Submitted to IEEE Transactions on Systems, Man, and Cybernetics

The Ant System:

Optimization by a colony of cooperating agents
Marco Dorigo, Vittorio Maniezzo, Alberto Colorni
Dipartimento di Elettronica e Informazione, Politecnico di Milano
Piazza Leonardo da Vinci 32, 20133 Milano, Italy
e-mail: [email protected]
[email protected]
[email protected]

Abstract

An analogy with the way ant colonies function has suggested the definition of a new computational

paradigm, which we call Ant System. We propose it as a viable new approach to stochastic combinatorial

optimization. The main characteristics of this model are positive feedback, distributed computation, and the use

of a constructive greedy heuristic. Positive feedback accounts for rapid discovery of good solutions, distributed

computation avoids premature convergence, and the greedy heuristic helps find acceptable solutions in the early

stages of the search process. We apply the proposed methodology to the classical Traveling Salesman Problem

(TSP), and report simulation results. We also discuss parameter selection and the early setups of the model, and

compare it with tabu search and simulated annealing using TSP. To demonstrate the robustness of the

approach, we show how the Ant System (AS) can be applied to other optimization problems like the

asymmetric traveling salesman, the quadratic assignment and job-shop scheduling. Finally we discuss the

salient characteristics – global data structure revision, distributed communication and probabilistic transitions of

the AS.

I. Introduction

In this paper we define a new general-purpose heuristic algorithm which can be used to solve different

combinatorial optimization problems. The new heuristic has the following desirable characteristics:

• It is versatile, in that it can be applied to similar versions of the same problem; for example, there is a

straightforward extension from the traveling salesman problem (TSP) to the asymmetric traveling salesman

problem (ATSP).

• It is robust. It can be applied with only minimal changes to other combinatorial optimization problems such

as the quadratic assignment problem (QAP) and the job-shop scheduling problem (JSP).
• It is a population based approach. This is interesting because it allows the exploitation of positive feedback

as a search mechanism, as explained later in the paper. It also makes the system amenable to parallel

implementations (though this is not considered in this paper).

These desirable properties are counterbalanced by the fact that, for some applications, the Ant System can

be outperformed by more specialized algorithms. This is a problem shared by other popular approaches like

simulated annealing (SA), and tabu search (TS), with which we compare the Ant System. Nevertheless, we

believe that, as is the case with SA and TS, our approach is meaningful in view of applications to problems

which, although very similar to well known and studied basic problems, present peculiarities which make the

application of the standard best-performing algorithm impossible. This is the case, for example, with the ATSP.

In the approach discussed in this paper we distribute the search activities over so-called "ants," that is,

agents with very simple basic capabilities which, to some extent, mimic the behavior of real ants. In fact,

research on the behavior of real ants has greatly inspired our work (see [10], [11], [21]). One of the problems

studied by ethologists was to understand how almost blind animals like ants could manage to establish shortest

route paths from their colony to feeding sources and back. It was found that the medium used to communicate

information among individuals regarding paths, and used to decide where to go, consists of pheromone trails. A

moving ant lays some pheromone (in varying quantities) on the ground, thus marking the path by a trail of this

substance. While an isolated ant moves essentially at random, an ant encountering a previously laid trail can

detect it and decide with high probability to follow it, thus reinforcing the trail with its own pheromone. The

collective behavior that emerges is a form of autocatalytic behavior1 where the more the ants following a trail,

the more attractive that trail becomes for being followed. The process is thus characterized by a positive

feedback loop, where the probability with which an ant chooses a path increases with the number of ants that

previously chose the same path.

Consider for example the experimental setting shown in Fig. 1. There is a path along which ants are walking

(for example from food source A to the nest E, and vice versa, see Fig. 1a). Suddenly an obstacle appears and

the path is cut off. So at position B the ants walking from A to E (or at position D those walking in the opposite

direction) have to decide whether to turn right or left (Fig. 1b). The choice is influenced by the intensity of the

pheromone trails left by preceding ants. A higher level of pheromone on the right path gives an ant a stronger

stimulus and thus a higher probability to turn right. The first ant reaching point B (or D) has the same probabil-

ity to turn right or left (as there was no previous pheromone on the two alternative paths). Because path BCD is

1 An autocatalytic [12], i.e. positive feedback, process is a process that reinforces itself, in a way that causes
very rapid convergence and, if no limitation mechanism exists, leads to explosion.

2
shorter than BHD, the first ant following it will reach D before the first ant following path BHD (Fig. 1c). The

result is that an ant returning from E to D will find a stronger trail on path DCB, caused by the half of all the

ants that by chance decided to approach the obstacle via DCBA and by the already arrived ones coming via

BCD: they will therefore prefer (in probability) path DCB to path DHB. As a consequence, the number of ants

following path BCD per unit of time will be higher than the number of ants following BHD. This causes the

quantity of pheromone on the shorter path to grow faster than on the longer one, and therefore the probability

with which any single ant chooses the path to follow is quickly biased towards the shorter one. The final result

is that very quickly all ants will choose the shorter path.

The algorithms that we are going to define in the next sections are models derived from the study of artificial

ant colonies. Therefore we call our system Ant System (AS) and the algorithms we introduce ant algorithms. As

we are not interested in simulation of ant colonies, but in the use of artificial ant colonies as an optimization

tool, our system will have some major differences with a real (natural) one:

• artificial ants will have some memory,

• they will not be completely blind,

• they will live in an environment where time is discrete.

E E E

D D

H Obstacle C H Obstacle C

B B

A A A
a) b) c)

Fig. 1. a) Ants follow a path between points A and E.

b) An obstacle is interposed; ants can choose to go around it following one of the two different
paths with equal probability.
c) On the shorter path more pheromone is laid down.

3
 Nevertheless, we believe that the ant colony metaphor can be useful to explain our model. Consider the

graph of Fig. 2a, which is a possible AS interpretation of the situation of Fig. 1b. To fix the ideas, suppose that

the distances between D and H, between B and H, and between B and D—via C—are equal to 1, and let C be

positioned half the way between D and B (see Fig. 2a). Now let us consider what happens at regular discretized

intervals of time: t=0, 1, 2, ... . Suppose that 30 new ants come to B from A, and 30 to D from E at each time

unit, that each ant walks at a speed of 1 per time unit, and that while walking an ant lays down at time t a

pheromone trail of intensity 1, which, to make the example simpler, evaporates completely and instantaneously

in the middle of the successive time interval (t+1, t+2).

At t=0 there is no trail yet, but 30 ants are in B and 30 in D. Their choice about which way to go is

completely random. Therefore, on the average 15 ants from each node will go toward H and 15 toward C (Fig.

2b).

At t=1 the 30 new ants that come to B from A find a trail of intensity 15 on the path that leads to H, laid by

the 15 ants that went that way from B, and a trail of intensity 30 on the path to C, obtained as the sum of the

trail laid by the 15 ants that went that way from B and by the 15 ants that reached B coming from D via C (Fig.

