Applied Sciences: Which Local Search Operator Works Best For The Open-Loop TSP?

applied
sciences
Article
Which Local Search Operator Works Best for the
Open-Loop TSP?
Lahari Sengupta *, Radu Mariescu-Istodor and Pasi Fränti
Machine Learning Group, School of Computing, University of Eastern Finland, 80101 Joensuu, Finland;
radum@cs.uef.fi (R.M.-I.); franti@cs.uef.fi (P.F.)
* Correspondence: lahari@cs.uef.fi; Tel.: +358-417232752

Received: 6 August 2019; Accepted: 17 September 2019; Published: 23 September 2019
Abstract: The traveling salesman problem (TSP) has been widely studied for the classical closed-loop
variant. However, very little attention has been paid to the open-loop variant. Most of the existing
studies also focus merely on presenting the overall optimization results (gap) or focus on processing
time, but do not reveal much about which operators are more efficient to achieve the result. In this
paper, we present two new operators (link swap and 3–permute) and study their efficiency against
existing operators, both analytically and experimentally. Results show that while 2-opt and relocate
contribute equally in the closed-loop case, the situation changes dramatically in the open-loop case
where the new operator, link swap, dominates the search; it contributes by 50% to all improvements,
while 2-opt and relocate have a 25% share each. The results are also generalized to tabu search and
simulated annealing.
Keywords: traveling salesman problem; open-loop TSP; randomized algorithms; local search;
NP–hard; O–Mopsi game
1. Introduction
The traveling salesman problem (TSP) aims to find the shortest tour for a salesperson to visit N
number of cities. In graph theory, a cycle including every vertex of a graph makes a Hamiltonian cycle.
Therefore, in the context of graph theory, the solution to a TSP is defined as the minimum-weight
Hamiltonian cycle in a weighted graph. In this paper, we consider the symmetric Euclidean TSP
variant, where the distance between two cities is calculated by their distance in the Euclidean space.
Both the original problem [1] and its Euclidean variant are NP–hard [2]. Finding the optimum solution
efficiently is therefore not realistic, even for the relatively small size of the input.
O–Mopsi (http://cs.uef.fi/o-mopsi/) is a mobile-based orienteering game where the player needs to
find some real-world targets, with the help of GPS navigation [3]. Unlike a prior instruction of visiting
order in the classical orienteering, O–Mopsi players have to decide the order of visiting targets by
themselves. The game finishes immediately when the last target is reached. The player does not need
to return to the start target. This corresponds to a special case of the orienteering problem [4]. Finding
the optimum order by minimizing the tour length corresponds to solving open-loop TSP [5], because
the players are not required to return to the start position. The open-loop variant is also NP–hard [2].
The solution to a closed-loop TSP is a tour; however, the solution to an open-loop TSP is an open path.
Hence, a TSP with the size N contains N links in its closed-loop solution and N-1 in its open-loop
solution (Figure 1). However, the optimum solution to an open-loop TSP is not usually achievable by
removing the largest link from the optimum solution to the closed-loop TSP (Figure 1).
Appl. Sci. 2019, 9, 3985; doi:10.3390/app9193985 www.mdpi.com/journal/applsci

