Page 1
1
Ant based probabilistic routing with pheromone and antipheromone mechanisms
(Manuscript to be submitted to the IEEE Transactions on Evolutionary Computation
Special Issue on Ant Algorithms and Swarm Intelligence)
Harilaos G. Sandalidis, Kostas Mavromoustakis and Peter Stavroulakis
Telecommunication Systems Institute of Crete (T.S.I.),
Iroon Polytehniou 37 str.,
73133 Chania, Crete, Greece.
Tel.: +30-821-28457, Fax: +30-821-28459
E-mail: [email protected] , [email protected]
Contacting author:
Prof. Peter Stavroulakis
Telecommunication Systems Institute of Crete (T.S.I.),
Iroon Polytehniou 37 str.,
73133 Chania, Crete, Greece.
Tel.: +30-821-28457, Fax: +30-821-28459
E-mail: [email protected]
Keywords: Adaptive routing, Ant system, Probabilistic routing, Pheromone andantipheromone principles
Page 2
2
Ant based probabilistic routing with pheromone andantipheromone mechanisms
Harilaos G. Sandalidis, Kostas Mavromoustakis and Peter Stavroulakis
Abstract−− Using the idea of probabilistic routing, calls in an ant based decentralized
scheme are not routed according to the largest probabilities in the pheromone tables but
randomly according to these probabilities. This principle can be particularly helpful in
order to further minimize possible node congestion problems. An additional
incorporation of the antipheromone mechanism in the operation of artificial ants helps
in better biasing the network. This paper examines the behavior of such a routing
scheme using a proper set of suitable metrics.
Index Terms−− Adaptive routing, Ant system, probabilistic routing, Pheromone and
antipheromone principles
I. INTRODUCTION
The objective of routing is to establish a successful connection between any two nodes in a
communication network maximizing network performance. The way that routing is
implemented influences several important network characteristics such as grade of service, cost
reliability etc. Roughly speaking, a good routing scheme has to be computationally efficient
and inexpensive to apply, capable to select the best route depending on various constraints, and
perform correctly in the face of unusual or unforeseen circumstances. Having all the above in
mind it becomes evident that the problem of finding a proper route is not an easy task. Real
networks consist of a large number of nodes and show to have a dynamically unpredictable
behavior mainly because traffic load is not uniformly distributed. It is therefore clear that the
Page 3
3
algorithm used for this task has to avoid overloaded nodes and activate alternative routes
whenever load conditions are drastically changing.
A lot of different routing schemes appeared in the literature and the most basic are
discussed in [1] and [2]. The majority of these mentioned schemes are suitable in finding short
routes according to defined optimization criteria but decisions are usually made off-line (non-
adaptive). Adaptive schemes are in principle more attractive because they can keep abreast of
any possible changes in network load and failures. Routing algorithms can be also classified
into centralized and decentralized. Centralized schemes allow a central controller to act as the
supervisor of the entire network. Although these routing methods seem to be quite attractive,
their application to real networks is time inefficient and algorithmically complex and therefore
they are used only in particular cases. Decentralized schemes involve a number of controllers
each of which supervises a part of a network, and consist of a modern alternative approach to
routing problems.
Heuristic techniques and especially computational intelligence (neural networks, fuzzy logic
etc.) have been extensively used in the past in the area of wireline communications mainly
because they can give fast and quite reliable solutions to difficult problems (routing, call
admission switching etc) [3]. However, the majority of them require the use of an explicitly
well defined cost function whose minimization leads to a sub-optimal route.
Swarm intelligence is a new challenging branch of artificial life which takes advantage of the
collective behavior of animals with limited intellectual faculties (insects, flocking birds, schools
of fish) to solve algorithmically complex problems. The way social insects behave can be
applied to implement effective decentralized routers without the use of such a function. In a
seminal work by Dorigo et al. [4], [5], [6], intelligent agents in the means of artificial ants
were applied for shortest path searching on constrained graphs. Artificial ants are laying small
Page 4
4
pieces of information (analogous to chemical pheromone trails laid by natural ants) to network
nodes. Using this kind of information, ants communicate with each other and they can find
short routes from their nest to a food source and vice versa. Based on similar principles,
Schoonderwoerd et al. [7], [8] proposed an ant-based routing model for wireline
communication networks whereas Di Caro et al. presented AntNet, a much more realistic
application to Internet routing [9], [10].