2c). The probability of choosing a path is therefore biased, so that the expected number of ants going toward C

will be the double of those going toward H: 20 versus 10 respectively. The same is true for the new 30 ants in D

which came from E.

This process continues until all of the ants will eventually choose the shortest path.

4
 E

D D
d=1 d=0.5

H C H C
d=1 d=0.5
B B

5
 Given a set of n towns, the TSP can be stated as the problem of finding a minimal length closed tour that

visits each town once. We call dij the length of the path between towns i and j; in the case of Euclidean TSP, dij
2 1/2
is the Euclidean distance between i and j (i.e., dij =[(xi -xj ) + (yi -yj )2 ] ). An instance of the TSP is given

by a graph (N,E), where N is the set of towns and E is the set of edges between towns (a fully connected graph

in the Euclidean TSP).

m = ∑i =1 b i (t)
n

Let bi(t) (i=1, ..., n) be the number of ants in town i at time t and let be the total number of

ants. Each ant is a simple agent with the following characteristics:

• it chooses the town to go to with a probability that is a function of the town distance “visibility” and

amount of trail present on the connecting edge;

• to force the ant to make legal tours, transitions to already visited towns are disallowed until a tour is

completed (this is controlled by a tabu list);

• when it completes a tour, it lays a substance called trail on each edge (i,j) visited.

Let τij(t) be the intensity of trail on edge (i,j) at time t. Each ant at time t chooses the next town, where it will

be at time t+1. Therefore, if we call an iteration of the AS algorithm the m moves carried out by the m ants in

the interval (t, t+1), then every n iterations of the algorithm (which we call a cycle) each ant has completed a

tour. At this point the trail intensity is updated according to the following formula

τij(t+n)=ρ.τij(t)+∆τij (1)

where

ρ is a coefficient such that (1 - ρ) represents the evaporation of trail between time t and t+n,
m
∆τ ij = ∑ ∆τ kij
k =1 (2)

∆τ kij
where is the quantity per unit of length of trail substance (pheromone in real ants) laid on edge (i,j) by the

k-th ant between time t and t+n; it is given by

 Q if k th ant uses edge (i, j) in its tour (between time t and t + n)

L -
∆τ ij = 
k k

 0 otherwise (3)

where Q is a constant and Lk is the tour length of the k-th ant.

6
 The coefficient ρ must be set to a value <1 to avoid unlimited accumulation of trail (see note 1). The

intensity of trail at time 0, τij(0), can be set to arbitrarily chosen values; in our experiments for every edge (i,j)

we set τij(0)=c, with c a small positive constant.

In order to satisfy the constraint that an ant visits all the n different towns, we associate with each ant a data

structure called the tabu list2, that saves the towns already visited up to time t and forbids the ant to visit them

again before n iterations (a tour) have been completed. When a tour is completed, the tabu list is used to

compute the ant’s current solution (i.e., the distance of the path followed by the ant). The tabu list is then

emptied and the ant is free again to choose. We define tabuk the vector containing the tabu list of the k-th ant,

tabuk the set obtained from the elements of tabuk, and tabuk(s) the s-th element of the list (i.e., the s-th town

visited by the k-th ant in the current tour).

We call visibility ηij the quantity 1/dij. This quantity is not modified during the run of the AS, as opposed to

the trail which instead changes according to the previous formula (1).

We define the transition probability from town i to town j for the k-th ant as

[ ] [ ]
 α β
τij (t) ⋅ ηij
 if j ∈ allowed k
 ∑ [τ (t )]α ⋅ [η ]β
pij (t) =  k ∈allowed
k ik ik
k

 0 otherwise (4)
where allowedk = {N - tabuk} and where α and β are parameters that control the relative importance of trail

versus visibility. Therefore the transition probability is a trade-off between visibility (which says that close

towns should be chosen with high probability, thus implementing a greedy constructive heuristic) and trail

intensity at time t (that says that if on edge (i,j) there has been a lot of traffic then it is highly desirable, thus

implementing the autocatalytic process).

2 Even though the name chosen recalls tabu search, proposed in [17,18], there are substantial differences
between our approach and tabu search algorithms. We mention here: (i) the absence in the AS of any
aspiration function, (ii) the difference of the elements recorded in the tabu list, permutations in the case of
tabu search, nodes in the AS (our algorithms are constructive heuristics, which is not the case of tabu
search).

7
III. The algorithms

Given the definitions of the preceding section, the so-called ant-cycle algorithm is simply stated as follows. At

time zero an initialization phase takes place during which ants are positioned on different towns and initial

values τij(0) for trail intensity are set on edges. The first element of each ant's tabu list is set to be equal to its

starting town. Thereafter every ant moves from town i to town j choosing the town to move to with a

probability that is a function (with parameters α and β, see formula (4)) of two desirability measures. The first,

the trail τij(t), gives information about how many ants in the past have chosen that same edge (i,j); the second,

the visibility ηij, says that the closer a town the more desirable it is. Obviously, setting α = 0, the trail level is no

longer considered, and a stochastic greedy algorithm with multiple starting points is obtained.

After n iterations all ants have completed a tour, and their tabu lists will be full; at this point for each ant k

the value of Lk is computed and the values

∆τkij are updated according to formula (3). Also, the shortest path
min
found by the ants (i.e., k Lk, k = 1, ..., m) is saved and all the tabu lists are emptied. This process is iterated

until the tour counter reaches the maximum (user-defined) number of cycles NCMAX, or all ants make the same

tour. We call this last case stagnation behavior because it denotes a situation in which the algorithm stops

searching for alternative solutions. We investigate this situation in Section IV.

Formally the ant-cycle algorithm is:

1. Initialize:
Set t:=0 {t is the time counter}
Set NC:=0 {NC is the cycles counter}
For every edge (i,j) set an initial value τij(t) for trail intensity and ∆τij= 0
Place the m ants on the n nodes

2. Set s:=1 {s is the tabu list index}

For k:=1 to m do
Place the starting town of the k-th ant in tabuk(s)

3. Repeat until tabu list is full {this step will be repeated (n-1) times}
Set s:=s+1
For k:=1 to m do k
Choose the town j to move to, with probability ij
p (t) given by equation (4)
{the k-th ant is now on town i=tabuk(s-1) at time t}
Move the k-th ant to the town j
Insert town j in tabuk(s)

4. For k:=1 to m do
Compute the length Lk of the tour described by tabuk
Update the shortest tour found

8
 For every edge (i,j)
For k:=1 to m do
 Q if (i, j) ∈ tour described by tabu
k
L k
∆τ i , j = k

 0 otherwise

∆τ ij : = ∆τ ij + ∆τijk
;

5. For every edge (i,j) compute τij(t+n) according to equation τij(t+n)=ρ.τij(t)+∆τij

Set t:=t+n
Set NC:=NC+1
For every edge (i,j) set ∆τij:=0

6. If (NC < NCMAX) and (not stagnation behavior)

then
Empty all tabu lists
Goto step 2
else
Print shortest tour
Stop

The complexity of the ant-cycle algorithm is Ο(NC.n2.m) if we stop the algorithm after NC cycles. In fact

step 1 is Ο(n2+m), step 2 is Ο(m), step 3 is Ο(n2.m), step 4 is Ο(n2.m), step 5 is Ο(n2), step 6 is Ο(n.m).