Appl. Sci.
Appl. Sci. 2019,
2019, 9,
9, 3985
x FOR PEER REVIEW 22 of 26
of 24
Appl. Sci. 2019, 9, x FOR PEER REVIEW 2 of 26
Figure 1. Difference between open- and closed-loop traveling salesman problem (TSP).
Differencebetween
Figure1.1.Difference
Figure betweenopen-
open-and
andclosed-loop
closed-looptraveling
travelingsalesman
salesmanproblem
problem(TSP).
(TSP).
In O–Mopsi, players are not required to solve the optimum tour, but finding a good tour is still
In O–Mopsi,
an essential part of players
the game. are not required to solve the
Experiments optimum tour, but finding a good tour is still
In O–Mopsi, players are not required toshowed solve the that, among tour,
optimum the players
but findingwho completed
a good toura isgame,still
an
only essential
14% had part
foundof the game. Experiments showed that, among the players who completed a game,
an essential part of thethe optimum
game. tour, even
Experiments if thethat,
showed numberamong of targets
the playerswas about 10, on average
who completed a game, [3].
only 14%
The optimum had found
tour isthe the optimum
stilloptimum
needed for tour, even
reference if the number of targets was about 10, on average [3].
only 14% had found tour, even if when analyzing
the number the performance
of targets was about of 10,the
on players.
average This[3].
The optimum
is usually made tour is still
asisa still needed
post-game for reference when analyzing the performance of the players. This is
The optimum tour needed analysis
for reference but can when also happen during
analyzing real-time of
the performance play.
theAplayers.
comparison
This
usuallybe made as a post-game tour analysis but can also happen during real-time play.creator
A comparison can anbe
iscan
usually made
madewith as a thepost-game length
analysis or with
but can thealsovisiting
happen order.
during Thereal-time
game play. A also needs
comparison
made
estimatorwith the
for the tourtourlength or with the visiting order. The game creator also needs an estimator for the
can be made with thelength, as it will
tour length orbewithpublished as a part
the visiting of the
order. Thegamegame info. If the game
creator is created
also needs an
tour length, as on
automatically it will
the be published
demand, TSP asmust
a part be ofsolved
the game info. IfThe
real-time. the game is created
optimality is automatically
highly desired, on
but
estimator for the tour length, as it will be published as a part of the game info. If the game is created
the demand, TSP
can sometimesonbethe must be
compromised solved real-time.
formust
the sake The optimality
of fast processing. is highly desired, but can sometimes be
automatically demand, TSP be solved real-time. The optimality is highly desired, but
compromised
The number for the sake of
of targets inforfast processing.
thethecurrent
can sometimes be compromised sake ofO–Mopsi instances varies from 4 to 27, but larger
fast processing.
The number
instances may of targets
appear. For in the
the current
smaller O–Mopsi
instances, theinstances
relatively varies
simple from 4 to 27, but largeralgorithm
branch-and-bound instances
The number of targets in the current O–Mopsi instances varies from 4 to 27, but larger
may appear.
is fast enough For the smaller
to produce instances,
the smaller
optimum the relatively
tour. However, simple branch-and-bound
for thesimple algorithm
longerbranch-and-bound is fast
instances, it mightalgorithm takeenough
from
instances may appear. For the instances, the relatively
to produce
minutes to the optimum
hours. Along tour.
with However,
the exact for the
solver, longer
a instances,
faster heuristic it might take
algorithm from
is minutes
therefore to
neededhours.to
is fast enough to produce the optimum tour. However, for the longer instances, it might take from
Along
find with
the the exact
solution more solver, a faster
quickly, in heuristic
spite of algorithm
the danger isoftherefore
occasionally needed to find the
resulting in solution
a more
sub-optimal
minutes to hours. Along with the exact solver, a faster heuristic algorithm is therefore needed to
quickly,
solution. inBesides
spite of the danger TSP
O–Mopsi, of occasionally resulting in a sub-optimal solution. Besides O–Mopsi,
find the solution more quickly, in arises spite ofinthe route
danger planning [6], orienteering
of occasionally resultingproblems [7], and
in a sub-optimal
TSP arises in
automated maproute planning[8].
generation [6], orienteering problems [7], and automated map generation [8].
solution. Besides O–Mopsi, TSP arises in route planning [6], orienteering problems [7], and
An
An example
example of
of O–Mopsi
O–Mopsi play is shown in Figure 2, with a comparison to the optimum tour
automated map generation [8]. play is shown in Figure 2, with a comparison to the optimum tour
(reference).
(reference). For analyzing,
For analyzing, we
weplay calculate several
calculate several reference
reference tours: optimum, fixed-start,
tours: optimum, fixed-start, and dynamic.
and dynamic.
An example of O–Mopsi is shown in Figure 2, with a comparison to the optimum tour
The
The optimum
optimum tour takes into account the fact
fact that
that the
the players can freely
freely choose
choose their their starting point.
point.
(reference). Fortour takes into
analyzing, weaccount
calculate the several referenceplayers
tours: can optimum, fixed-start, starting
and dynamic.
Since
Since selecting
selectingtour the best
best starting
the takes starting point
pointthe is
is challenging
challenging [9], we
we also calculate two two alternative reference
The optimum into account fact that the[9], players also
cancalculate
freely choose alternative
their starting reference
point.
tours
tours from
from the
the location
location where
where the
the player
player started.
started. Fixed-start
Fixed-start is
is the
the optimum
optimum tour
tour using
using this
this starting
starting
Since selecting the best starting point is challenging [9], we also calculate two alternative reference
point.
point.fromThe dynamic
The the
dynamic tour
tour follows
follows thetheplayer’s
player’s choices
choices from
from thethestarting point but re-calculates the new
tours location where the player started. Fixed-start is thestarting
optimum point tourbut re-calculates
using this starting the
optimum
new optimum every time
every the player
time deviates
the player from
deviates optimumfromfrompath.
optimum Three different
path. Three evaluation
different measures
evaluation are
point. The dynamic tour follows the player’s choices the starting point but re-calculates the
obtained
measures using
are these three
obtained usingtours:
these gap, mismatches,
three tours: and
gap, mistakes; see
mismatches, and [10].
mistakes; see [10].
new optimum every time the player deviates from optimum path. Three different evaluation
measures are obtained using these three tours: gap, mismatches, and mistakes; see [10].
2. Examples of O–Mopsi playing (left), the final tour (middle)

