Computers & Industrial Engineering 88 (2015) 229–236

Contents lists available at ScienceDirect

Computers & Industrial Engineering

journal homepage: www.elsevier.com/locate/caie

Profit maximization of TSP through a hybrid algorithm

Aditi Khanra a, Manas Kumar Maiti b,⇑, Manoranjan Maiti c,*
a
Bohichberia High School (H.S.), Bohichberia, Purba Medinipur, West Bengal 721649, India
b
Department of Mathematics, Mahishadal Raj College, Mahishadal, Purba-Medinipur, WB 721628, India
c
Department of Applied Mathematics with Oceanology and Computer Programming, Vidyasagar University, Paschim-Medinipur, WB, India

a r t i c l e i n f o a b s t r a c t

Article history: Here a new model of Traveling Salesman Problem (TSP) with uncertain parameters is formulated and
Received 24 November 2014 solved using a hybrid algorithm. For this TSP, there are some fixed number of cities and the costs and time
Received in revised form 12 June 2015 durations for traveling from one city to another are known. Here a Traveling Salesman (TS) visits and
Accepted 24 June 2015
spends some time in each city for selling the company’s product. The return and expenditure at each city
Available online 15 July 2015
are dependent on the time spent by the TS at that city and these are given in functional forms of t. The
total time limit for the entire tour is fixed and known. Now, the problem for the TS is to identify a tour
Keywords:
program and also to determine the stay time at each city so that total profit out of the system is maxi-
Ant colony optimization
Particle swarm optimization
mum. Here the model is solved by a hybrid method combining the Particle Swarm Optimization (PSO)
Travel cost and Ant Colony Optimization (ACO). The problem is divided into two subproblems where ACO and
Travel time PSO are used successively iteratively in a generation using one’s result for the other. Numerical experi-
Profit ments are performed to illustrate the models. Some behavioral studies of the models and convergences
Hybrid algorithm of the proposed hybrid algorithm with respect to iteration numbers and cost matrix sizes are presented.
Ó 2015 Published by Elsevier Ltd.

1. Introduction TSP (PTSP) (Bianchi, Gambardella, & Dorigo, 2002), etc. are worth
mentioning. In TSP with precedence constraints, there exists an
The TSP (Applegate, Bixby, Chvtal, & Cook, 2007) is one of the order in which the cities should be visited. In Asymmetric TSP,
most widely studied NP-hard combinatorial optimization problem cost of traveling from vertex (node/city) v i to v j is not equal to
which cannot be solved exactly in polynomial time. It is also an the cost of traveling from vertex v j to v i . In Stochastic TSP, each
important research topic. This problem is, for a given set of n cities vertex is visited with a given probability and goal is to minimize
with travel costs (or distances) cij between each pair of cities the expected distance/cost of a priori tour. In TSP with time
i; j 2 ½1 : n, to determine a minimum cost (or distance) circuit windows, each vertex is visited within a specified time windows.
(Hamiltonian circuit or cycle) passing through each vertex once PTSP is a TSP problem where each customer has a given
and only once. Every such tour together with a start city can probability of requiring a visit.
be characterized by the permutation of all cities as they are During the last decades, several algorithms emerged to approx-
visited along the tour. TSP has important applications to real imate the optimal solution of TSP such as Tabu search method
world problems, such as vehicle routing problem, mixed chinese (Fiechter, 1994), Neural networks (Leung, Jin, & Xu, 2004;
postman problems, and printed circuit board punching sequence Masutti & Castro, 2009), Simulated annealing (Lo & Hus, 1998),
problems. Genetic algorithms (GA) (Cheng & Gen, 1994; Cheng, Gen, &
Different types of TSPs have been solved by the researchers Sasaki, 1995; Liu, 2010; Majumdar & Bhunia, 2011; Nagata &
during last two decades. Among these, TSP with time windows Soler, 2012; Yang, Wu, Lee, & Liang, 2008), Ant Colony
(Focacci, Lodi, & Milano, 2002), Stochastic TSP (Chang, Wan, & Optimization (ACO) (Bai, Yang, Chen, Hu, & Pan, 2013; Cheng &
OOI, 2009; Liu, 2010), Double TSP (Petersen & Madsen, 2009), Mao, 2007; Dorigo & Gambardlla, 1997; Ghafurian & Javadian,
Asymmetric TSP (Majumdar & Bhunia, 2011), TSP with prece- 2011; Ibanez & Blum, 2010), and Particle Swarm Optimization
dence constraints (Moon, Ki, Choi, & Seo, 2002), Probabilistic (PSO) (Chen & Chien, 2011; Lin, Chen, & Lin, 2009).
ACO is an important soft computing technique for solving opti-
⇑ Corresponding authors. mization problems. In ACO, the behavior of real ants to find the
E-mail addresses: aditikhanra@gmail.com (A. Khanra), manasmaiti@yahoo.co.in shortest path between their nest and food sources, has been used.
(M.K. Maiti), mmaiti2005@yahoo.co.in (M. Maiti). Several ACO algorithms are available to solve the well-known

http://dx.doi.org/10.1016/j.cie.2015.06.018
0360-8352/Ó 2015 Published by Elsevier Ltd.
230 A. Khanra et al. / Computers & Industrial Engineering 88 (2015) 229–236