Because we have experimentally found a linear relation between the number of towns and the best number of

ants (see Section V-A), the complexity of the algorithm is Ο(NC.n3).

We also experimented with two other algorithms of the AS, which we called ant-density and ant-quantity

algorithms [6]. They differ in the way the trail is updated. In these two models each ant lays its trail at each

step, without waiting for the end of the tour. In the ant-density model a quantity Q of trail is left on edge (i,j)

every time an ant goes from i to j; in the ant-quantity model an ant going from i to j leaves a quantity Q/dij of

trail on edge (i,j) every time it goes from i to j. Therefore, in the ant-density model we have

Q if the k - th ant goes from i to j between time t and t + 1

∆τ kij =
0 otherwise (5)

and in the ant-quantity model we have

Q if the k - th ant goes from i to j between time t and t + 1

d
∆τ kij =  ij
 (6)
 0 otherwise

9
 From these definitions it is clear that the increase in trail intensity on edge (i,j) when an ant goes from i to j is

independent of dij in the ant-density model, while it is inversely proportional to dij in the ant-quantity model

(i.e., shorter edges are made more desirable by ants in the ant-quantity model).

IV. Experimental study 1: Parameter setting and basic properties

We implemented the three algorithms (ant-cycle, ant-density and ant-quantity) of the AS and investigated their

relative strengths and weaknesses by experimentation. Since we have not yet developed a mathematical analysis

of the models, which would yield the optimal parameter setting in each situation, we ran simulations to collect

statistical data for this purpose.

The parameters considered here are those that affect directly or indirectly the computation of the probability

in formula (4):

• α: the relative importance of the trail, α•0;

• β: the relative importance of the visibility, β•0;

• ρ: trail persistence, 0•ρ<1 (1-ρ can be interpreted as trail evaporation);

• Q: quantity of trail laid by ants.

The number m of ants has always been set equal to the number n of cities (see Section V-A for the

explanation). We tested several values for each parameter while all the others were held constant (over ten

simulations for each setting in order to achieve some statistical information about the average evolution). The

default value of the parameters was α=1, β=1, ρ=0.5, Q=100. In each experiment only one of the values was

changed, except for α and β, which have been tested over different sets of values, as discussed at the end of this

section. The values tested were: α∈{0, 0.5, 1, 2, 5}, β∈{0, 1, 2, 5}, ρ∈{0.3, 0.5, 0.7, 0.9, 0.999} and Q∈{1,

100, 10000}. Preliminary results, obtained on small-scale problems, have been presented in [6], [7], and [12],

[13]; all the tests reported in this section are based, where not otherwise stated, on the Oliver30 problem, a 30-

cities problem described in [34]3. All the tests have been carried out for NCMAX = 5000 cycles and were

averaged over ten trials.

To compare the three models we first experimentally determined the parameters best values for each

algorithm, and then we ran each algorithm ten times using the best parameters set. Results are shown in Table

I. Parameter Q is not shown becasue its influence was found to be negligible.

3 In [34] genetic algorithms were applied to solve the Oliver30 problem; they could find a tour of length
424.635. The same result was often obtained by ant-cycle, which also found a tour of length 423.741.

10
 Table I. Comparison among ant-quantity, ant-density, and ant-cycle. Averages over 10 trials.

Best parameter set Average result Best result

ant-density α=1, β=5, ρ=0.99 426.740 424.635

ant-quantity α=1, β=5, ρ=0.99 427.315 426.255

ant-cycle α=1, β=5, ρ=0.5 424.250 423.741

Both the ant-density and the ant-quantity models have given worse results than those obtained with ant-

cycle. The reason is to be found in the kind of feedback information which is used to direct the search process.

Ant-cycle uses global information, that is, its ants lay an amount of trail which is proportional to how good the

solution produced was. In fact, ants producing shorter paths contribute a higher amount of trail than ants whose

tour was poor. On the other side, both ant-quantity and ant-density use local information. Their search is not

directed by any measure of the final result achieved. Therefore, it is not surprising that they gave worse

performance results (details can be found in [6]).

The optimal value ρ = 0.5 in ant-cycle can be explained by the fact that the algorithm, after using the greedy

heuristic to guide search in the early stages of computation, starts exploiting the global information contained in

the values τij of trail. Ant-cycle needs therefore to have the possibility to forget part of the experience gained in

the past in order to better exploit new incoming global information.

Given that we found ant-cycle to be superior to the other two algorithms, we decided to deepen our

understanding of the ant-cycle alone. Figures 3, 4, and 5 present traces of a typical run of the ant-cycle

algorithm applied to the Oliver30 problem. In particular, Fig.3 shows the length of the best found tour at each

cycle, and Fig.4 the standard deviation of the tour lengths of the population at each cycle of the same run. Note

how in the early cycles the AS identifies good tours which are subsequently refined in the rest of the run. Since

the standard deviation of the population’s tour lengths never drops to zero, we are assured that the algorithm

actively searches solutions which differ from the best-so-far found, which gives it the possibility of finding

better ones. The search for better solutions is carried on in selected regions of the search space determined by

the trail resulting from preceding cycles. This can be observed in Fig.5, in which the vertical axis shows the

average node branching of the problem’s graph. Although the graph is initially fully connected, those arcs

whose trail level falls below a (very small) value ε, which makes their probability of being chosen by ants

negligible, are removed. The node branching of node i is therefore given by the number of edges which exit

from node i and which have a trail level higher than ε. Note how at the beginning of the run an ant could go

11
from any node to any other (except for tabu list constraints), while at the end the possible choices are

significantly reduced.

Best tour length

600

500

400

Cycles
300
0 500 1000 1500 2000 2500 3000

Fig. 3. Evolution of best tour length (Oliver30). Typical run.

Tour length
standard deviation
80

70

60

50

40

30

20

10 Cycles
0
0 500 1000 1500 2000 2500 3000

Fig. 4. Evolution of the standard deviation of the population’s tour lengths (Oliver30). Typical run.

Average node branching

30

25

20

15

10

5
Cycles
0
0 500 1000 1500 2000 2500 3000

12
 Fig. 5. Evolution of the average node branching of the problem’s graph (Oliver30). Typical run.

The same process can be observed in the graphs of Fig. 6, where the AS was applied to a very simple 10-

cities problem (CCA0, from [20]), and which depict the effect of ants search on the trail distribution. In the

figure the length of the edges is proportional to the distances between the towns; the thickness of the edges is

proportional to their trail level. Initially (Fig. 6a) trail is uniformly distributed on every edge, and search is only

directed by visibilities. Later on in the search process (Fig. 6b) trail has been deposited on the edges composing

good tours, and is evaporated completely from edges which belonged to bad tours. The edges of the worst tours

actually resulted to be deleted from the problem graph, thus causing a reduction of the search space.

