNet rate isconsidered to be 1

2of ant generation rate in the Standard AntNet) (Fig. 6).

Table 5 shows that, the Modified AntNet has a better performance. It reduces the delayby 20% and 13%, and jitter by 68% and 73%, respectively. It also has lower overhead andpacket loss rate.

In above Experiments 1–5, only two nodes, one node for sending and another forreceiving the data traffic, are selected at any given time. Above experiments reveal that thedistance and variety of alternate routing paths between these two nodes is very important

ARTICLE IN PRESS

Table 5


20–50 50–120

Modified AntNet

Mean of delay 0.099682 0.101489

Mean of jitter 0.000481 0.000340

#Received packets 21802.1 51728.5

Loss rate 26.89% 26.09%

#Ant packets 4903.5 6252.7

Standard AntNet

Mean of delay 0.124919 0.116574

Mean of jitter 0.001542 0.001299

#Received packets 21699.8 51216.7

Loss rate 27.43% 26.85%

#Ant packets 7871.8 10295.3

0.16

0.15

0.14

0.13

0.12

0.11

0.1

dela

y (s

ec)

20 30 40 50 60

simulation time (sec)

0.16

0.15

0.14

0.13

0.12

0.11

0.1

dela

y (s

ec)

20 30 40 50 60

simulation time (sec)

end to end dealy (Standard AntNet)end to end dealy (Modified AntNet)



and can effectively influence individual results. This is necessary in order to isolate theinfluence of the helping ants in improving various types of single routing tasks. However,in order to study the overall effect of helping ants in better information distribution andimproved nodal coordination/collaboration, Experiment 6 is devised in which all nodessimultaneously send/receive.

Experiment 6: In this experiment all nodes are actively sending/receiving nodes andupdate their routing tables simultaneously. Unlike previous experiments 1–5, we are notconcerned with individual node/link failure; we are instead concerned with the routingperformance under normal network conditions. After the initial 20 s for reaching steadystate, each node transfers CBR traffic to all other nodes in network for the next 40 s.Packet size is 256 bytes @ 0.205Mb/s. (ant generation rate in the Modified AntNet rate isconsidered to be 1

2of ant generation rate in the Standard AntNet.)

Fig. 7 illustrates an example run for both Modified AntNet (left) and the standardalgorithm (right). Table 6 lists the overall mean delay, mean jitter and overall packet loss

ARTICLE IN PRESS

Table 6


20–60

Modified AntNet

Mean of delay 0.127360

Mean of jitter 0.007688

#Received packets 418635

Loss rate 0.4237

#Ant Packets 45976

Standard AntNet

Mean of delay 0.132145

Mean of jitter 0.007818

#Received packets 412332

Loss rate 0.4324

#Ant Packets 73527


rate for both the Modified AntNet as well as the standard algorithm. From the abovetable, it can be inferred that the proposed algorithm reduces end-to-end delay withsignificantly fewer ant packets (38% fewer).

5. Statistical analysis

Since the statistical distribution of the population may not be normal, Mann Whitneytest [16] is used here for statistical comparison of the proposed approach versus theconventional. The null hypothesis in Mann–Whitney test determines if means of twopopulations have no statistically significant difference; in other words, null hypothesisH0 : m1 � m2 ¼ 0 or m1 ¼ m2.

As shown in Table 7, above statistical analysis reveals that the proposed algorithmsignificantly reduces mean packet delay during link failures in both experiments 1 and 3,i.e. two different routes. Furthermore, this difference may remain significant even when theant generation rate (overhead) for the proposed algorithm is reduced to 50% of standardalgorithm as in Experiments 4. The proposed algorithm, however, loses it superiorperformance, i.e. reducing packet delay with reduced overhead, when there are severalalternative and equally good paths as in Experiment 2. In this experiment, a betterdistribution of information does not result in better performance as there are several goodpaths, and therefore both algorithms find good path equally fast, i.e. the problem is tootrivial. Considering Experiments 1–4, therefore, it can be concluded that Modified AntNetdecreases the delay after failure particularly when good routes are few.

Analysis of Experiments 5 and 6 further confirm the superiority of the proposedalgorithm when a new link is introduced to the network or when all nodes are activelysending and receiving. Statistical test of significance confirms improved packet deliverytime under both test conditions. Furthermore, analysis of Tables 8 and 9 reveals that theproposed method either significantly reduces jitter and loss rate under some routingconditions or does not significantly alter their behavior. This conclusion is made whileoverhead is significantly reduced for the proposed algorithm in Experiments 2 and 4.

ARTICLE IN PRESS

Table 8

Probability of the difference of jitter occurring as a result of chance. The difference is significant if po0:05

20–45 s 45–85 s 85–120 s

Experiment 1 0.0749 0.0268 0.3669

Experiment 2 0.3372 0.121 0.3974

Experiment 3 0.0351 0.1357 0.2148

Experiment 4 0.3372 0.0228 0.3372

Experiment 5 0.0446 0.0003

Experiment 6 0.3669

Table 9

Probability of the difference of loss rate occurring as a result of chance. The difference is significant if po0:05

20–45 s 45–85 s 85–120 s

Experiment 1 0.0559 0.017 0.0129

Experiment 2 0.4562 0.1736 0.0351

Experiment 3 0.2236 0.2611 0.0268

Experiment 4 0.2483 0.3121 0.121

Experiment 5 0.0808 0.0375

Experiment 6 0.2358

Table 7

Probability of the difference of delay occurring as a result of chance. The difference is significant if po0:05