Figure 2. (middle) and comparison
comparison to optimum
optimum
tour (right).
Figure 2. Examples of O–Mopsi playing (left), the final tour (middle) and comparison to optimum
tour (right).
Appl. Sci. 2019, 9, 3985 3 of 24
Being NP–hard, exact solvers of TSPs are time-consuming. Despite the huge time consumption,
exact solvers were prime attractions to researchers in earlier days. Dantzig, Fulkerson, and Johnson [11],
Held and Karp [12], Padberg and Rinaldi [13], Grötschel and Holland [14], Applegate et al. [15],
and others developed different algorithms to produce exact solutions. Laporte [16] surveyed exact
algorithms for TSP. By the year 2006, the Concorde solver solved a TSP instance with 85900 cities [17].
Along with these, different heuristic algorithms have been developed to find very fast near-optimal
solutions. The local search is one of the most common heuristics to produce near-optimum results [18].
It has two parts: tour construction and tour improvement. The tour construction generates an initial
solution that can be far beyond the optimum. A significant amount of research developed different
tour construction algorithms, such as the savings algorithm by Reference [19], polynomial–time 3/2
approximation algorithm by Reference [20], insertion [21], greedy [21], and the nearest neighbor
algorithm [21]. Among these, we use the greedy algorithm and a random arrangement for the
tour construction.
Being an iterative process, the tour improvement step modifies the initial solution with a small
improvement in each iteration. The key component in the local search is the choice of the operator,
which defines how the current solution is modified to generate new candidate solutions. Most used
operators are 2-opt [22] and its generalized variant, called k–opt [23], which is empirically the most
effective local search algorithm for large TSPs [24]. Okano et al. [25] analyzed the combination of 2-opt
with different well-known construction heuristics. Several researchers studied local search for TSP and
different optimization problems [16,21,26–32], even for vehicle routing application [33,34].
Although existing studies have analyzed the local search algorithms extensively, it is still not
known which operators are effective for open-loop cases. Besides this, the existing literature mainly
presents optimization results of algorithms instead of providing a detailed study on different operators.
As operators are the backbone of the local search algorithms, we need a detailed study on them to
know how to increase the productivity of the algorithm and how to avoid the local minima.
In this paper, we study several operators and their combinations in the context of local search.
We aim at answering which of them works best in terms of the quality and efficiency. We consider four
operations, of which two are existing and two are new:
• Relocate [35]
• 2-opt [22]
• 3–permutation (new)
• Link swap (new)
Between the new two operators, the former also applies to the more common closed-loop TSP,
but the second is specifically tailored for the open-loop problem. We also propose an algorithm to
achieve a global minimum in most problem instances of size up to 31, preferably in real-time.
We study the performance of these operators when applied separately, and when mixed.
We consider random, best improvement, and first improvement search strategies. We first report how
successful the operators are in equal conditions, and whether the choice of the best operator changes
towards the end of the iterations. We study two initialization techniques: random and heuristic,
and the effect of re-starting the search. Results are reported for two open-loop datasets (O–Mopsi,
Dots), see Table 1. The O–Mopsi dataset contains real-world TSP instances and the Dots dataset is
a computer-generated random dataset for an experimental computer game. Since it is an outdoor
dataset, generating the former one is relatively difficult. Therefore, the available number of instances
are much lower than Dots. However, both datasets need real-time reference solutions.
The rest of the paper is organized as follows. In Section 2, we describe the local search operators
we use in this study. We define these operators with illustrations and study their abilities to produce
unique improved solutions. In Section 3, we provide results of the tests carried on with these operators
on Table 1 datasets. Based on these results, we introduce a new method to solve open-loop TSPs in this
section. In Section 4, we evaluate the quality and efficiency of our proposed algorithm. In Section 5,
Appl. Sci. 2019, 9, 3985 4 of 24
we test stochastic variants, aiming to improve our algorithm. Lastly, in Section 6, we conclude our
findings in this study.
Table 1. Datasets used in this study.
Dataset: Type: Distance: Instances: Sizes:

O-Mopsi 1 Open loop Haversine 145 4-27
Dots 1 Open loop Euclidean 4226 4-31
1 http://cs.uef.fi/o-mopsi/datasets/.
2. Local Search
The local search is a way to upgrade an existing solution to a new candidate solution by small
changes. A local search operator seeks and finds a small modification to improve the solution.
The search approach of a local search operator can be one of several of different types. In the first and
best improvement strategy, the first and best candidate is always chosen among all the candidates,
respectively. For these two cases, we continue the search until saturation, which means no further
improvement is found. In a random search strategy, candidates are chosen randomly until some
constant iteration. In the local search, a better solution can almost always be found.
We next study the following four operators:
• Relocate [35]
• 2-opt [22]
• 3–permute (new) Link swap (new)
2.1. Relocate
Changing the order of nodes can produce different solutions. Relocate is the process where a
selected node (target) is moved from its current position in the tour to another position (destination).
Hence, the position of the selected node is relocated. Each relocation of a node produces one outcome.
Gendreau, Hertz, and Laporte [35] developed the GENI method to insert a new node by breaking a
link. We use this idea and create relocate as a tour improvement operator. Figure 3 illustrates relocate
using an O-Mopsi example, where target A is moved between B and C. First, we take A from its current
position by removing links PA and AQ, and by adding link PQ to close the gap. Then, A is added to its
new position by removing link BC, and by adding BA and AC. As an effect, the tour becomes 30 m
shorter. The size of the search neighborhood is O(N2 ) because there are N possible targets and N-2
destinations to choose
Appl. Sci. 2019, 9, x FORfrom, in total.
PEER REVIEW 5 of 26
Figure 3. Working of the relocate operator and an example of how it improves the tour in an O–Mopsi
Figure 3. Working of the relocate operator and an example of how it improves the tour in an
gameO–Mopsi
by 30 meters.
game by 30 meters.
2.2. Two-Optimization (2-opt)