In the Schoonderwoerd et al. model, artificial ants lay artificial pheromones on network
nodes in order to find optimum routes with non-heavy traffic load for network calls and
achieve load balancing. Simulation results applied to a synchronous digital hierarchy network
used by the British Telecom, showed that the ant based routing is a quite effective
decentralized scheme compared to the software agents' approach proposed in [11].
In the Schoonderwoerd et al. model, calls are routed according to the largest probabilities in
the pheromone tables that exist in every node of the network. That is they only follow a
particular route which in some cases may be not the optimal one due to possible congestion
problems which often appear in overloaded networks. An additional extension to the above-
mentioned routing scheme is the consideration of probabilistic routing where calls are not
routed according to the largest probabilities in the pheromone tables but randomly according to
these probabilities. Therefore calls can now select more than one path to reach their final
destination. In order to overcome possible drawbacks of this routing scheme and make it
competitive against the previous ant based approaches, an additional mechanism in the way
artificial ants operate, called antipheromone, is used.
In this paper we examine the application of the ant based probabilistic routing to a proper
network topology. Section II gives a brief outline of the Schoonderwoerd et al. model. Section
III describes the basic principles of the proposed routing scheme. The performance of this
Page 5
5
approach is examined through a number of various metrics which are good indicators of the
network performance on the one hand (throughput, call failures) and the ants' behavior
characteristics on the other (mean pheromone quantity etc.). Detailed simulation results are
presented in section III.
II. THE SCHOONDERWOERD ET AL. MODEL
Ant colony optimization algorithms have been inspired by the foraging behavior of real ant
colonies and specialized for the solution of minimum cost or shortest path problems on graphs.
The way they are operate has been extensively studied in [6] and [12]. Real ants deposit
pheromone on the ground while walking which influence the rest of them to follow, in
probability, the same route. Using this mechanism they can find the shortest path between two
nodes (nest, food). Dorigo et al, were first to apply the above simple principle to solve the
famous NP-complete problem of travelling salesman (TSP) using artificial pheromone trails.
These correspond to specific distributed numeric information which are affected by the ants to
reflect their experience gained while solving a particular problem. [4], [6].
Schoonderwoerd et al. [7] and [8] have developed a quite similar model to cope with the
dynamic aspects of routing problems in communication networks. Their main intention was to
propose an efficient decentralized routing scheme in order to route calls via parts of the
network that have spare capacity. It has to be noted that in real communication networks
which are characterized by nonuniform traffic distribution between their nodes, the shortest
routes between two nodes is not necessarily the ideal solution.
In the Schoonderwoerd et al. model every node has a routing table which indicates the
direction of a call in order to reach the final destination. Artificial ants modify the table entries
and thus continually affect the current network state. For this reason, routing tables are called
Page 6
6
pheromone tables. As an example, figure 1 shows a typical pheromone table for node 1 for a
simple network configuration with 5 nodes. Ants behave as particular calls that properly
polarize the network and achieve load balancing. By this way the changing load over the whole
network is efficiently distributed and hence the number of lost calls is minimized. This
decentralized routing scheme appears to be quite attractive since it allows the parallel
operation of ants and calls that are activated in the network.