NP-Hard TSP. Dorigo and Gambardlla (1997) described an artificial the TS is to fix the tour program and the stay times at each city
ant colony capable of solving the TSP. In 2007, Cheng and Mao pre- so that total profit out of the tour is maximum. Thus TSP is formu-
sented a modified ant colony system for solving the TSP with time lated as maximization problem with deterministic returns and stay
windows. Ibanez and Blum proposed a Beam-ACO which is a expenditures.
hybrid method combining ACO with beam search to solve TSP in In comparison to the usual TSP, here the proposed TSP consists
2010. In 2011, Ghafurian and Javadian proposed an ACO algorithm of two sub-optimization models – (i) minimization of total travel
for solving fixed destination multi-depot multiple traveling costs between the cities and (ii) allocation of stay times at the cities
salesmen problems. In 2013, Bai et al. and Pan proposed a model so that total outcome due to stay is maximum. Then a trade off
inducing max–min ant colony optimization for Asymmetric TSP. between these two sub-models is made. For these two fold TSP
PSO is a heuristic optimization technique based on swarm intel- optimization problem, a hybrid algorithm combining the algorithms
ligence that is inspired by the behavior of birds flocking. It was first of ACO and PSO is designed and applied successfully. Here ACO
proposed by Kennedy and Eberhart (Eberhart & Kennedy, 1995; and PSO are used successively and iteratively in a generation using
Kennedy & Eberhart, 1995) and has received significant attention. one’s result for the other. The proposed TSP is illustrated with
Many studies modified the PSO algorithm to improve its perfor- numerical examples. Some interesting model behaviors are
mance in continuous optimization. Among them, one important presented with different sizes of cost matrices. It is shown that
improvement was adding the inertia weights into the updating more stay at stations does not fetch more profit. Some parametric
rules of PSO (Shi & Eberhart, 1998). Ratnaweera, Halgamuge, and studies of the proposed hybrid algorithm with respect to iteration
Watson (2004) used time-varying strategies of inertia weights. numbers and cost matrices sizes are also presented.
Later, Clerc and Kennedy (2002) presented a construction factor Rest of the paper is organized as follows. Models are formulated
to form a new inertia weights, which has proven effective and is in Section 2. In Section 3, Hybrid ACO–PSO system is described. In
the current state of the art. A large number of different algorithms Section 4, experimental results are presented. Finally a brief dis-
have been developed to find suboptimal solutions for the TSP in cussion, models’ behavioral and different sensitivity studies and
polynomial time. conclusions are respectively drawn in Sections 5–7.
Inspite of the above developments, there are some lacunas/gaps
in forming the realistic TSPs, These are:
2. Model formulation
 Though the visit of a TS is organized in order to get a return for
the company, till now, none has considered the returns at each 2.1. Classical TSP for minimum total travel cost (Model1A)
city out of the tour.
 Most of the TSPs are concerned with the minimization of tour In a classical two-dimensional TSP, TSP can be represented as
cost or travel time. But, for selling or canvassing a product, a graph G ¼ ðV; EÞ, where V ¼ 1; 2; . . . ; N is the set of nodes and E is
TS has to spend some time at each city and incurs some expen- the set of edges. A salesman has to travel N cities at minimum cost.
diture for this. This also has been overlooked by the TSP In this tour, salesman starts from a city, visits all the cities exactly
researchers. once and comes to the starting city using minimum cost. Let cij be
 Normally, a TS is asked by his/her company to finish the entire the cost for traveling from i-th city to j-th city. Then the model is
activities including the tour(travel and stay times) within a mathematically formulated as (Dantzig, Fulkerson, & Johnson,
specified time limit. This constraint has been taken into account 1954):
by very few investigators only. 9
Determine xij ; i ¼ 1; 2; . . . N; j ¼ 1; 2; . . . N: >
 Though ultimate goal of a company is to make the profit >
>
>
>
through the sales representative, till now, no TSP has been for- X
N X
N >
>
to minimize Z ¼ xij cij ; >
>
mulated as a profit maximization problem considering returns >
>
i¼1 j¼1 >
>
and expenditures. >
>
>
>
 In order to bridge the above gaps, one new TSP has been formu- X
N >
>
=
subject to xij ¼ 1; j ¼ 1; 2; . . . ; N
lated and solved. ð1Þ
i¼1 >
>
>
>
The above assumptions are realistic and have applications in X
N >
>
>
>
medicine producing firms with respect to their medical represen- xij ¼ 1; i ¼ 1; 2; . . . ; N >
>
>
>
tatives and salesmen. For a medical firm, normally a medical j¼1 >
>
XX >
>
>
representative/salesman is asked to tour a number of xij 6 jSj  1; 8S  V; xij 2 f0; 1g: >
>
;
towns/sub-towns within a limited time. In this process, he/she i2S j2S
fixes the paths and spends some times at the towns/stations in
such way that total outcome/benefit of the tour is maximum. where xij ¼ 1 if the salesman travels from city-i to city-j, otherwise
He/she incurs some costs for travel, pays for stay at stations and xij ¼ 0.
earns indirectly through Doctors/Medical shops by canvassing Let ðx1 ; x2 ; . . . ; xN ; x1 Þ be a complete tour of a salesman, where
and presentation. All these are done within the limited time fixed xi 2 f1; 2; . . . ; Ng for i = 1, 2, . . . , N and all xi ’s are distinct. Then
by the medical firm. The proposed TSP incorporating the above the above model reduces to
assumptions are most appropriate for the real life medical house 9
problems and incorporated in the proposed model. Determine a complete tour ðx1 ; x2 ; . . . ; xN ; x1 Þ >
=
In this paper, for the proposed TSP, a TS visits some fixed X
N1
ð2Þ
number of cities following the TSP rule and spends sometime at to minimize Z ¼ cxi ;xiþ1 þ cxN ;x1 >
;
i¼1
each visited city. He/she earns some returns and incurs some
expenditure at each city and these are stay time dependent in Classical TSP with time for minimum total travel time
some functional forms of time. Here total allowable time for the (Model1B): Let t ij be the time for traveling from i-th city to j-th
entire tour including stay times is fixed. Now, the problem for city. In this case, the model formulation is the same as ð1Þ except
 A. Khanra et al. / Computers & Industrial Engineering 88 (2015) 229–236 231