Croes [22] first introduced the 2-optimization method, which is a simple and very common
operator [36]. The idea of 2-opt is to exchange the links between two pairs of subsequent nodes as
shown in Figure 4. It has also been generalized to k–opt (L–K heuristics) [23] and implemented by
Figure
Appl. Sci. 2019,3. Working of the relocate operator and an example of how it improves the tour in an5 of 24
9, 3985
O–Mopsi game by 30 meters.
2.2. Two-Optimization
2.2. Two-Optimization (2-opt)
(2-opt)
Croes [22] first introduced the 2-optimization method, which which is a simple and very common
operator
operator [36]. The idea of 2-opt
The idea of 2-opt is is to exchange
exchange the the links between two pairs of subsequent nodes as
between
shown in Figure 4. It It has
has also
also been
been generalized
generalized to to k–opt (L–K heuristics)
heuristics) [23] and implemented by
Helsgaun [37].
[37]. Johnson
Johnson and
andMcGeoch
McGeoch[21] [21]explained
explainedthoroughly
thoroughlythethe
2-opt andand
2-opt 3–opt
3–optprocesses. In the
processes. In
case of 3–opt, the only difference from 2-opt is to exchange the links between three
the case of 3–opt, the only difference from 2-opt is to exchange the links between three pairs of pairs of nodes.
The method
nodes. The was originally
method developed developed
was originally for the closed-loop
for the TSP, but it applies
closed-loop TSP, tobut
the itopen-loop
applies TSP as
to the
well.
open-loop TSP as well. In addition, we also consider dummy nodes paired with one of the twoa
In addition, we also consider dummy nodes paired with one of the two terminal nodes. Such
way, terminal
terminal nodes.nodes
Suchcan alsoterminal
a way, change into an can
nodes internal
also node.
change into an internal node.
Figure
Figure 4.
4. Three
Three examples
examples of
of how
how the
the 2-optimization
2-optimization method
method can
can improve the tour.
The 2-opt
2-optapproach
approachremoves crossings
removes in the path.
crossings in the Panpath.
and Xia
Pan[38]and
highlighted
Xia [38]this cross-removing
highlighted this
effect in their work.
cross-removing However,
effect in their 2-opt
work.is However,
more than 2-opt
the cross-removal
is more thanmethod, as shown in method,
the cross-removal Figure 4. as
shownIn inFigure
Figure5, 2-opt
4. is illustrated with two real examples from O-Mopsi. Here, AB and CD are the
two links
In Figureinvolved in is
5, 2-opt theillustrated
operation. In two
with the modified tour, from
real examples theseO-Mopsi.
links are Here,
replaced
AB byandAC
CDand
are BD.
the
In the first example, the original links cross, which is then deleted and the length of the
two links involved in the operation. In the modified tour, these links are replaced by AC and BD. Inoverall tour is
shortened from 1010 m to 710 m. The second example shows that 2-opt can also improve the tour,
even when the original links do not cross.
The 2-opt operator works with links. A tour consists of N-1 links, so there are (N-1)*(N-2) possible
pairs of links. Thus, the size of the neighborhood is O(N2 ).
the first example, the original links cross, which is then deleted and the length of the overall tour is
shortened from 1010 m to 710 m. The second example shows that 2-opt can also improve the tour,
even when the original links do not cross.
Appl. Sci.
The2019, 9, 3985
2-opt 6 of 24
operator works with links. A tour consists of N-1 links, so there are (N-1)*(N-2)
possible pairs of links. Thus, the size of the neighborhood is O(N2).
Figure 5. Two examples of the 2-opt method in O-Mopsi.
2.3. Three-Node Permutation

2.3. Three-Node Permutation (3-permute)
(3-permute)
Permutation
Permutation betweenbetween the the order
order of
of the
the nodes
nodes is is another
another potentially
potentially useful
useful operator
operator forfor the
the local
local
search.
search. Although
Although the the idea
ideaisisquite
quitestraightforward,
straightforward,we weare not
are aware
not awareof of
its its
existence
existencein literature, so we
in literature, so
will briefly
we will consider
briefly it here.
consider As per
it here. Asthe
pertheory, three three
the theory, nodesnodes
(A, B, (A,
C) can
B, C)generate six different
can generate orders.
six different
Therefore, given angiven
orders. Therefore, existing sequencesequence
an existing (triple) of(triple)
three nodes,
of threeABC, we ABC,
nodes, can create
we canfivecreate
new solutions:
five new
ACB, BAC, BCA, CAB, and
solutions: ACB, BAC, BCA, CAB, and CBA. CBA.
In
In Figure
Figure 6, 6, an
an original
original andand its
its five
five alternatives
alternatives areare shown. The real
shown. The real O-Mopsi
O-Mopsi game game where
where this
this
example originates from is also shown. Here, the tour length is reduced
example originates from is also shown. Here, the tour length is reduced by 170 m when changing by 170 m when changing the
original order (ABC) to the better one
the original order (ABC) to the better one (BAC). (BAC).
There
There areare N-1
N-1 possible
possible triples,
triples, including
including twotwo dummies
dummies after
after the
the terminal
terminal points. The size
points. The size of
of the
the
neighborhood
neighborhood is, is,therefore, 6*(N-1)==O(N),
therefore,6*(N-1) O(N),which
which is is
significantly smaller
significantly smaller than thatthat
than of the
of relocate and
the relocate
Appl.
2-opt Sci. 2019, 9, x FOR PEER REVIEW
operators. 7 of 26
and 2-opt operators.
Figure
Figure6.
6. All
Allsix
six combinations 3–permute and
combinations of 3–permute andan
anO–Mopsi
O–Mopsigame
gameexample
examplethat
that
is is improved
improved byby
170170
m.
m.
2.4. Link Swap