Roughly speaking this decentralized scheme works as follows:
At every simulation time step every node generates an ant with a random destination. Ants
are walking according to probabilities given in pheromone tables. Their speed, unless they are
delayed on a particular node, is one node per simulation time steps. When an ant arrives to a
node, the entry corresponding to the node from which the ant came is increased by the
quantity:
P
PPold
∆+∆+=Ρ
1 (1)
where ûp is the pheromone increase. The other entries in the table of this node are decreased
accordingly following the formula:
P
PP old
∆+=
1 (2)
Ants are dying when they reach their destination. In order to force ants to choose short
lengths, pheromone increase reduces progressively with the age of the ant using the equation:
bt
a
age
+=∆Ρ (3)
Page 7
7
where tage is the number of time steps passed since the birth of the ant. a and b are proper
constants chosen by the designer. Possible congestion problems are faced by delaying ants at
congested nodes according to the formula:
[ ]sdeCdelay ⋅−⋅= (4)
where C and d are constants and s is the spare capacity of the node. The following constant
values were found to consist a suitable set: a= 0.08, b=0.005, C=80, d=0.0755 (See [7]). The
incorporation of delay reduces the ant flow rate from the congested node to its neighboring
ones permitting pheromones for other alternative choices to increase rapidly and makes
delayed ants older. Therefore they have less influence on pheromone tables.
Calls are operating independently and in parallel with ants. When a call is set up every node
has a certain probability to be the destination node. These probabilities are generated at the
start of every run and lie between 0.01 and 0.07. At every time step a call is generated by
following Poisson distribution and the duration is exponentially distributed with an average of
170 steps. In order to determine the path for a call from a specific node to a destination, the
route, where intermediate nodes have large pheromone, is selected. This route is valid if the
destination can be reached unless one of the nodes on the route is congested. If that happens,
the call is blocked.
Hence there is an indirect interaction between calls and ants in the means that ants affect the
routing tables and calls influence the load on nodes which effect the ants with the delay
mechanism. A more detailed description of the method is given in [7].
III. ANT BASED PROBABILISTIC ROUTING
An additional extension to the above-mentioned system is the consideration of probabilistic
routing where calls are not routed according to the largest probabilities in the pheromone
Page 8
8
tables but randomly according to these probabilities. Therefore calls can now select more than
one path to reach their final destination. This principle can be particularly helpful in avoiding
possible congestion problems in overloaded networks. In order to overcome possible
drawbacks of this routing scheme and make it competitive against the previous ant based
approaches, some additional mechanisms have to be taken into account. Such mechanisms can
be incorporated into the characteristics of artificial ants so that the enhanced ant entities are
able to properly polarize the network in order the probabilistic routing have at least
comparative performance with the routing scheme proposed by Schoonderwoerd et al.
A way of achieving this is by using antipheromone mechanisms which, differently from
pheromones, are used for reducing the values of probabilities maintained in the pheromone
tables. In this approach every ant has a random destination and moves from one node to
another on the basis of the probabilities included in the pheromone table. When an ant reaches
its destination node, the probabilities included in the pheromone table, which correspond to the
node from which they come from, are renewed. With this renewal the pheromone either
increases according to how short is the distance or decreases when the distance that the ant has
covered is longer than the one that has already been recorded (at the pheromone table of the
starting point for the particular destination). The increase and decrease of pheromone can be
illustrated in figure 2.
Noise can be considered in order to handle the so called shortcut and blocking problems.
The shortcut problem occurs when a new shorter route becomes suddenly available and the
blocking one when an older route is no longer valid. In both cases, it is difficult for the artificial
ants that realize only the quantity of laid pheromone, to have an adequate dynamic reaction and
find the new proper routes.
Page 9
9
Noise avoids freezing the routes that remain static for a long time and suddenly change and
also encourages the easier discovery of a proper route that appears suddenly due to the release
from congestion of a node. Using that consideration we can handle more easily shortcut and
blocking problems. For this reason a noise factor f is considered. An ant has a probability f to
select a pure random path and (1-f) to choose a path according to the pheromone tables on the
nodes. This property can be clearly shown in figure 3. Let's suppose that an ant is launched on
node 6. Then it has f probability to select a random node between 3, 7, 11, 12 and 13 and (1-f)
probability to select an intermediate node among the previous ones according to the
pheromone tables following the principles of the probabilistic ant routing explained before. An
indicative value 5% noise factor was used by Schoonderwoerd et al. in their model and was
also chosen in our simulation. It must be noted that values of noise factor greater than 6% are
shown to have a dramatic increase of call failures and hence decrease the performance of the
routing scheme.