that the objective function is replaced by the following. formulated The defined single objective for the crisp TSP is solved by the
as: hybrid algorithm developed in Section 3.
)
X
N X
N
To minimize Z ¼ xij t ij ð3Þ
i¼1 j¼1 3. Hybrid ACO–PSO system
All other constraints remain unaltered.
Let ðx1 ; x2 ; . . . ; xN ; x1 Þ be a complete tour of a salesman, where During the last decade, nature inspired intelligence became
xi 2 f1; 2; . . . ; Ng for i = 1, 2, . . . , N and all xi ’s are distinct. Then increasingly popular through the development and utilization
the above model reduces to of intelligent paradigms in advance information systems design.
9 Among the most popular nature inspired approaches, when task
Determine a complete tour ðx1 ; x2 ; . . . ; xN ; x1 Þ > is to optimize with in complex decisions of data or information,
=
X
N1
ð4Þ ACO and PSO draw significant attention. The foraging behavior of
To minimize Z ¼ txi ;xiþ1 þ txN ;x1 >
; real ants have inspired the development of a large number of
i¼1
ant-based algorithms (AA) by the researchers to solve mainly
combinatorial optimization problems defined over discrete
2.2. Proposed TSP for maximum profit (Model2) search spaces. One of the first behaviors studied by ethologists
was the ability of ants to find the shortest path between their
Here a model is considered where the salesman spends some nest and food source. Forager ants decide which path to follow
time at each city to convince the customers and hence due to this, based on the pheromone concentrations on different paths.
makes some profit and incurs some expenditure also. Amount of Paths with a larger pheromone concentration have a higher
profit depends on the duration of time he spends in the city. He probability of being selected. As more ants follow a specific trail,
bears some expenditure and makes profit per unit time from the the desirability of that path is reinforced by more pheromone
city. Let cij and t ij be the cost and time respectively for traveling being deposited by the foragers, which attracts more ants to
from i-th city to j-th city. ti be the spent time in i-th city. oi ðtÞ is follow that path. Mimicking this behavior of ant, a class of
the output/return ei ðtÞ, is the expenditure where optimization algorithms named ACO has been developed by
oi ðtÞ ¼ ai þ bi t  ci t2 and ei ðtÞ ¼ e0i t for spending t time by TS in several authors (Bai et al., 2013; Cheng & Mao, 2007; Dorigo &
the i-th city. Here a constraint on total time used by the salesman Gambardlla, 1997; Ghafurian & Javadian, 2011; Ibanez & Blum,
is imposed. He can at most use H units of time for his total tour. So 2010).
the model is mathematically formulated as: PSO algorithm was an intelligent technology first presented
9 in 1995 by Eberhart and Kennedy, and it was developed
Determine xij ; and ti ; i ¼ 1; 2; . . . N; j ¼ 1; 2; . . . N: >
> under the inspiration of behavior laws of bird flocks, fish
>
>
>
> schools and human communities. PSO algorithms has been
X
N X
N X
N >
>
to maximize Z ¼ ðoi  ei Þ  xij cij >
>
>
> developed by several authors (Chen & Chien, 2011; Lin et al.,
>
>
i¼1 i¼1 j¼1 >
> 2009).
>
>
>
> Both algorithms PSO and ACO have some merits and demerits
X
N >
>
subject to xij ¼ 1; j ¼ 1; 2; . . . ; N >
> in solving combinatorial optimization problems. Here combining
>
>
i¼1 >
> the features of these algorithms, a hybrid algorithm is proposed
=
ð5Þ to solve TSP in some modified form. The algorithm can solve
X
N
>
xij ¼ 1; i ¼ 1; 2; . . . ; N >
> single-objective crisp TSP. In the algorithm, initially ant colony
>
>
j¼1 >
> system is used to produce a set of paths (tours) for the salesman
>
>
XX >
> which is a set of potential solutions for the ACO part of the
>
xij 6 jSj  1; 8S  V xij 2 f0; 1g >
>
> algorithm. For each path of the salesmans stay times at different
>
>
i2S j2S >
> cities to maximize the profit from the system is determined by
>
>
>
> the PSO algorithm. To do this, for each path a set of stay times
X
N X
N X
N >
>
such that t ij xij þ ti 6 H >
>
>
; at different nodes are randomly generated. Each set of stay time
i¼1 j¼1 i¼1 for the path is a particle for the PSO. A set of such particles is
where xij ¼ 1 if the salesman travels from city-i to city-j, otherwise generated for a path which is initial swarm of the PSO for that
xij ¼ 0, H = total allowable time for the entire tour. particular path. PSO operations are made on this set of particles
Let ðx1 ; x2 ; . . . ; xN ; x1 Þ be a complete tour for the salesman, where a finite number of times to determine the optimum stay time at
xi 2 f1; 2; . . . ; Ng for i = 1, 2, . . . , N and all xi ’s are distinct. Then the different nodes for the path. In this way, PSO operation is done
above model reduces to for every path of the salesman (ant) to determine optimal stay
9 times at different nodes for the respective path. The hybrid
Determine a complete tour ðx1 ; x2 ; . . . ; xN ; x1 Þ and ðt 1 ; t 2 ; . . . ; t N Þ >
> ACO–PSO system is presented below. In the algorithm, sij repre-
>
>
>
> sents amount of pheromone lies on the path between node i and
To maximize Z ¼ ðo1 ðtÞ þ o2 ðtÞ þ    þ oN ðtÞÞ >
>
>
> node j; iter1 and iter2 represent iteration counters, maxiter and
" # >
>
XN1 = maxgen represent maximum iteration number of the ACO algo-
 cxi ;xiþ1 þ cxN ;x1 þ ðe1 ðtÞ þ e2 ðtÞ þ    þ eN ðtÞÞ rithm and maximum iteration number in PSO part where
>
>
i¼1 >
> maxiter ¼ 500 and maxgen ¼ 10; n (n ¼ 50) represent number of
>
>
>
>
X
N1 >
> ants or population size and N (N ¼ 10) represents number of
subject to t xi ;xiþ1 þ t xN ;x1 þ ðt 1 þ t 2 þ    þ t N Þ 6 H >
>
; nodes/cities in the problem. Here a set of particles of size M
i¼1
ðM ¼ 10Þ is used. The algorithm of the hybrid ACO–PSO is as
ð6Þ follows.
232 A. Khanra et al. / Computers & Industrial Engineering 88 (2015) 229–236