Solving the open-loop variant of TSP allows us to introduce a new operator that is not possible
with the closed-loop problem. Link swap is analogous to relocate, hence we consider one link, AB,
Figure 6. All six combinations of 3–permute and an O–Mopsi game example that is improved by 170
Appl. Sci.
m. 2019, 9, 3985 7 of 24
2.4. Link Swap

2.4. Link Swap
and swap it into a new location in the tour, see Figure 7. However, to keep the new solution as a
and swap it into a new location in the tour, see Figure 7. However, to keep the new solution as a valid
valid TSP tour, we are limited to three choices: replace AB by AT2, BT1, or T1T2. As a result, one or
TSP tour, we are limited to three choices: replace AB by AT2 , BT1 , or T1 T2 . As a result, one or both of
both of the nodes A and B will become a new terminal node.
the nodes A and B will become a new terminal node.
In the O–Mopsi example in Figure 7, we can see that a single link swap operation improves the
In the O–Mopsi example in Figure 7, we can see that a single link swap operation improves the
solution by reducing the tour length up to 137 m.
solution by reducing the tour length up to 137 m.
We have N–1 choices for a link to be removed, and for each of them, we have three new
We have N–1 choices for a link to be removed, and for each of them, we have three new alternatives.
alternatives. Thus, the size of the neighborhood is 3*(N–1) = O(N). Link swap is a special case of
Thus, the size of the neighborhood is 3*(N–1) = O(N). Link swap is a special case of 3–opt and relocate
3–opt and relocate operator, but as the size of the neighborhood is linear, it is a faster operation
operator, but as the size of the neighborhood is linear, it is a faster operation than both 3–opt and
than both 3–opt and relocate operator.
relocate operator.
Figure 7.
Figure Original tour
7. Original tour and
and three
three alternatives
alternatives created
created by
by link
link swap
swap (left). In each
(left). In each case,
case, the
the link
link xx is
is
replaced by a new link connecting to one or both terminal nodes. In the O–Mopsi example
replaced by a new link connecting to one or both terminal nodes. In the O–Mopsi example (right), (right),
the tour
the tour is
is improved
improved byby 137m.
137m.
2.5. The Uniqueness of the Operators
2.5. The Uniqueness of the Operators
All operators create a large number of neighbor solutions. The set of solutions generated by
one operator can be similar to the set of another operator. Given the same initial solution, if all the
candidate solutions generated by one operator were a subset of the solutions generated by another
operator, then this approach would be redundant. Next, we study the uniqueness of these methods.
Relocate: The highlighted target in the start tour in Figure 8 can swap to six different positions.
By analyzing these six results, we can observe that the tours 1 to 3 are also possible outcomes of
3–permute. Tours 1 and 2 are also possible outcomes of 2-opt and tour 1 can be an outcome of link
swap. However, the rest of the tours are unique, and only possible to produce by the relocate operator.
Only the relocate operator can produce the best result in this example (tour 4).
Byoperator,
analyzing these six results, we can observe that the tours 1 to 3 are also possible outcomes of
then this approach would be redundant. Next, we study the uniqueness of these methods.
3–permute. Tours
Relocate: The 1 and 2 are also
highlighted possible
target in the outcomes ofFigure
start tour in 2-opt 8and
cantour
swap1 to
cansixbedifferent
an outcome of link
positions.
swap. However,
By analyzing the six
these restresults,
of thewe tours
can are unique,
observe that and only 1possible
the tours to 3 are to
alsoproduce
possibleby the relocate
outcomes of
operator. Only the relocate operator can produce the best result in this example (tour
3–permute. Tours 1 and 2 are also possible outcomes of 2-opt and tour 1 can be an outcome of link 4).
swap.
Appl. Sci.However,
2019, 9, 3985 the
rest of the tours are unique, and only possible to produce by the relocate
8 of 24
operator. Only the relocate operator can produce the best result in this example (tour 4).
Figure 8. Six neighboring solutions created by relocate, of which three are unique.
2-opt: In thisFigure 8. Six neighboring