20–45 s 45–85 s 85–120 s

Experiment 1 0.2148 0.0375 0.2358

Experiment 2 0.0188 0.4562 0.484

Experiment 3 0.484 0.0228 0.0934

Experiment 4 0.2148 0.0057 0.3121

Experiment 5 0.0104 0.0228

Experiment 6 0.0322


6. Conclusion

Pheromone propagation, in addition to pheromone aggregation and evaporation, playsa key role in natural ant colonies. The proposed algorithm introduces helping ants toimitate this natural process of information sharing to improve an existing network routingtechnique in terms of speed of convergence. Here, we consider mean delay as the mostimportant parameter for network convergence in general, and NSFNet in particular. Theproposed modification demonstrates a method by which mean delay and networkoverhead can be significantly reduced without a higher loss rate or jitter. The severalexperiments, as reported in this paper, indicate an enhanced method of informationdistribution resulting in faster network convergence with smaller overhead, and otherimproved performance metrics such as delay, jitter, and throughput. Statistical analysisreveals that the proposed Modified AntNet achieves better behavior under differentconditions of the network. For example, when a network is disturbed by a certain link


going down and coming back up, the algorithm reduces the packet delay rate by reducingthe time needed for convergence. Experiments also indicate that the performance of theModified AntNet is significantly improved when paths are long or alternative paths arefew. Furthermore, under normal state of network operation, i.e. when all nodes areactively engaged in sending/receiving information, the Modified AntNet significantlyreduces the overall mean end-to-end delay.

References

[1] C. Hedrick, Routing Information Protocol RFC 1058, June 1998.

[2] J. Moy, OSPF version2, RFC 1247, July 1991.

[3] A. Amin, A.R. Mikler, Agent-based distance vector routing: a resource efficient scalable approach to routing

in large communication networks, J. Systems Software 71 (3) (2004).

[4] R. Schoonderwoerd, O. Holland, J. Bruten, Ant-like agent for load balancing in telecommunications

networks, in: Proceeding of the First International Conference on Autonomous Agents, February 1997,

Marina del Rey, CA, pp. 209–216.

[5] G. Di Caro, M. Dorigo, Mobile agent for adaptive routing, Technical report IRIDIA, 1997.

[6] B. Baran, R. Sosa, AntNet routing algorithm for data networks based on mobile agents, Inteligencia

Artificial, Revista Iberoamericana de Inteligencia Artificial 12 (2001) 75–84.

[7] I. Kassabalidis, M.A. El-Sharkawi, R.J. Marks, P. Arabshahi, A.A. Gray, Adaptive-SDR: adaptive swarm-

based distributed routing, Proceedings of the 2002 International Joint Conference on Neural Networks, vol.

1, Honolulu, HI, 2002, pp. 351–354.

[8] S. Marwaha, C. khong Tham, D. Serinvasan, A novel routing protocol using mobile agents and reactive

route discovery for ad hoc wireless networks, hhttp://www.ece.nus.edu.sg/stfpage/eletck/papers/

ShivaICON_conf8a51.pdfi.

[9] Z. Subing, L. Zemin, A Qos routing algorithm based on ant algorithm, Proceedings of the 25th Annual IEEE

Conference on Local Computer Networks 2000, LCN 2000, Tampa, FL, pp. 574–578.

[10] A. Soltani, M.-R. Akbarzadeh-T, M. Naghibzadeh, Introducing helping ants in AntNet and implementation

on NSFNet, (in Farsi) Proceedings of the 2004 Iranian Conference on Electrical Engineering, May 2004,

Mashhad, Iran.

[11] M.-R. Akbarzadeh-T, A. Soltani, M. Naghibzadeh, A novel AntNet-based algorithm with application to

NSFNET, Proceedings of the 2005 International Conference on Modeling, Simulation and Optimization,

February 2005, Sharjah, UAE.

[12] G. DiCaro, F. Ducatelle, L.M. Gambardella, Swarm intelligence for routing in mobile ad hoc networks, in:

Proceedings of the IEEE Swarm Intelligence Symposium, Pasadena, USA, June 8–10, 2005.

[13] G. Di Caro, F. Ducatelle, L.M. Gambardella, AntHocNet: an ant-based hybrid routing algorithm for mobile

ad hoc networks, Proceedings of PPSN VIII—Eight International Conference on Parallel Problem Solving

from Nature, Birmingham, UK, September 18–22, Lecture Notes in Computer Science, vol. 3242, Springer,

Berlin, 2004.

[14] S. Brueckner, Return from the ant: synthetic ecosystems for manufacturing control, Dissertation in

Computer Science. Humboldt University Berlin, Berlin, Germany, 2000.

[15] G. DiCaro, M. Dorigo, AntNet: Distributed stigmergic control for communications networks, J. Artif. Intell.

Res. 9 (1998) 317–365.

[16] C. Spatz, J.O. Johnston, Basic Statistics, Brooks/Cole Publishing Company, Pacific Grove, CA, 1976.

http://www.ece.nus.edu.sg/stfpage/eletck/papers/ShivaICON_conf8a51.pdf

http://www.ece.nus.edu.sg/stfpage/eletck/papers/ShivaICON_conf8a51.pdf

Documents

Helping ants for adaptive network routing