Hybrid Algorithm (HA): (c) Path Construction: To construct a path X k for k-th ant, fol-
lowing steps are followed:
1. Set iter1 ¼ 0 and initialize maxiter; maxgen.
2. Initialize pheromo sij for i ¼ 1; 2; . . . ; N and j ¼ 1; 2; . . . ; N. i. Let NS ¼ f1; 2; . . . ; Ng and l ¼ 1.
3. Do ii. xkl = A random element from the set NS.
4. Construct path of n ants, i.e., create n tours iii. Let NS ¼ NS  fxkl g.
X i ¼ ðxi1 ; xi2 ; . . . ; xiN ; xi1 Þ; i ¼ 1; 2; . . . ; n using sij . iv. Let node i be the present position of the ant, i.e., xkl ¼ i. Then
/*For k-th path X k ¼ ðxk1 ; xk2 ; . . . ; xkN ; xk1 Þ determine next node j 2 NS is selected by the ant with a probability pij
proportion of optimum stay times at different cities given by the formula.
ðOTPk1 ; OTPk2 ; . . . ; OTPkN Þ to maximize profit for the path saij
using PSO. So ðOTPk1 ; OTPk2 ; . . . ; OTPkN Þ is a particle for pij ¼ P
j2NS ijsa
the PSO*/
5. Do for each path k where a is a positive constant used to amplify the influence
6. Randomly generate a set of particles of size M for of pheromone concentrations. Roulette–Wheel selection pro-
the k-th path TP ki ¼ ðTP ki1 ; TP ki2 ; . . . ; TPkiN Þ, i ¼ 1; 2; . . . M, cess (Michalewicz, 1992) is used for the purpose.
where each TPkij 2 ð0:0001; 1Þ. TP ki is i-th particle for the v. l ¼ l þ 1; xkl ¼ j.
swarm of path X k . For each TPki stay time at j-th node tkij vi. if l < N goto step (iii).
is calculated as below where T ¼ H  T k is the total stay
time at different nodes, H is the total tour time and T k is n-such paths are constructed for n ants.
travel time for path X k . (d) For each path X k do the following operations:
t kij ¼ T  TPkij =ðTPki1 þ TP ki2 þ    þ TPkiN Þ for
(i) Swarm initialization: For each tour, a set of proportion of
j = 1, 2, . . . , N and i = 1, 2, . . . , M. So stay time at different
stay times at different nodes are generated. Each set of pro-
cities for k-th path is represented by the vector
portion of stay times at different nodes is considered as a
t ki ¼ ðt ki1 ; tki2 ; . . . ; tkiN Þ.
particle. A set of M such particles are randomly generated
7. For each particle TP ki do
which is swarm for the path. PSO operations are made on
8. initialize velocity V ki .
this swarm to improve the profit from the path. For every
9. End for
path total tour completion time H is given. Total stay time
10. iter2 ¼ 0
in different nodes due to k-th path T ¼ H  T K , where T K is
11. Do
the travel time of path X k . i-th particle of the swarm, is con-
12. For each particle TPki do
sidered as N component vector TP ki ¼ ðTP ki1 ; TP ki2 ; . . . ; TPkiN Þ,
13. Find the personal best position PBTP ki .
where each TPki 2 ð0:0001; 1Þ. TPki is i-th particle for the
14. End for
swarm of path X k . For each particle velocity is initialized
15. Find the global best position GBTPk of the
between V max and V min . For each TPki stay time at different
swarm.
nodes are calculated as tkij ¼ T  ðTPkij Þ=ðTPki1 þ TPki2 þ
16. For each particle TPki do
   þ TPkiN Þ for j = 1, 2, . . . , N and i = 1, 2, . . . , M. so
17. Update the velocity V ki .
t ki ¼ ðtki1 ; tki2 ; . . . ; tkiN Þ.
18. Update the position TP ki .
(ii) Find global best position: Initially profit of the path X k due
19. End for
to different particles TP ki are calculated. The particle which
20. iter2 ¼ iter2 þ 1
gives maximum profit is considered as initial value of
21. while(iter2 < maxgen).
GBTP k . After iteration when particles get new positions
22. Determine stay times of X k from GBTP k
GBTP k is updated if a position gives better profit than GBTP k .
23. Calculate profit for X k .
(iii) Find personal best position: Initial position of a particle is
24. End Do.
taken as initial personal best position PBTPki . After each iter-
25. Made pheromone evaporation.
ation if new position gives better profit then PBTPki is
26. Updated pheromone for all the paths.
replaced by new position.
27. Find best solution.
(iv) Updating velocity: For each particle (i.e., stay times at
28. iter1 ¼ iter1 þ 1
different nodes) of a path velocity is updated using the
29. While (iter1 < maxiter).
following equation.
30. Output: best solution.
31. End Algorithm. V kij ðt þ 1Þ ¼ w  V kij ðtÞ þ C 1 r 1jðtÞ ½PBTP ki  PPBki 
þ C 2 r 2jðtÞ ½GBTPk  PPBki 
The graphical representation of this algorithm is given in Fig. 1 where PPBki is present position of i-th particle of swarm
of Section 3.2. k; V kij ðtÞ is velocity of particle i in j-th dimension at time step
t and V kij ðt þ 1Þ is velocity of particle i in j-th dimension at
3.1. Procedures for the proposed hybrid algorithm time step t þ 1. C 1 and C 2 are positive acceleration constant
and here C 1 ¼ 1:49618, C 2 ¼ 1:79618 and w ¼ 0:7298. r 1j ðtÞ
(a) Representation: Here a complete tour on N cities represents a and r 2j ðtÞ  Uð0; 1Þ are random values in ½0; 1.
path of an ant, i.e., a potential solution for the ACO part of the algo- (v) Updating position: Each particle of k-th swarm is updated
rithm. A ‘n dimensional integer vector’, X k ¼ ðxk1 ; xk2 ; . . . ; xkN Þ, is used using the following equation
to represent a solution, where xk1 ; xk2 ; . . . ; xkN represent N consecu-
tive cities in a tour and k ¼ 1; 2; . . . ; n where n is number of ants. PPBki ðt þ 1Þ ¼ PPBki ðtÞ þ V ki ðt þ 1Þ
(b) Pheromone Initialization: As aim of a TSP is to maximize
the profit from a tour, i.e., to minimize the cost of a tour, it is (e) Find fittest solution: To find fittest solution X k of n solu-
assumed that initial value of pheromone sij ¼ 1=c1:5 ij . tions, following step is followed:
 A. Khanra et al. / Computers & Industrial Engineering 88 (2015) 229–236 233