method, two linkssolutions created by
are removed andrelocate,
two newof which
onesthree
three are
areare unique. An example is
unique.
created.
shown2-opt:
in Figure 9, where the first removed linkand
is fixed as the red one, which is highlighted, and
2-opt: InIn this
thismethod,
method,two twolinks areare
links removed
removed two two
and new new
ones ones
are created. An example
are created. is shown
An example is
theinsecond
Figure one is varied. The resulting six new tours are shown. The other operators can generate
shown in9,Figure
where9,the first removed
where link is fixed
the first removed linkasisthe red as
fixed one,
thewhich is highlighted,
red one, and the second
which is highlighted, and
four
one ofis them asThe
varied. well; however, two tours
of them are unique. The best outcome is uniquely generated
them as by
the second one is resulting
varied. The six new
resulting arenew
six shown.
tours Theareother operators
shown. can generate
The other operatorsfour
canofgenerate
the 2-opt
well;
four of
operator.
however,
them astwo of them
well; are unique.
however, The best
two of them are outcome
unique. The is uniquely generated
best outcome by the 2-opt
is uniquely operator.
generated by
the 2-opt operator.
Figure 9. Six neighboring solutions created by 2-opt, of which two are unique.
3-permute: The example in Figure 10 reveals that the operator does not produce even a single
3-permute:
unique The example
result. Therefore, in relocate
if both Figure 10 andreveals thatused,
2-opt are the operator doesis not
this operator produceRelocate
redundant. even a cansingle
unique 3-permute:
result. The example
Therefore, if in Figure
both relocate10 reveals
and 2-optthat
aretheused,
operator
this does not
operator produce
is even a Relocate
redundant. single
find
uniquefourresult.
out ofTherefore,
the five solutions, and 2-optand
if both relocate can2-opt
find three.
are used, this operator is redundant. Relocate
can find A four out
potential of the
benefit five
of thesolutions,
3–permute and 2-opt
operator can
is find
that three.
it needs
can find four out of the five solutions, and 2-opt can find three.to explore only a smaller neighborhood,
A potential
O(N),Ain potential
comparison benefit of2 the
to O(Nof) of 3–permute
relocate operator
and 2-opt. However, is that it needs
this it
smaller to exploreisonly a smaller
benefit the 3–permute operator is that needsneighborhood
to explore onlynotaextensive
smaller
neighborhood,
enough, O(N),theinsolutions
as it misses comparisonshownto O(N2 ) 8ofand
2
in Figures relocate
9. Our andand 2-opt. also
experiments However, this smaller
show that
neighborhood, O(N), in comparison to O(N ) of relocate 2-opt. However, this the local
smaller
neighborhood
search is not
using 3-permute extensive
is consistently outperformed by the other operators. We, therefore, do9.
enough, as it misses the solutions shown in Figures 8 and Our
not
neighborhood
Appl. is not
Sci. 2019, 9, x FOR PEERextensive
REVIEW enough, as it misses the solutions shown in Figures 8 and 9. Our 9 of 26
experiments
consider thisalso showfurther.
operator that the local search using 3-permute is consistently outperformed by the
experiments also show that the local search using 3-permute is consistently outperformed by the
other
other operators. We, therefore,do
operators. We, therefore, donot
not consider this operator
consider this operatorfurther.
further.
Figure 10. Six neighboring solutions created by 3–permute, but none of them is unique.
Link swap: This operator creates three alternatives for a selected link. Two of these are redundant
with 2-opt,
Link but the
swap: oneoperator
This which connects
createsthe two alternatives
three previous terminal
for apoints is unique
selected (Figure
link. Two of11). It isare
these
redundant with 2-opt, but the one which connects the two previous terminal points is unique
(Figure 11). It is also the best one. Link swap is therefore worthwhile in cases when the terminal
points are close to each other.
Link
Appl. swap:
Sci. 2019, This
9, 3985operator creates three alternatives for a selected link. Two of these 9 of 24are
redundant with 2-opt, but the one which connects the two previous terminal points is unique
(Figure 11). It is also the best one. Link swap is therefore worthwhile in cases when the terminal
also the best one. Link swap is therefore worthwhile in cases when the terminal points are close to
points are close to each other.
each other.
Figure11.
Figure Three neighboring
11.Three neighboringsolutions
solutionscreated by link
created by swap, of which
link swap, of the one connecting
which the terminalthe
the one connecting
nodes is unique.
terminal nodes is unique.
3. Performance of a Single Operator
3. Performance of a Single Operator
We next test the performance of the different operators within a local search. For this, we need
We next
an initial testtothe
tour performance
start with and aofsearch
the different
strategy.operators within
These will a local search.
be discussed For this, by
first, followed we the
need
anexperimental
initial tour to start with and a search strategy. These will be discussed first, followed by the
results.
experimental results.
3.1. Initial Solution
3.1. Initial Solution
If the operator and search strategy are good, the choice of initialization should not matter much
forIfthe final
the result. We,
operator andtherefore, considerare
search strategy onlygood,
two choices:
the choice of initialization should not matter
much
• for the final result. We, therefore, consider only two choices:
Random
• • Random
Heuristic (greedy)
• Heuristic (greedy)
Random
Random initializationselects
initialization selectsthe
the nodes
nodes one
one by
by one
one randomly
randomlytotocreate
createthe
theinitial tour.
initial tour.
A straightforward method for better initialization is greedy heuristic. It selects
A straightforward method for better initialization is greedy heuristic. It selects a random a random node
node
as a starting point and always takes the closest unvisited node as the next target. Because
as a starting point and always takes the closest unvisited node as the next target. Because of solving of solving
open-loop TSP, the starting point also matters in this case. We therefore systematically try all nodes as
open-loop TSP, the starting point also matters in this case. We therefore systematically try all nodes
the starting point and generate N candidate tours using the greedy heuristics. Among N candidate
as the starting point and generate N candidate tours using the greedy heuristics. Among N
tours, we select the shortest tour as the initial solution. Algorithm 1 presents the pseudocode for this.
candidate tours, we select the shortest tour as the initial solution. The pseudocode for this is the
following:
Algorithm 1: Finding the heuristic initial path
HeuristicPath () Algorithm 1: Finding the heuristic initial path
InitialPath ← infinity
HeuristicPath ()
FOR node ← 1 TO N DO
InitialPath  infinity
NewPath ← GreedyPath ( node )
FOR) <node
IF Length ( NewPath ( 1InitialPath
Length TO N DO ) THEN
NewPath  GreedyPath ( node )
InitialPath ← NewPath
RETURN InitialPath IF Length ( NewPath ) < Length ( InitialPath ) THEN
GreedyPath ( startNode )
NewPath [1] ← startNode
FOR i ← 2 TO N DO
NewPath [i] ← Nearest node from startNode
startNode ← NewPath [i]
RETURN NewPath
3.2. Search Strategy