The ant based probabilistic routing method has been applied to a network topology of 25
nodes shown in figure 4. Each node is represented as a class structure containing various
parameters (identification of the node, adjacent nodes, capacity etc). Nodes are interconnected
to each other following the structure of figure 4 and a possible such interconnection is shown
in figure 5. Every node has a maximum number of capacities, which in our case is 40 storage
units. (See figure 6).
The initialization guidelines follow the ones proposed by Schoonderwoerd et al. in their
model [7]. The network is initialized with equal probabilities for neighbor nodes in every
pheromone table and this takes place for a fixed period of time steps before the calls are
applied. During initialization the artificial ants bias the pheromone tables so that proper short
Page 10
10
routes can be easily found. With this manner, the behavior of generated calls is examined with
a greater amount of fidelity.
IV. SIMULATION RESULTS
A simulation program has been implemented using the programming languages of Java and
C. We have chosen these specific languages in order to combine the benefits of Java as a visual
programming language permitting inheritance with the fast execution of C based algorithms.
The program follows the general guidelines proposed in [7] and [8] briefly discussed in
sections II and III.
The results regarding the various metrics in both the absence and presence of noise
presented in the next section, are the mean value of 100 repeated trials between 0 and 15.000
approximately simulation time steps. This procedure was followed in order to avoid possible
variability problems between runs due to the initial random selection of call probabilities,
random generations and lengths of calls.
In order to examine the performance of the ant-based probabilistic routing various metrics
have been taken into account. In general the metrics used can be classified into two categories:
The first one has to do with the network performance whereas the second contains parameters
about the ant heuristic algorithm.
At first we examine the case where no noise is included. Figure 7 shows the efficiency of
this routing scheme for various simulation time steps regarding the number of call failures. The
number of call failures is a basic metric used also in Schoonderwoerd et al. formulation [7]. As
it can be shown, the maximum number of call failures occurs at about 1000 time steps. After
2500 steps failures are getting steady (smaller variations of failures). The rate in which calls
Page 11
11
arrive in the system and the variations in throughput are shown in figures 8 and 9 respectively.
The mean delay per node appears in figure 10.
For comparison reasons we have also simulated and examined the performance of the
Schoonderwoerd et al. routing scheme using the same model assumptions. Figures 11, 12 and
13 show the number of call failures, the arrival rate and the throughput. More analytical results
are presented in [13]. The two routing schemes show to have comparable performance (The
performance of the ant based probabilistic routing enhanced with the antipheromone
mechanisms is slightly better). However the fact that the new scheme allows the possibility of
calls to follow different paths leads to a better link utilization which may be advantageous in
the case of overloaded networks.
Figures 14, 15 and 16 show the performance of the new scheme under the effect of noise. A
noise factor 5% was used. The incorporation of noise decreases slightly the performance of the
routing scheme asthis can be easily noticed.
As regards to the second category various interesting metrics have been taken into account.
Figure 17 shows the mean number of ants per each node of the network. The mean number of
ants used for the whole network is shown in figure 18. It can be easily seen that after 8000
time steps this number is about 500. The mean value of pheromone quantity for all routing
tables in every node is shown in figure 19.
V. CONCLUSIONS
This paper examined the behavior of an ant based decentralized router using an adequate set
of commonly acceptable and some newly introduced metrics. The routing scheme presented is
an extension of the one presented in [7]. A further mechanism of antipheromone has been
considered in the operation of the artificial ant and calls are now following more than one path
Page 12
12
according to the values of pheromone tables. In the case of wireline communication networks
this adaptation helps in congestion avoidance and achieves an adequate grade of resources
utilization.
The heuristic algorithm presented can be easily compared with other routing methods used
in networks operating with some commonly used protocols such as ALOHA, Ethernet, FDDI
and DQDB. A possible application of the scheme to the areas shown in figure 20 is also an
interesting topic of further research.
REFERENCES
[1] A. S. Tanenbaum, Computer Networks, Prentice-Hall 3nd edition, 1996.