For maximization of crisp, objective a solution X k dominates X j (g) Pheromone Updating: Once all ants have been constructed,
if objective value of X k is grater than the objective value their complete tour pheromone is increased on the paths through
of X j . which the ants move. Depending upon the nature of the problem,
(f) Pheromone Evaporation: For evaporation of pheromone, after a complete tour pheromone is increased using the following
the following formula is used rules.
For first crisp problem, if P k be the profit of path X k , then for this
sij ¼ ð1  qÞsij path sxki xkiþ1 is increased by 1=P bk , where b is a parameter used to
best fit the updating function.
where q is in ½0; 1. The constant q specifies the rate at which
pheromone evaporates, causing ants to forget previous
decisions. 3.2. Graphical representation of proposed hybrid algorithm

Fig. 1. Graphical representation of proposed hybrid algorithm.

234 A. Khanra et al. / Computers & Industrial Engineering 88 (2015) 229–236

Table 1
Travel costs/times between different cities.

i/j 0 1 2 3 4 5 6 7 8 9
0 –/– 25/5 28/5 32/4 20/6 26/5 37/3 8/7 29/5 20/6
1 37/3 –/– 20/6 28/5 35/3 40/2 30/5 42/2 28/6 4/7
2 42/2 28/6 –/– 30/5 25/7 35/4 9/9 32/5 40/2 30/5
3 28/6 30/5 7/9 –/– 20/8 25/7 30/5 35/4 22/8 37/3
4 37/3 22/8 35/4 30/5 –/– 20/8 25/7 30/5 9/9 28/6
5 25/7 30/5 25/7 8/9 28/6 –/– 32/5 40/2 32/5 30/5
6 28/6 25/7 30/5 22/8 37/3 40/2 –/– 10/9 32/5 20/8
7 20/8 5/9 32/5 40/2 35/4 25/7 40/2 –/– 22/8 37/3
8 30/5 40/2 35/4 25/7 20/8 22/8 37/3 32/5 –/– 28/6
9 28/6 30/5 28/6 20/8 11/7 32/5 37/3 40/2 30/5 –/–

values, ðcl ; cu Þ and ðtl ; tu Þ (using a relation with travel cost for the
Table 2 present problem). The above ranges differ for different problems.
Returns ðoi Þ’s with constants ai ; bi ; ci and expenditure ðe0i Þ in different cities. Random data sets are generated using rand() function in C++
i/j 0 1 2 3 4 5 6 7 8 9
language. For the present problem, t ij s are obtained as