For the search strategy we consider the following three choices:
• Best improvement
• First improvement
• Random search
Appl. Sci. 2019, 9, 3985 10 of 24
Best improvement studies the entire neighborhood and selects the best one among all solutions.
This can be impractical, especially if the neighborhood is large. The first improvement studies the
neighborhood only until it founds any solution that provides improvement. It can be more practical as
it moves on with the search quicker. Random search studies the neighbors in random order, and similar
to the first improvements, accepts any new solution that improves.
We also consider repeated (multi-start) local search, which simply re-starts the search from scratch
several times. This helps if the search is susceptible to getting stuck into a local minimum [39].
Moreover, every repeat is essentially a random attempt to converge to a local optimum; parallelization
can be applied by running different repeats on different execution threads. O'Neil & Burtscher [40]
and Al-Adwan et al. [41] preferred repeated local search, for TSP and Xiang et al. [42] for matrix
multiplication to overcome local optima.
A pseudocode that covers all the search variants with all the operators is given in Algorithm 2.
The only parameter is the number of iterations for which the search continues. First and best
improvement searches can also be terminated when the entire neighborhood is searched without
further improvement. This corresponds to the hill-climbing strategy.
Algorithm 2: Finding the improved tour by all the operators for different search strategies
LocalSearch ( InitialPath, OperationType )
Path ← InitialPath
SWITCH SearchType
CASE RANDOM
FOR i ← 1 TO Iterations DO
NewPath ← NewRandomSolution ( Path, OperationType )
IF Length ( NewPath ) < Length ( Path ) THEN
Path ← NewPath
CASE FIRST
Path ← NewFirstSolution ( Path )
CASE BEST
Path ← NewBestSolution ( Path )
RETURN Path
NewRandomSolution ( Path, OperationType )
SWITCH OperationType
CASE RELOCATE
Target ← Random ( 1, N )
Destination ← Random ( 1, N )
NewPath ← relocate ( Path, Target, Destination )
CASE 2OPT
FirstLink ← Random ( 1, N )
SecondLink ← Random ( 1, N )
NewPath ← 2Opt ( Path, FirstLink, SecondLink )
CASE LINKSWAP
Link ← Random ( 1, N )
NewPath ← linkSwap ( Path, Link )
RETURN NewPath
NewFirstSolution ( Path, OperationType )
FOR i ← 1 TO N DO
FOR j ← 1 TO N DO
CASE RELOCATE
NewPath ← relocate ( Path, i, j )
CASE 2OPT
NewPath ← 2Opt ( Path, i, j )
Appl. Sci. 2019, 9, 3985 11 of 24

Path ← NewPath
RETURN Path
IF OperationType = LINKSWAP
NewPath ← linkSwap ( Path, i )
Path ← NewPath
RETURN Path
NewBestSolution ( Path, OperationType )

FOR i ← 1 TO N DO
FOR j ← 1 TO N DO
CASE RELOCATE
NewPath ← relocate ( Path, i, j )
CASE 2OPT
NewPath ← 2Opt ( Path, i, j )
Path ← NewPath
IF OperationType = LINKSWAP
NewPath ← linkSwap ( Path, i )
Path ← NewPath
RETURN Path
3.3. Results With A Single Operator

We first study how close the operators can improve paths to the optimum solutions using O–Mopsi
and Dots datasets. Optimum paths were computed by branch-and-bound. Two results are reported:
• Gap (%)
• Instances solved (%)
The gap is the length of the optimum path divided by the length of the tour found by the local
search (0% indicates optimum). The second result is the number of instances for which an optimum
solution is found by the local search. We also express it in a percentage value, which indicates 100%
means that the method finds the optimum solutions for all instances. We execute every operator
individually with all search strategies on all O-Mopsi game instances. Repeated search is applied only
for the random search with random initialization.
We combine the results of O-Mopsi and Dots datasets and report those in Table 2. In the table,
we first write the gap values, and the percentage values of solved instances are written inside the
parentheses. The 2-opt is the best individual operator. It reaches a 2% gap and solves 57% of all
instances when starts from a random initialization, and 0.8% (73% instances) using the heuristic
initialization. The repeats are important. With 25 repeats, the 2-opt solves 97% of all instances, and the
mean gap is almost 0%.
The better initialization provides significant improvement in all cases, which indicates the search
gets stuck into a local minimum. The choice of the search strategy, however, has only minor influence.
We, therefore, consider only the random search in further experiments.
Appl. Sci. 2019, 9, 3985 12 of 24
Table 2. Combined results of single and subsequent application of random improvement with different
operators averaged over O–Mopsi and Dots instances. The percentage of solved instances are shown
inside the parenthesis. The number of iterations was fixed to 10000 for randomized search, whereas
first and best improvements were iterated until convergence.
Single Operator Double Operators Mix of Three