[2] J. Walrand, Communication Networks: A First Course, McGraw Hill 2nd edition,
1998.
[3] A. Vasilakos A and W. Pedrycz, Computational Intelligence in Telecommunications
Networks, CRC Press, 2000.
[4] M. Dorigo, V. Maniezzo and A. Colorni, "Positive feedback as a search strategy",
Tech. Rep. No.91-016, Dip Elettronica, Politecnico di Milano, Italy 1991.
[5] M. Dorigo, Ottimizzazione, Apprendimento Automatico, ed Algoritmi Basati su
Metafora Naturale (Optimization, Learning and Natural Algorithms Ph.D.Thesis,
Politecnico di Milano, Italy (in Italian), 1992.
[6] Dorigo M., V. Maniezzo and A. Colorni, "The Ant System: Optimization by a Colony
of Cooperating Agents". IEEE Transactions on Systems, Man, and Cybernetics–Part
B, vol. 26, No. 1, 1996, pp. 29–41.
[7] R. Schoonderwoerd, O. Holland, J. Bruten and L. Rothkrantz, "Ant-based Load
Balancing in Telecommunications Networks", Adaptive Behavior, vol. 5, 1997,
pp. 169-207.
Page 13
13
[8] R. Schoonderwoerd, O. Holland and J. Bruten, "Ant-like Agents for Load Balancing in
Telecommunications Networks", Proceedings of Agents'97, Marina del Rey, CA,
1997, pp. 209-216.
[9] Di Caro G. and M. Dorigo, "AntNet: A Mobile Agents Approach to Adaptive
Routing", Tech. Rep. IRIDIA/97-12, Uni-ver-sité Libre de Bruxelles, Belgium, 1997.
[10] Di Caro G. and M. Dorigo, "AntNet: a mobile agents approach to adaptive routing",
Ninth Dutch/Belgian Artificial Intelligence Conference, Antwerp, Belgium, 1997,
pp. 12–13.
[11] S. Appleby and S. Stewart, Mobile software agents for control in telecommunications
networks, BT Technology Journal, vol. 12, 1994, pp. 104-113.
[12] E. Bonabeau and G. Théraulaz, "Swarm smarts", Scientific American March 2000.
[13] H.G. Sandalidis, K. Mavromoustakis and P. Stavroulakis, "Performance Measures of
an Ant based Decentralised Routing Scheme for Circuit Switching Communication
Networks", Special Issue On Computational Intelligence In Telecommunications
Networks, Softcomputing Journal, Springer Verlag (To be published)
Page 14
14
Fig. 1: Pheromone (routing) table for node 1
Fig. 2: Consideration of antipheromone mechanism
Page 15
15
2
1
5
4 2 3
61 2
1 11 6
1 3
1 4
1 52 4
1 7
1 9
1 8
2 0
2 1
2 2
2 5
7
89
3
1 0
an t launc h
Fig. 3: An example of noise addition
2
1
5
4 2 3
61 2
1 11 6
1 3
1 4
1 52 4
1 7
1 9
1 8
2 0
2 1
2 2
2 5
7
89
3
1 0
25 node S DH N etw ork
Fig. 4: Topology of the network used in simulation
Page 16
16
1RGH L
LG
QHLJKERXUKV
FDSDFLW\
VSDUH FDSDFLW\
SKHURPRQH WDEOH
DQWV
1RGH L��
LG
QHLJKERXUKV
FDSDFLW\
VSDUH FDSDFLW\
SKHURPRQH WDEOH
DQWV
1RGH L��
LG
QHLJKERXUKV
FDSDFLW\
VSDUH FDSDFLW\
SKHURPRQH WDEOH
DQWV
1RGH L�N
LG
QHLJKERXUKV
FDSDFLW\
VSDUH FDSDFLW\