ai 150 60 140 70 80 110 90 100 60 140 tij ¼ random no in ½t l ; tu  þ ðcn =cij Þ

bi 24 33 60 47 65 64 31 18 110 116
ci 2 1 3 4 2 3 2 1 4 5 where t l ¼ 0; t u ¼ 5; cn ¼ 200. The returns and expenditures at dif-
ðe0i Þ 8 9 12 7 13 10 7 8 14 6 ferent cities for large TSPs are assumed as the returns oi ðtÞs and
expenditures ei ðtÞs at the i-th city following the earlier expressions.
4. Numerical illustration Here ai ’s, bi ’s, ci ’s and ci0 s for first ten cities are given in Table 2 and
for other cities, there values are repeated in the same order at the
4.1. Input test data for different models (1A,1B and 2) interval of 10 cities, i.e., values at (10th, 20th, . . .),
(11th, 21st, . . .), . . . , (29th, 39th, . . .) cities are same as the values
The models (Model1A, Model1B, Model2) are illustrated for ten at 0th, 1st, . . . , 9th cities respectively. With these input data, TSPs
cities (N = 10). The assumed values of travel costs and times of sizes ð20  20Þ; . . . ; ð60  60Þ and ð100  100Þ are solved by the
between different cities are presented in Table 1 (times are pre- proposed HA and the optimum results are presented in Table 4.
sented under slash(/)). The time dependent return and expenditure Here, it is observed that the behaviors of ‘Return’, ‘Expenditure’
of the i-th city are oi ðtÞ ¼ ai þ bi t  ci t2 and ei ðtÞ ¼ e0i t respectively and ‘Profit’ for large-size problems are same as the ð10  10Þ TSP
in crisp form. The values in oi ðtÞ and ei ðtÞ are presented in Table 2. (Table 3). For all problems, ‘Returns’ and ‘Expenditures’ increase
For these models, the total tour time is limited to H hours. with total tour time, but their rates of increase are such that the
total profit increases initially and after sometime, it decreases.
4.2. Solutions of different TSP Thus the proposed hybrid algorithm works well even for larger
TSPs.
The Models 1A and 1B which are respectively the minimiza-
tion of total travel cost and travel time (without stay times) are 5. Discussion
solved by ACO only. These results are presented in Table 3. The
Model-2 is solved for the above input data by the proposed HA out- In Table 3, the results of Models-1A and -1B are as per our
lined in section 3. The optimum results are presented in Table 3. expectations. When travel cost is minimized (Model-1A), the
Here H is the total tour time in hours, PN model type, ST Stay time, minimum travel cost is 111$ but required travel time is very high
TC Travel cost, TT Travel time. (80 h). On the contrary, for Model-1B, it is reversed i,e. travel time
is minimum (27 h) but the travel cost is highest (376$). For
4.3. TSP with virtual data set Model-2 when total tour time (H) is fixed with variable stay times
at different cities, TS tries to earn as much as possible allocating
Here we generate large TSP problems with virtual data set, possible maximum stay time at stations out of the available total
travel costs ðcij Þ and travel times ðtij Þ for 20, 30, 40, 50, 60 and 100 tour time as it is a problem of maximization of profit. In this pro-
cities. These values are randomly generated within some specified cess, TS adopts the minimum tour time accepting highest travel

Table 3
Total tour times (H hours), Optimum paths, Stay Time (ST hours), Travel Costs (TC in $), Travel Time (TT hours), Returns (R), Expenditure (E) and Profit (in $) for different models.

PN H Path TT ST TC R E Profit
1A – 0, 7, 1, 9, 4, 8, 5, 3, 2, 6 80 – 111 – – –
1B – 6, 4, 0, 2, 8, 1, 5, 7, 3, 9 27 – 376 – – –
2 80 7, 3, 9, 6, 4, 0, 2, 8, 1, 5 27 1.06, 3.57, 8.57, 2.25, 8.58, 1.76, 5.75, 10.24, 4.49, 6.74 376 3643.41 546.30 2721.11
100 4, 0, 6, 2, 8, 1, 5, 7, 3, 9 31 11.14, 2.85, 4.38, 7.18, 11.09, 8.35, 7.64, 2.08, 4.17, 10.11 352 3970.68 697.83 2920.85
120 6, 5, 1, 9, 4, 0, 7, 3, 2, 8 47 5.06, 8.02, 9.14, 10.32, 11.55, 3.10, 2.83, 4.35, 7.15, 11.47 254 4030.69 734.30 3042.40
140 5, 3, 2, 0, 7, 1, 9, 4, 8, 6 64 8.21, 4.64, 7.51, 3.09, 3.40, 9.83, 10.51, 12.12, 11.46, 5.22 171 4072.67 762.80 3138.87
160 3, 2, 6, 0, 7, 1, 9, 4, 8, 5 80 4.55, 7.48, 5.54, 3.53, 4.32, 10.91, 10.66, 12.59, 11.81, 8.62 111 4121.81 800.45 3210.36
165 2, 6, 0, 7, 1, 9, 4, 8, 5, 3 80 8.00, 6.01, 4.01, 4.99, 12.00, 11.00, 13.00, 11.99, 8.99, 5.01 111 4172.95 845.96 3216.00
170 9, 4, 8, 5, 3, 2, 6, 0, 7, 1 80 11.21, 13.56, 12.25, 9.45, 5.26, 8.23, 6.83, 4.37, 5.78, 13.07 111 4213.12 891.64 3210.48
190 1, 9, 4, 8, 5, 3, 2, 6, 0, 7 80 17.21, 12.06, 15.57, 13.26, 10.74, 6.19, 9.56, 8.62, 6.72, 10.07 111 4273.82 1075.42 3087.40
210 3, 2, 6, 7, 1, 9, 0, 4, 8, 5 81 7.29, 10.89, 10.57, 13.94, 21.06, 12.79, 8.56, 17.41, 14.24, 12.25 122 4179.04 1250.11 2806.92
230 2, 6, 7, 1, 9, 0, 4, 8, 5, 3 81 12.36, 12.72, 18.19, 24.82, 13.60, 10.59, 19.64, 15.36, 13.36, 8.37 122 3920.14 1435.01 2363.13
250 6, 7, 1, 9, 0, 4, 8, 5, 3, 2 81 14.75, 22.27, 29.39, 14.36, 12.62, 21.41, 16.26, 14.80, 9.26, 13.88 122 3495.28 1618.41 1754.87
 A. Khanra et al. / Computers & Industrial Engineering 88 (2015) 229–236 235

Table 4
Total tour times (H hours), Stay Time (ST hours), Travel Costs (TC in $), Travel Time (TT hours), Returns (R), Expenditure (E) and Profit (in $) for different models.