10 7 10 7
9 9

8 8

3 3

2 6 2 4 6
4

1 1
5 5

a) b)

Fig. 6. Evolution of trail distribution for the CCA0 problem.

a) Trail distribution at the beginning of search.
b) Trail distribution after 100 cycles.

Besides the tour length, we also investigated the stagnation behavior, i.e. the situation in which all the ants

make the same tour. This indicates that the system has ceased to explore new possibilities and no better tour

will arise. With some parameter settings we observed that, after several cycles, all the ants followed the same

tour despite the stochastic nature of the algorithms because of a much higher trail level on the edges comprising

that tour than on all the others. This high trail level made the probability that an ant chooses an edge not

belonging to the tour very low. For an example, see the Oliver30 problem, whose evolution of average

branching is presented in Fig. 7. In fact, after 2500 cycles circa, the number of arcs exiting from each node

sticks to the value of 2, which – given the symmetry of the problem – means that ants are always following the

same cycle.

13
Average node branching
30

25

20

15

10

5 2

0 Cycles
0 500 1000 1500 2000 2500 3000

Fig. 7. Average node branching of a run going to stagnation behavior (Oliver30). Typical run obtained setting
α=5 and β=2.

This led us to also investigate the behavior of the ant-cycle algorithm for different combination of parameters

α and β (in this experiment we set NCMAX=2500). The results are summarized in Fig. 8, which was obtained

running the algorithm ten times for each set of parameters, averaging the results and ascribing each averaged

result to one of the three following different classes.

• Bad solutions and stagnation. For high values of α the algorithm enters the stagnation behavior very quickly

without finding very good solutions. This situation is represented by the symbol ∅ in Fig. 8;

• Bad solutions and no stagnation. If enough importance was not given to the trail (i.e., α was set to a low

value) then the algorithm did not find very good solutions. This situation is represented by the symbol ∞.
• Good solutions. Very good solutions are found for α and β values in the central area (where the symbol used

is ●). In this case we found that different parameter combinations (i.e., (α=1, β=1),(α=1, β=2),(α=1,

β=5),(α=0.5, β=5)) resulted in the same performance level: the same result (the shortest tour known on the

Oliver30 problem) was obtained in approximately the same number of cycles.

14
 ● - The algorithm finds the best known solution without entering the stagnation behavior.
∞ - The algorithm doesn't find good solutions without entering the stagnation behavior.
∅ - The algorithm doesn't find good solutions and enters the stagnation behavior.

The results obtained in this experiment are consistent with our understanding of the algorithm: a high value

for α means that trail is very important and therefore ants tend to choose edges chosen by other ants in the past.

On the other hand, low values of α make the algorithm very similar to a stochastic multigreedy algorithm.

In Fig. 9 we present the new optimal tour4 found using the experimentally determined optimal set of

parameters values for the ant-cycle algorithm, α=1, β=5, ρ=0.5, Q=100. This tour is of length 423.741 and

presents two inversions, 2–1 and 25–24, with respect to the best tour published in [34].

Fig. 9. The new best found tour obtained with 342 cycles of the ant-cycle algorithm for the Oliver30 problem
(α=1, β=5, ρ=0.5, Q=100), real length = 423.741, integer length = 420.

The major strengths of the ant-cycle algorithm can be summarized as:

• Within the range of parameter optimality the algorithm always finds very good solutions for all the tested

problems (Oliver30 and other problems which will be presented later).

• The algorithm quickly finds good solutions (see Fig. 10; for a comparison with other heuristics, see Section

VI); nevertheless it doesn't exhibit stagnation behavior, i.e. the ants continue to search for new possibly

better tours.

4 This result is not competitive with results obtained by special-purpose algorithms [2].

15
• With increasing dimensions the sensitivity of the parameter values to the problem dimension has been found

to be very low.

Best tour length

600

500

400

Cycles
300
0 500 1000 1500

Fig. 10. The algorithm finds good values for Oliver30 very quickly and the new optimal value (423.741) after
NC=342 cycles.

We partially tested the ant-cycle algorithm on the Eilon50 and Eilon75 problems [14] with a limited number

of runs and with a number of cycles bounded by NCMAX=3000. Under these restrictions we never got the best-

known result, but a quick convergence to satisfactory solutions was maintained for both the problems.

V. Experimental study 2: Extensions and advanced properties

In this section we discuss experiments which have deepened our understanding of the ant-cycle algorithm. We

study how synergy affects the algorithm performance (Section V-A). We compare the performance of ant-cycle

when all the ants are initially positioned on a unique starting point with the performance obtained when each ant

starts from a different town (Section V-B). Finally, we study the effects of an elitist strategy which increases the

importance of the ant that found the best tour (Section V-C), and the change in performance of the AS when the

problem dimension increases (Section V-D).

A. Synergistic effects

We ran a set of experiments to assess both the impact of the number m of ants, and the importance of

communication through trail, on the efficiency of the solving process. In this case, the test problem involved

finding a tour in a 4x4 grid of evenly spaced points: this is a problem with a priori known optimal solution (160

if each edge has length 10, see Fig. 11).

16
 —
10
—

Fig. 11. An optimal solution for the 4x4 grid problem.

The result was that there is a synergistic effect in using many ants and using the trail communication system;

that is, a run with n ants is more search-effective with communication among ants than with no communication.

In case of communicating ants, there is an "optimality point" given by n•m in which the synergistic effects

reach a maximum. The results are shown in Figs. 12 and 13.

In Fig. 12 we compare a situation in which ants do not communicate (α=0), with a situation in which they

communicate (α=1). Results show that communication is indeed exploited by the algorithm. In Fig. 13 we

report on an experiment in which the 4x4 grid problem was solved with m ∈ {4, 8, 16, 32, 64}. The abscissa

shows the total number of ants used in each set of runs, the ordinate shows the so-called one-ant cycles, that is,

the number of cycles required to reach the optimum, multiplied by the number of ants used (in order to evaluate

the efficiency per ant, and have comparable data). The algorithm has always been able to identify the optimum

with any number m • 4 of ants. Tests run on a set of r x r grid problems (r = 4, 5, 6, 7, 8) have substantiated our

hypothesis that the optimal number of ants is close to the number of cities (n•m); this property was used in the

assessment of the computational complexity (Section III).

Best tour length

240

220 Best tour length

200 240

220
180 168
200
160 180
160
140 160

140
120
Cycles 120
Cycles
100 100
0 400 800 1200 1600 2000 0 400 800 1200 1600 2000

a) α=0 b) α=1

Fig. 12. Synergy: Communication among ants (α>0) improves performance.

17
 1200

Number of one-ant cycles

1000

800

600

400

200

0
4 8 16 32 64
Number m of ants

Fig. 13. Number of one-ant cycles required to reach optimum as a function of the total number of ants for the
4x4 grid problem. Results are averaged over five runs.

A second set of tests has been carried out with 16 cities randomly distributed (16 cities random graph). Again

we found that the optimal performance was reached with 8-16 ants, a number comparable with the dimension of

the problem to be solved.