SKHURPRQH WDEOH
DQWV
Fig. 5: Parameters used in every node
Node
node capacity of 40 calls
node capacity 1node capacity 2
node capacity 40
Fig. 6: Capacity of a node
Page 17
17
0
3
6
9
12
15
18
0 500 820 3000 7000 11000 12900 15000
6LPXODWLRQ VWHSV
&DOOIDLOXUHV
Fig.7. Call failures vs time steps
0
0,5
1
1,5
2
2,5
3
3,5
4
4,5
0 45 150
500
790
820
1000
3000
5000
7000
9000
1100
0
1230
0
1290
0
1300
0
1500
0
6LPXODWLRQ VWHSV
&DOOV
Linear
Fig. 8. Rate of call arrival
Page 18
18
0
0,5
1
1,5
2
2,5
0 25 45 50 150
250
500
750
790
810
820
950
1000
2000
3000
4000
5000
6000
7000
8000
9000
1000
0
1100
0
1200
0
1230
0
1260
0
1290
0
1295
0
1300
0
1400
0
1500
0
6LPXODWLRQ WLPH VWHSV
6HUYHGFDOOV
Linear
Fig.9. Throughput
02468
101214
50 250
750
810
950
2000
4000
6000
8000
1000
0
1200
0
1260
0
1295
0
1400
0
7LPH
0HDQ'HOD\SHU
QRGH
Fig. 10. Mean delay per node
Page 19
19
0
2
4
6
8
10
12
14
16
18
50 250
750
810
950
2000
4000
6000
8000
1000
0
1200
0
1260
0
1295
0
1400
0
&DOO)DLOXUHV
Figure 11: Call Failures vs. simulation time steps (Schoonderwoerd et. al. model)
00,5
11,5
22,5
33,5
44,5
0 45 150
500
790
820
1000
3000
5000
7000
9000
1100
0
1230
0
1290
0
1300
0
1500
0
6LPXODWLRQ 6WHSV
FDOOV
Linear
Figure 12: Rate of call arrival (Schoonderwoerd et al. model)
Page 20
20
0
0,5
1
1,5
2
2,5
25 50 250
750
810
950
2000
4000
6000
8000
1000
0
1200
0
1260
0
1295
0
1400
0
6LPXODWLRQ WLPH VWHSV
6HUHGFDOOV
/LQHDU
Fig. 13. Throughput for the Schoonderwoerd et al. model
0
5
10
15
20
25
30
025
079
095
011
5012
7013
9014
8030
0040
0050
0080
00
1100
0
1260
0
1300
0
1515
0
6LPXODWLRQ WLPH VWHSV
&DOOIDLOXUHV
Fig. 14. Call failures for 5% noise
Page 21
21
0
0,5
1
1,5
2
2,5
3
3,5
4
4,5
25 50 250
750
810
950
2000
4000
6000
8000
1000
0
1200
0
1260
0
1295
0
1400
0
6LPXODWLRQ WLPH VWHSV
1XPEHURIFDOOV
Arrived callsServed calls
Fig. 15. Comparison of arrived and served calls for 5% noise
0
2
4
68
10
12
14
50 250
750
810
950
2000
4000
6000
8000
1000
0
1200
0
1260
0
1295
0
1400
0
6LPXODWLRQ WLPH VWHSV
0HDQ'HOD\SHUQRGH
Fig. 16. Mean delay per node for 5% noise
Page 22
22
0
5
10
15
20
25
50 250
750
810
950
2000
4000
6000
8000
1000
0
1200
0
1260
0
1295
0
6LPXODWLRQ WLPH VWHSV
1XPEHURI$QWV
Fig. 17. Mean number of ants per node
0100200300400500600
50 250
750
810
950
2000
4000
6000
8000
1000
0
1200
0
1260
0
1295
0
6LPXODWLRQ WLPH VWHSV
1XPEHU
RI$QWV
Fig. 18. Mean number of ants for the whole network.
Page 23
23
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
50 250
750
810
950
2000
4000
6000
8000
1000
0
1200
0
1260
0
1295
0
1400
0
6LPXODWLRQ WLPH VWHSV
3KHURPRQHTXDQWLW\
Fig. 19. Mean pheromone value used for all nodes.
Fig. 20. Possible application of the ant based system