Sl no. Problem size H TT ST TC Returns (R) Expenditure (E) Profit

1 20 400 238 162 409 8227.72 1603.78 6214.22
500 336 164 327 8254.64 1620.60 6307.04
600 382 218 295 8472.53 2105.71 6071.82
2 30 700 444 256 524 12,412 2524.07 9363.93
800 534 266 451 12,493 2596.18 9445.82
900 561 339 430 12,691 3299.95 8961.79
3 40 800 472 328 829 16302.48 3196.69 12276.79
900 558 342 698 16450.70 3335.28 12417.42
1000 644 356 681 16518.70 3452.68 12388.02
4 50 1100 674 426 985 20498.65 4129.34 15384.31
1200 780 420 851 20,536 4123.27 15462.62
1300 809 491 810 20958.9 4772.76 15376.14
5 60 1300 820 480 1173 24248.45 4711.31 18364.14
1400 904 496 999 24335.02 4827.35 18508.86
1500 954 546 1003 24829.67 5321.18 18505.49
6 100 2700 1003 1797 1596 36918.66 8321.34 27001.32
2900 1778 1122 1553 38075.9 8101.29 28421.64
3000 1965 1035 1682 38538.64 8645.28 28211.36

cost. This practice is followed up to H = 80 h. After this, as H

increases, TS makes a trade off between the savings due to stay
and tour cost. To reduce the travel cost, TS increases the travel
time, i.e. takes more time than the minimum travel time and the
rest time (from the total tour time) as stay-time at different sta-
tions, because more time is available now. By this arrangement,
profit increases with H up to H = 165 h. After this, profit decreases
when H is more because stay at different stations is not beneficial
any more, this is more clearly explained in Fig. 2.

6. Model’s behavioral studies

In this section, changing some model’s parameters, the behavior

of the models are determined and presented graphically.
Fig. 2. Tour time vs profit for Model-2 for 10 city.
6.1. Profit vs tour time

For Model2, maximum profits are evaluated and plotted against

different tour times in Fig. 2. Here profit is maximum at TT ¼ 165.
It is interesting to note that for the model, as total tour time
increases, profit also increases up to certain tour time and after
that time profit decreases with the increase of tour time. In this
process of evaluation, a TS first tries to make his total return
P
( oi ) maximum allotting the required total stay time for this from
the tour time and then suitable tour path is selected against the
travel time which is equal to available remaining time (i.e.
total tour time  total stay time). Obviously, if the available travel
time is more, TS adopts the path which gives comparatively mini-
mum tour cost. But when the total tour time is very high, TS adopts
the lowest tour cost path and the rest time is spent as total stay
time. As we know, if the stay time at each city exceeds a certain
limit, then expenditure is more than the return and due to these
reasons, we observe that for the model, profit goes down after cer-
Fig. 3. Problem size vs iteration for Model-2.
tain tour time, this study will be very helpful for the managers to
fix/allot the total tour times for TSs.
onwards, number of iterations required for optimum values
6.2. Model size vs iteration number increases sharply. But when matrix size increases from ð5  5Þ size
to ð10  10Þ size, the increase in iteration number is not much
Here a study is made for the number of iterations required for sharp. On the other hand, when matrix size increases from
optimum results against different sizes of cost matrices. For the ð60  60Þ size to ð100  100Þ size, the increase in iteration number
Model2, optimum results were derived with ð5  5Þ; . . . ; is very sharp. This is as per our expectation. As the problem is a
ð10  10Þ; ð20  20Þ; . . . ; ð60  60Þ; ð100  100Þ matrices and the NP-hard problem, the complexity and time increase with size of
corresponding iteration numbers were noted. These are plotted the cost matrices and hence the number of iterations for optimum
in Fig. 3 for Model2. It is to be noted that ð10  10Þ matrix size results sharply increases with model sizes.
236 A. Khanra et al. / Computers & Industrial Engineering 88 (2015) 229–236