B. Initialization

We tested whether there is any difference between the case in which all ants at time t=0 are in the same city and

the case in which they are uniformly distributed5. This experiment was run in order to know whether particular,

possibly problem-specific, ant distributions are more effective than the uniform distribution, which in the

previous experiments was used as the default starting configuration of ants. We used ant-cycle applied to the 16

cities random graph, to the 4x4 grid, and to the Oliver30 problem. In all three cases, uniformly distributing ants

resulted in better performance.

We also tested whether an initial random distribution of the ants over the cities performed better than a

uniform one; results show that there is no significant difference between the two choices, even though the

random distribution obtained slightly better results.

5 We say ants are uniformly distributed if there is, at time t=0, the same integer number of ants on every town
(this forces m to be a multiple of n).

18
C. Elitist strategy

We use the term "elitist strategy" (because in some way it resembles the elitist strategy used in genetic

algorithms [19]) for the modified algorithm in which at every cycle the trail laid on the edges belonging to the

best-so-far tour is reinforced more than in the standard version. We added to the trail of each arc of the best tour

a quantity e.Q/L*, where e is the number of elitist ants6 and L* is the length of the best found tour. The idea is

that the trail of the best tour, so reinforced, will direct the search of all the other ants in probability toward a

solution composed by some edges of the best tour itself.

The test were carried out again on the Oliver30 problem (the run was stopped after NCMAX = 2500 cycles)

and results indicated that there is an optimal range for the number of elitist ants: below it, increasing their

number results in better tours discovered and/or in the best tour being discovered earlier; above it, the elitist ants

force the exploration around suboptimal tours in the early phases of the search, so that a decrease in

performance results. Fig. 14 shows the outcome of a test on the Oliver30 problem where this behavior is

evident.

Local optima:
2500
425.82
2250
Number of cycles to reach

423.91
2000
the local optimum

423.74
1750

1500

1250
1000

750
500

250
0
0 1 2 4 8 12 16 20 30
Number e of elitist ants
Fig. 14. Number of cycles required to reach a local optimum related to the number of elitist ants used
(Oliver30). Results are averaged over five runs.

6 In our case the effect of an ant is to increment the value of the trail on edges belonging to its tour; therefore
in our case the equivalent of “saving” an individual is to reinforce its contribution.

19
D. Increasing the problem dimensions

The algorithm complexity presented in Section III, O(NC⋅n3), says nothing about the actual time required to

reach the optimum. The experiment presented in this section is devoted to investigating the efficiency of the

algorithm for increasing problem dimensions. Results are reported in Table II for the case of similar problems

with increasing dimensions (r x r grids with the edge length set to 10, as in Fig. 11). It is interesting to note that,

up to problems with 64 cities, the algorithm always found the optimal solution.

Table II. Time required to find optimum as a function of problem dimension. Results are averaged over five
runs.

n Best Average number of Time required to

Problem (dimension) solution cycles to find the find the optimum7
optimum (seconds)

4x4 16 160 5.6 8

5x5 25 254.1 13.6 75
6x6 36 360 60 1020
7x7 49 494.1 320 13440
8x8 64 640 970 97000

VI. Comparison with other heuristics

In this section we compare the efficacy of our algorithm to that of other heuristics, both tailored and general-

purpose.

A. Comparison with TSP-tailored heuristics

In this section we compare ant-cycle with the heuristics contained in the package "Travel" [4]. This package

represents the distances between the cities as an integer matrix and so we implemented an analogous

representation in our system8. The results of the comparisons on Oliver30 are shown in Table III, where the first

column is the length of the best tour identified by each heuristic, and the second column is the improvement on

the solution as obtained by the 2-opt heuristic (the 2-opt heuristic is an exhaustive exploration of all the

permutations obtainable by exchanging 2 cities). Comparisons have been carried out also with the Lin-

7 Tests were run on a IBM-compatible PC with 80286 Intel processor.

8 In this case distances between towns are integer numbers and are computed according to the standard code
proposed in [31].

20
Kernighan [27] improvement of the first-column solutions, which has been able to reduce the length of any tour

to 420 (or 421, depending on the starting solution provided by the basic algorithms).

Note how ant-cycle consistently outperformed 2-opt, while its efficacy – i.e., the effectiveness it has in

finding very good solutions – can be compared with that of Lin-Kernighan. On the other hand, our algorithm

requires a much longer computational time than any other tested special-purpose heuristic.

Table III. Performance of the ant-cycle algorithm compared with other approaches. Results are averaged over
ten runs, and rounded to the nearest integer.

basic9 2-opt L-K

Ant-cycle 420 - -
Near Neighbor 587 437421
Far Insert 428 421420
Near Insert 510 492420
Space Filling Curve 464 431421
Sweep 486 426420
Random 1212 663421

As a general comment of all the tests, we would like to point out that, given a good parameter setting (for

instance α=1, β=5, ρ=0.5, Q=100, e=8), our algorithm consistently found the best known solution for the

Oliver30 problem, and converged quickly towards satisfactory solutions. It always identified for Oliver30 the

best-known solution of length 423.741 in less than 400 cycles, and it took only •100 cycles to reach values

under 430. The algorithm never fell into the stagnation behavior. In fact, the average branching was always

greater than 2, and the average length of tours was never equal to the best tour found but remained somewhat

above it. This indicates that the ants followed different tours.

B. Comparison with general-purpose heuristics

We also compare ant-cycle with other general-purpose heuristics. This comparison is more fair to the AS, which

in fact is a general-purpose heuristic, and not a specialized algorithm for the TSP. To run the comparisons, we

implemented a Simulated Annealing (SA) [1], and a Tabu Search (TS) [16], [18]; we let each of them run 10

times on the Oliver30 data. SA used the annealing function T(t+1)=αT(t), with α=0.99; TS was implemented

with tabu list length varying in [20, 50]. TS and SA, and the AS as well, were allowed to run for 1 hour on a

IBM-compatible PC with 80386 Intel processor. The results are presented in Table IV.

9 The name "basic" means the basic heuristic, with no improvement.

21
 Table IV. Performance of AS compared to TS and SA on the Oliver30 problem.
Results are averaged over ten runs using integer distances.

Best Average Std.dev.

AS 420 420.4 1.3

TS 420 420.6 1.5

SA 422 459.8 25.1

Results show that the AS for this problem was as effective as TS and better than SA, when running under the

same hardware and time constraints.

VII. Generality of the approach

As we said in Section I, the AS is both versatile and robust. Versatility is exemplified by the ease with which

AS can be applied to the asymmetric TSP (ATSP), a particular kind of TSP (Section VII-A). Robustness is

exemplified by the possibility of using the same algorithm, although appropriately adapted, to solve other

combinatorial optimization problems like the quadratic assignment problem (QAP), and the job-shop

scheduling problem (JSP) (Section VII-B).

A. Versatility: The ATSP

The asymmetric traveling salesman problem is a TSP in which the distance between two nodes is not

symmetric (i.e., in general dij•dji). The ATSP is more difficult than the TSP; in fact, while symmetric TSP can

be solved optimally even on graphs with several thousand nodes, ATSP instances, and particularly ATSP

instances where the distance matrix is almost symmetric, can be solved to the optimum only on graphs with a

few dozen nodes [26], [16].