Bianchi, L., Gambardella, L. M., & Dorigo, M. (2002) pp. 883–892. An ant colony
optimization approach to the probabilistic traveling salesman problem (2439).
Berlin Heidelberg: Springer-Verlag.
Chang, T., Wan, Y., & OOI, W. T. (2009). A stochastic dynamic travelling salesman
problem with hard time windows. European Journal of Operational Research,
198(3), 748–759.
Chen, S., & Chien, C. (2011). parallelized genetic ant colony systems for solving the
traveling salesman problem. Expert Systems with Applications, 38, 3873–3883.
Cheng, R., & Gen, M. (1994). Crossover on intensive search and traveling salesman
problem. Computers and Industrial Engineering (14), 485–488.
Cheng, R., Gen, M., & Sasaki, M. (1995). Film-copy deliverer problem using genetic
algorithms. Computers and Industrial Engineering, 29(14), 549–553.
Cheng, C.-B., & Mao, C.-P. (2007). A modified ant colony system for solving the
travelling salesman problem with time windows. Mathematical and Computer
Modelling, 46, 1225–1235.
Clerc, M., & Kennedy, J. (2002). The particle swarm-explosion, stability, and
convergence in a multidimensional complex space. IEEE Transactions on
Evolutionary Computation, 6(1), 58–73.
Dantzig, G. B., Fulkerson, D. R., & Johnson, S. M. (1954). Solution of a large-scale
traveling salesman. Operations Research, 2, 393–410.
Dorigo, M., & Gambardlla, L. M. (1997). Ant colonies for the travelling salesman
problem. BioSystems, 43, 73–81.
Eberhart, R. C., & Kennedy, J. (1995). A new optimizer using particle swarm theory.
In Proceedings of the sixth international symposium on micromachine and human
science (pp. 39–43).
Fig. 4. Convergence of the algorithm due to different H For Model-2 for 10 city.
Fiechter, L. (1994). A parallel tabu search algorithm for large traveling salesman
problems. Discrete Applied Mathematics, 51(3), 243–267.
Focacci, F., Lodi, A., & Milano, M. (2002). A hybrid exact algorithm for the TSPTW.
6.3. Iteration number vs profit INFORM Journal on Computing, 14(4), 403–417.
Ghafurian, S., & Javadian, N. (2011). An ant colony algorithm for solving fixed
destination multi-depot multiple traveling salesmen problems. Applied Soft
Here, changes in profit for a ð10  10Þ size crisp TSP (Model-2) Computing, 11, 1256–1262.
are depicted against different iteration numbers in Fig. 4 for differ- Ibanez & Blum, C. (2010). Beam-ACO for the travelling salesman problem with time
ent tour times (H). It is interesting to observe from Fig. 4 that dur- windows. Computers and Operations Research, 37, 1570–1583.
Kennedy, J., & Eberhart, R. C. (1995). Particle swarm optimization. In IEEE intentional
ing initial iterations, profit increases sharply for all H-values and conference on neural networks, Perth, Australia (pp. 1942–1948).
after a certain number of iterations, the rate of increase in profit Leung, K. S., Jin, H. D., & Xu, Z. B. (2004). An expanding self-organizing neural
is quit slow. This particular stage in iteration number increases network for the traveling salesman problem. Neurocomputing, 6, 267–292.
Lin, C. J., Chen, C. H., & Lin, C. T. (2009). A hybrid of cooperative particle swarm
with total tour time (H). In all cases, the rate of profit sharply optimization and cultural algorithm for neural fuzzy networks and its
increases up to the iteration number (3–40). prediction applications. IEEE Transactions on Systems, Man, and Cybernetics,
Part C: Applications, 39(1), 55–68.
Liu, Y. H. (2010). Different initial solution generators in genetic algorithms for
7. Conclusions solving the probabilistic traveling salesman problem. Applied Mathematics and
Computation, 216, 125–137.
Lo, C. C., & Hus, C. C. (1998). Annealing framework with learning memory. IEEE
In this investigation, a new type of TSP, i.e., a real life TSP is Transactions on System, 28, 113.
proposed and solved. Here, TSP Models are formulated as Majumdar, J., & Bhunia, A. K. (2011). Genetic algorithm for asymmetric traveling
salesman problem with imprecise travel time. Computational and Applied
maximization problems. The model is illustrated with numerical
Mathematics, 235, 3063–3078.
examples and some interesting expected results are derived. For Masutti, T. A. S., & Castro, L. N. (2009). A self-organizing neural network using ideas
the first time, a hybrid algorithm – a combination of ACO and from the immune system to solve the traveling salesman problem. Information
PSO has been successfully applied for the above mentioned TSP. Sciences, 179, 1454–1468.
Michalewicz, Z. (1992). Genetic algorithms + data structures = evolution programs.
Here, the model formulation and algorithm are quite general. Berlin: Springer.
These can be applied to formulate and solved solid TSPs and Moon, C., Ki, J., Choi, G., & Seo, Y. (2002). An ancient genetic algorithm for the
3-dimensional assignment problems. The present TSP can also be traveling salesman problem with precedence constraints. European Journal of
Operational Research, 140, 606–617.
extended to Fuzzy, Rough, Fuzzy-rough, Random, etc. environments. Nagata, Y., & Soler, D. (2012). A new genetic algorithm for the asymmetric traveling
TSPs with time windows also can be formulated and solved by the salesman problem. Expert Systems with Applications, 39, 8947–8953.
present algorithm. Petersen, H. L., & Madsen, O. B. G. (2009). The double travelling salesman problem
with multiple stack-formulation and heuristic solution approaches. European
Journal of Operational Research, 198, 339–347.
References Ratnaweera, A., Halgamuge, S., & Watson, H. (2004). Self-organizing hierarchical
particle swarm optimizer with time varying accelerating coefficients. IEEE
Transactions on Evolutionary Computation, 8(3), 240–255.
Applegate, D. L., Bixby, R. E., Chvtal, V., & Cook, W. J. (2007). The traveling salesman Shi, Y., & Eberhart, R. C. (1998). A modified particle swarm optimizer. IEEE
problem. Princeton University Press. International Conference on Evolutionary Computation, 69–73.
Bai, J., Yang, G., Chen, Y., Hu, L., & Pan, C. (2013). A model induced max–min ant Yang, J., Wu, C., Lee, H. P., & Liang, Y. (2008). Solving traveling salesman problems
colony optimization for asymmetric traveling salesman problem. Applied Soft using generalized chromosome genetic algorithm. Natural Science, 18, 887–892.
Computing, 13, 1365–1375.