The application of the AS to the ATSP is straightforward, as no modifications of the basic algorithm are

necessary. The computational complexity of a cycle of the algorithm remains the same as in the TSP

application, as the only differences are in the distance and trail matrices which are no longer symmetric. We

chose as test problem the RY48P problem [16], a very difficult problem instance with a distance distribution

that is hard to solve even with tailored heuristics and branch and bound procedures. We ran AS 5 times on it,

each time for 4000 cycles. The average length of the best found tour was 14899, that is 3.3% longer than the

optimal one. The average number of cycles to find this result was 1517.

22
B. Robustness: QAP and JSP

Let's now consider the robustness of the AS approach. Many combinatorial problems can be solved by the AS.

To apply the autocatalytic algorithm to a combinatorial problem requires defining:

1) an appropriate graph representation with search by many simple agents for the problem;

2) the autocatalytic (i.e. positive) feedback process;

3) the heuristic that allows a constructive definition of the solutions (which we also call "greedy force");

4) the constraint satisfaction method (that is, the tabu list).

This has been done for two well-known combinatorial optimization problems – Quadratic Assignment (QAP)

and Job-Shop Scheduling (JSP) – each time obtaining an adapted version of the AS that could effectively handle

the relative problem. The most difficult (and ad hoc) tasks to face when applying the AS are to find an

appropriate graph representation for the problem to be solved and a greedy force as heuristic.

1) Quadratic Assignment Problem:

A QAP of order n is the formalization of the problem that arises when trying to assign n facilities to n locations.

Formally the problem is usually defined by two nxn (symmetric) matrices:

D = {dij} is the distance between location i and location j;

F= {fhk} is the flow (of information, products or some other quantity) between facility h and facility k;

A permutation π can be interpreted as an assignment of facility h=π(i) to location i, for each i=1,..,n. The

problem is then to identify a permutation π of both row and column indexes of the matrix F that minimizes the

total cost:
n
min z = ∑ d ijf π( i)π( j)
i, j=1

To apply AS to QAP we used the same algorithm as in the case of the TSP, after having studied an

approximation of the QAP objective function that allows a problem representation on the basis of a single

matrix. The QAP objective function was expressed by a combination of the "potential vectors" of distance and

flow matrices. The potential vectors, D and F, are the row sums of each of the two matrices. Consider the

following example:

23
 0 1 2 3 6 0 60 50 10 120
D= 1 0 4 5 → D= 10 F = 60 0 30 20 → F=
110
2 4 0 6 12 50 30 0 50 130
3 5 6 0 14 10 20 50 0 80

From the two potential vectors, a third matrix S is obtained, where each element is computed as sih=di.fh, di

and fh being elements of the potential vectors.

720 660 780 480

S = 1200 1100 1300 800
1440 1320 1560 960
1680 1540 1720 1120

The ants choose the node to move to using the inverse of the values of S as visibility data, ηih=1/sih, thus

interpreting each element sih as the choice of assigning to location i the facility h (so that, for example, 480 on

the first row is linked to the choice of assigning to location 1 facility 4). The algorithm then goes on as in the

case of the TSP, with trail laid down on the arc of the digraph whose "cost" matrix is S.

We compared AS, and a version of the AS with a local optimizing routine, with many other well know

heuristics (see [28] for more details). Experiments were run on IBM-compatible PCs with a 80286 Intel

processor, and were stopped after one hour time. The test problems used are those known as Nugent problems

[29], Elshafei [15], and Krarup [25]. As can be seen in Table V, the performance of AS was always very good.

Ant system always found a result within 5% of the best known, while AS with local optimization always found,

except for the Nugent 30 problem, the best known solution.

Table V. Comparison of the AS with other heuristic approaches. Results are averaged over five runs. Best
known results are in bold.

Nugent Nugent Nugent Elshafei Krarup

(15) (20) (30) (19) (30)
Best known 1150 2570 6124 17212548 88900
Ant System (AS) 1150 2570 6232 18122850 92490
AS with local optimization 1150 2570 6128 17212548 88900
Simulated Annealing 1150 2570 6128 17937024 89800
Tabu Search 1150 2688 6124 17212548 90090
Genetic Algorithm 1160 2654 6784 17640548 108830
Evolution Strategy 1168 2570 6308 19600212 97880
Sampling & Clustering 1150 2598 6154 17212548 88900

2) Job-shop Scheduling Problem

The JSP can be described as in the following. A set of M machines and a set of J jobs are given. The j-th job

(j=1, ..., J) consists of an ordered sequence (chain) of operations from a set O={... ojm ...}. Each operation ojm

24
∈ O belongs to job j and has to be processed on machine m for djm consecutive time instants. N=|O| is the total

number of operations. The problem is to assign the operations to time intervals in such a way that no two jobs

are processed at the same time on the same machine and the maximum of the completion times of all operations

is minimized [22].

To apply the AS to JSP we chose the following representation. A JSP with M machines, J jobs and operation

set O is represented as an undirected, weighted graph Q=(O’,A), where O’=O"{o0} and A is the set of arcs that

completely connect the nodes of O and that connect o0 with the first operation of each job. Note that graph Q is

not the graph with cliques representing machines that is usually utilized to represent the JSP. Node o0 is

necessary in order to specify which job will be scheduled (N −in1)case several jobs have their first operation on
Nfirst,
+J
the same machine. We have therefore N+1 nodes and 2 arcs, where all the nodes are pairwise

connected except o0, which is connected only to the first operation of each job. Each arc is weighted by a pair of

numbers, {τkl, ηkl}. The first, τkl, is the trail level, while the second is the visibility ηkl, and is computed

according to a desirability measure derived from a greedy problem specific heuristic like the Longest

Processing Time or the Shortest Completion Time. The order in which the nodes are visited by each ant

specifies the proposed solution. For instance, consider a 3x2 problem (3 jobs, 2 machines): it would be

represented in our system by the graph presented in Fig. 15. We suppose the first machine processes operations

1, 3, 5, and the second one the others.

1 4
o o
o0 =0
o =1
11
0
2
o12=4
o o o5 o 21=2
o22=5
o31=3
o32=6
o 6
o
3
Fig. 15. AS graph for a 3 jobs and 2 machines JSP.

All ants are initially in o0; later on they have to identify at each step a feasible permutation of the remaining

nodes. To cope with this problem, transition probabilities have to be slightly modified with respect to those

computed according to formula (4): in order to have a feasible permutation it is in fact necessary to define the

set of allowed nodes in any step not only through the tabu list, but also in a problem-dependent way. For each

ant k, let Gk be the set of all the nodes still to be visited and Sk the set of the nodes allowed at the next step.

25
Initially Gk={1, 2, 3, 4, 5, 6} and Sk={1, 2, 3}. Transition probabilities are computed on the basis of formula

(4), where the set of allowed nodes is equal to Sk. When a node is chosen, it is appended to the tabu list and

deleted from Gk and from Sk; if the chosen node is not the last in its job then its immediate successor in the job

chain is added to Sk. This procedure ensures the possibility to always produce a feasible solution, possibly the

optimal one. The process is iterated until Gk=∅. At the end, the order of the nodes in the permutation given by

the tabu list specifies the solution proposed by ant k. The trails can thus be computed in the usual way and they

are laid down as specified by the ant cycle algorithm.

For example, suppose that an ant yielded the solution π=(0, 1, 4, 6, 2, 3, 5); this would direct the order of the

operations imposing the precedences {(1,3), (3,5), (1,5)} and {(4,6), (6,2), (4,2)}, respectively.

This approach has been implemented and successfully applied to JSP instances of dimension 10x10 and

10x15 (10 jobs, 15 machines). For each of these problems we always obtained a solution within 10% of the

optimum [8], which can be considered a promising result.

VIII. Discussion on some AS characteristics

A major issue in defining any distributed system is the definition of the communication protocol. In the AS a set

of ants communicate by modifications of a global data structure: after each tour the trail left on each ant’s tour

will change the probability with which the same decision will be taken in the future. A heuristic also guides ants

in the early stages of the computational process, when experience has not yet accumulated into the problem

structure. This heuristic automatically loses importance (remember the coefficient ρ related to evaporation) as

the experience gained by ants, and saved in the problem representation, increases.

One way to explain the behavior of AS on the TSP problem is the following. Consider the transition matrix
pkij (t)
pk(t) of ant k: every element is the transition probability from town i to town j at time t as defined by
pk (0)
equation (4). At time t=0 each ij is proportional to ηij, i.e., closer towns are chosen with higher

probability. As the process evolves, pk(t) changes its elements according to (1) and (4). The process can

therefore be seen as a space deformation, in which path cost is reduced between towns which are connected by

edges with a high amount of traffic, and, conversely, path cost is incremented between towns connected by

edges with low traffic levels. From simulations we observed that the matrix pk(t), at least in the range of

optimality for our parameters, converges to a state10 that is very close to stationary (i.e., variations in the

transition matrix pk(t) are very small). When this state is reached the behavior of the ants is dependent on the

10 The stochastic process that rules the evolution of the matrix pk(t) is a Markov process with infinite memory.

26
kind of transition matrix obtained. We observed two situations: in the most rare one, occurring (as we saw in

Section IV) for particular parameter settings, only two transition probabilities are significantly higher than zero

in every row and therefore all the ants choose the same edge at each step and no new tour is searched. In the

most common situations instead, most of the rows have only a few transition probabilities with a significant

value. In these cases search never stops, even if the number of significant transitions is highly reduced, with

respect to the initial situation. Consider for example Fig. 16, obtained as the steady-state transition matrix for a

randomly generated 10-town problem: the area of each circle is proportional to the corresponding value of the

transition probability. An ant in town 1 has a very high probability to go either to town 5 (near 50%) or to town

2 (near 35%), and a low probability of choosing any other edge. A similar analysis holds for ants in any other

town; from towns 9 and 10, for example, any destination is equally probable.

Fig. 16. The steady-state transition matrix for a randomly generated 10-town problem.

Another way to interpret how the algorithm works is to imagine having some kind of probabilistic

superimposition of effects: each ant, if isolated (that is, if α=0), would move with a local, greedy rule. This

greedy rule guarantees only locally optimal moves and will practically always lead to bad final results. The

reason the rule doesn't work is that greedy local improvements lead to very bad final steps (an ant is constrained

to make a closed tour and therefore choices for the final steps are constrained by early steps). So the tour

followed by an ant ruled by a greedy policy is composed of some (initial) parts that are very good and some

(final) parts that are not. If we now consider the effect of the simultaneous presence of many ants, then each one

contributes to the trail distribution. Good parts of paths will be followed by many ants and therefore they

receive a great amount of trail. On the contrary, bad parts of paths are chosen by ants only when they are

obliged by constraint satisfaction (remember the tabu list); these edges will therefore receive trail from only a

few ants.

27
IX. Conclusions

This paper introduces a new search methodology based on a distributed autocatalytic process and its

application to the solution of a classical optimization problem. The general idea underlying the Ant System

paradigm is that of a population of agents each guided by an autocatalytic process induced by a greedy force.

Were an agent alone, both the autocatalytic process and the greedy force would tend to converge to a

suboptimal tour with exponential speed. When agents interact it appears that the greedy force can give the right

suggestions to the autocatalytic process and facilitate quick convergence to very good, often optimal, solutions

without getting stuck in local optima. We have speculated that this behavior could be due to the fact that

information gained by agents during the search process is used to modify the problem representation and in this

way to reduce the region of the space considered by the search process. Even if no tour is completely excluded,

bad tours become highly improbable, and the algorithms search only in the neighborhood of good solutions.

The main contributions of this paper are the following.

(i) We employ positive feedback as a powerful search and optimization tool. The idea is that if at a given

point an agent (ant) has to choose between different options and the one actually chosen results to be

good, then in the future that choice will appear more desirable than it was before11.

(ii) We show how synergy can arise and be useful in distributed systems. In the AS the effectiveness of the

search carried out by a given number of cooperative ants is greater than that of the search carried out by

the same number of ants, each one acting independently from the others.

(iii) We show how to apply the AS to different combinatorial optimization problems. After introducing the

AS by an application to the TSP, we show how to apply it to the ATSP, the QAP, and the JSP.

At the present stage of our investigation we do not have any proof of convergence or bound on the time

required to find the optimal solution. Nevertheless simulations strongly support the following conjecture: with

good parameter settings the algorithm consistently converges in a polynomial number of steps to satisfactory

solutions. We believe our approach to be a very promising one because of its generality (it can be applied to

many different problems, see Section VII), and because of its effectiveness in finding very good solutions to

difficult problems.

11 Reinforcement of this nature is used by the reproduction-selection mechanism in evolutionary algorithms

[23], [30], [33]. The main difference is that in evolutionary algorithms it is applied to favour (or disfavor)
complete solutions, while in AS it is used to build solutions.

28
 Related work can be classified in the following major areas:

• studies of social animal behavior;

• research in "natural heuristic algorithms";

• stochastic optimization.

As already pointed out the research on behavior of social animals is to be considered as a source of

inspiration and as a useful metaphor to explain our ideas. We believe that, especially if we are interested in

designing inherently parallel algorithms, observation of natural systems can be an invaluable source of

inspiration. Neural networks [32], genetic algorithms [23], evolution strategies [30], [33], immune networks [3],

simulated annealing [24] are only some examples of models with a "natural flavor". The main characteristics,

which are at least partially shared by members of this class of algorithms, are the use of a natural metaphor,

inherent parallelism, stochastic nature, adaptivity, and the use of positive feedback. Our algorithm can be

considered as a new member of this class. All this work in "natural optimization" [12] fits within the more

general research area of stochastic optimization, in which the quest for optimality is traded for computational

efficiency.

Acknowledgments

We would like to thank two of the reviewers for the many useful comments on the first version of this paper.

We also thank Thomas Bäck, Hughes Bersini, Jean-Louis Denebourg, Frank Hoffmeister, Mauro Leoncini,

Francesco Maffioli, Bernard Manderik, Giovanni Manzini, Daniele Montanari, Hans-Paul Schwefel and Frank

Smieja for the discussions and the many useful comments on early versions of this paper.

References

[1] E.H.L.Aarts, J.H.M.Korst, Simulated Annealing and Boltzmann Machines, John Wiley, 1988.

[2] J.J.Bentley, "Fast Algorithms for Geometric Traveling Salesman Problems," ORSA Journal on

Computing, 4 (4), 387–411, 1992.

[3] H.Bersini, F.J.Varela, "The Immune Recruitment Mechanism: a Selective Evolutionary Strategy,"

Proceedings of the Fourth International Conference on Genetic Algorithms, Morgan Kaufmann, 520–

526, 1991.

[4] S.C.Boyd, W.R.Pulleyblank, G.Cornuejols, Travel Software Package, Carleton University, 1989.

29
[5] R.E.Burkhard, "Quadratic Assignment Problems," European Journal of Operations Research, 15, 283–

289, 1984.

[6] A.Colorni, M.Dorigo, V.Maniezzo, "Distributed Optimization by Ant Colonies," Proceedings of the

First European Conference on Artificial Life, Paris, France, 134–142, 1991.

[7] A.Colorni, M.Dorigo, V.Maniezzo, "An Investigation of some Properties of an Ant Algorithm,"

Proceedings of the Parallel Problem Solving from Nature Conference (PPSN 92), Brussels, Belgium,

Elsevier Publishing, 509-520, 1992.

[8] A.Colorni, M.Dorigo, V.Maniezzo, M.Trubian, "Ant system for Job–Shop Scheduling," Belgian

Journal of Operations Research, Statistics and Computer Science, 1994, in press.

[9] A.Colorni, M.Dorigo, F.Maffioli, V.Maniezzo, G.Righini, M.Trubian, "Heuristics from Nature for

Hard Combinatorial Problems," Technical Report No. 93–025, Dip. Elettronica e Informazione,

Politecnico di Milano, Italy, 1993.

[10] J.L.Denebourg, J.M.Pasteels, J.C.Verhaeghe, "Probabilistic Behaviour in Ants: a Strategy of Errors?,"

Journal of Theoretical Biology, 105, 259–271, 1983.

[11] J.L.Denebourg, S.Goss, "Collective patterns and decision-making," Ethology, Ecology & Evolution, 1,

295–311, 1989.

[12] M.Dorigo, Optimization, Learning and Natural Algorithms, Ph.D. Thesis, Dip. Elettronica e

Informazione, Politecnico di Milano, Italy, 1992 (in Italian).

[13] M.Dorigo, V.Maniezzo, A.Colorni, "Positive feedback as a search strategy," Technical Report n. 91-

016, Politecnico di Milano, 1991.

[14] S.Eilon, T.H.Watson-Gandy, N.Christofides, "Distribution management: mathematical modeling and

practical analysis," Operational Research Quarterly, 20, 37–53, 1969.

[15] Elshafei A.E., "Hospital Layout as a Quadratic Assignment Problem", Operational Research

Quarterly, 28, 167–179, 1977.

[16] M.Fischetti, P.Toth, "An Additive Bounding Procedure for the Asymmetric Travelling Salesman

Problem," Mathematical Programming, 53, 173–197, 1992.

[17] F.Glover, "Tabu Search—Part I," ORSA Journal on Computing, 1 (3), 190–206, 1989.

[18] F.Glover, "Tabu Search—Part II," ORSA Journal on Computing, 2 (1), 4–32, 1990.

[19] D.E.Goldberg, Genetic Algorithms in Search, Optimization & Machine Learning, Addison-Wesley,

Reading, MA, 1989.

[20] B.Golden, W.Stewart, "Empiric analysis of heuristics," in The Travelling Salesman Problem, E. L.

Lawler, J. K. Lenstra, A. H. G. Rinnooy-Kan, D. B. Shmoys eds., New York:Wiley, 1985.

30
[21] S.Goss, R.Beckers, J.L.Denebourg, S.Aron, J.M.Pasteels, "How Trail Laying and Trail Following Can

Solve Foraging Problems for Ant Colonies," in Behavioural Mechanisms of Food Selection,

R.N.Hughes ed., NATO-ASI Series, vol. G 20, Berlin:Springer-Verlag, 1990.

[22] R.L.Graham, E.L.Lawler, J.K.Lenstra, A.H.G.Rinnooy Kan, “Optimization and Approximation in

Deterministic Sequencing and Scheduling: a Survey”, in Annals of Discrete Mathematics, 5, 287-326,

1979.

[23] J.H.Holland, Adaptation in Natural and Artificial Systems, Ann Arbor: The University of Michigan

Press, 1975.

[24] S.Kirkpatrick, C.D.Gelatt, M.P.Vecchi, "Optimization by Simulated Annealing," Science, 220, 671–

680, 1983.

[25] J. Krarup, P.M. Pruzan, "Computer-aided Layout Design", Mathematical Programming Study, 9, 85–

94, 1978.

[26] E.L.Lawler, J.K.Lenstra, A.H.G.Rinnooy-Kan, D.B.Shmoys eds., The Travelling Salesman Problem,

New York:Wiley, 1985.

[27] S.Lin, B.W.Kernighan, "An effective Heuristic Algorithm for the TSP," Operations Research, 21,

498–516, 1973.

[28] V. Maniezzo, L.Muzio, A.Colorni, M.Dorigo, “The Ant System Applied to the Quadratic Assignment

Problem,” Tech.Rep.94-058, Dip. Elettronica e Informazione, Politecnico di Milano, Milano, 1992 (in

Italian). Submitted to Rivista di Ricerca Operativa.

[29] C.E. Nugent, T.E. Vollmann, J. Ruml, "An Experimental Comparison of Techniques for the

Assignment of Facilities to Locations," Operations Research, 16, 150–173, 1968.

[30] I.Rechenberg, Evolutionsstrategie, Fromman-Holzbog, Stuttgart, Germany, 1973.

[31] G.Reinelt, TSPLIB 1.0, Institut für Mathematik, Universität Augsburg, 1990.

[32] D.E.Rumelhart, J.L.McLelland, Parallel Distributed Processing: Explorations in the Microstructure of

Cogniton, Cambridge, MA: MIT Press, 1986.

[33] H.P. Schwefel, "Evolutionsstrategie und Numerische Optimierung," Ph.D. Thesis, Technische

Universität Berlin, 1975. Also available as Numerical Optimization of Computer Models, J.Wiley,

1981.

[34] D.Whitley, T.Starkweather, D.Fuquay, "Scheduling Problems and Travelling Salesman: the Genetic

Edge Recombination Operator," Proceedings of the Third International Conference on Genetic

Algorithms, Morgan Kaufmann, 1989.

31