Repeated Repeated (25 times) Second
Initialization Random Heuristic Repeated (25 times)
(25 times) Operator:
Link
Relocate 2-opt Subseq-uent Mixed
swap
First 7% (30%) 1.0% (70%) - - - - - -
Best 6% (34%) 0.9% (72%) - - - - - -
Reloc-ate
0.18% 0.007% 0.032% 0.010% 0.005%
Random 6% (34%) 0.9% (71%) -
(90%) (98%) (97%) (98%) (99.7%)
First 2% (54%) 1.0% (70%) - - - - - -
2-opt Best 2% (57%) 0.8% (73%) - - - - - -
0.02% 0.009% 0.018% 0.005% 0.005%
Random 2% (52%) 0.8% (73%) -
(97%) (98%) (97%) (99%) (99.7%)
Link Best 44% (5%) 2.9 (59%) - - - - - -
swap 1.00% 0.019% 0.010% 0.006% 0.005%
Random 11% (31%) 3.0 (58%) -
(77%) (98%) (97%) (99%) (99.7%)
3.4. Combining the Operators

Our primary goal was to find an optimum solution for all game instances. The best combination
(2-opt with a random search using 25 repeats) missed it for only 3% of instances. Nevertheless, these are
relatively easy instances; therefore, they should all be solvable by the local search. We, therefore,
consider combining different operators. We do this by applying a single run of the local search with one
operator, and then continue the search with another operator. We consider all pairs and the subsequent
combinations of all three operators.
We tested all combinations but report only the results of 25 repeats using the random improvement
(random initialization). From the results in Table 2, we can see that all combinations manage to solve a
minimum of 97% of all instances and have a gap of 0.032% or less. The mixed variant with the order
(2-opt + relocate + link swap) found the optimum solution for 99% cases. In this case, 2-opt solves
already a 97% of all instances, then the consecutive relocate makes it 98%, and finally link swap makes
it 99%. This shows that the operators have complementary search spaces; when one operator stops
progressing, another one can still improve further. Examples are shown in Figure 12.
Figure12.
Figure Operatorshave
12.Operators havecomplementary
complementarysearch
searchspaces.
spaces.
After being stopped, it is also possible that a single operator can start to work again when
other operators make some improvement in the meantime. Hence, these works of other operators
make the first operator alive again. Figure 13 has three such examples, where the first operator
entraps into a local optimum. Then, other operators recover the process from the local optimum
Appl. Sci. 2019, 9, 3985 13 of 24
Figure 12. Operators have complementary search spaces.
After being stopped, it is also possible that a single operator can start to work again when other
After being stopped, it is also possible that a single operator can start to work again when
operators make some improvement in the meantime. Hence, these works of other operators make the
other operators make some improvement in the meantime. Hence, these works of other operators
first operator alive again. Figure 13 has three such examples, where the first operator entraps into a
make the first operator alive again. Figure 13 has three such examples, where the first operator
local optimum. Then, other operators recover the process from the local optimum and improvement
entraps into a local optimum. Then, other operators recover the process from the local optimum
continues. Again, the first operator starts working later.
and improvement continues. Again, the first operator starts working later.
Figure 13. A single operator can re-work after being stuck with the help of other operators’ improvement.
Figure 13. A single operator can re-work after being stuck with the help of other operators’
improvement.
In addition, we consider also mixing the operators randomly as follows. We perform only a
single run of the local search for 10,000 iterations. However, instead of fixing the operator beforehand,
we choose it randomly at every iteration. The pseudocode of the mixed variant is demonstrated in
Algorithm 3. This random mixing solves all instances with having a 0% gap if we repeat the process
for 25 times.
Algorithm 3: Finding a solution by mixing local search operators

RandomMixing ( NumberOfIteration, Path )
FOR i ← 1 TO NumberOfIteration DO
operation ← Random ( NodeSwap, 2opt, LinkSwap )
NewPath ← NewRandomSolution ( Path, operation )
Path ← NewPath
RETURN Path
Lawler et al. [43] stated that local search algorithms get trapped in sub-optimal points. Here, we find
that the operators, even when mixed, could not solve all the instances from any random initial solution.
Figure 14 shows three examples where the search was stuck to a local optimum from where none of
Path  NewPath
RETURN Path
Lawler et al. [43] stated that local search algorithms get trapped in sub-optimal points. Here,
Appl. Sci. 2019, 9, 3985 14 of 24
we find that the operators, even when mixed, could not solve all the instances from any random
initial solution. Figure 14 shows three examples where the search was stuck to a local optimum
the operators
from where none can of
process further. In
the operators these
can cases,further.
process the repeated random
In these cases,mixed local search
the repeated was mixed
random needed
to solve these instances.
local search was needed to solve these instances.
Figure
Figure14.
14.Examples
Examplesofoflocal
localoptima.
optima.Repeats
Repeatsovercome
overcomethese.
these.
4.4.Analysis
AnalysisofofRepeated
RepeatedRandom
RandomMixed
MixedLocal
LocalSearch
Search
Wenext
We nextanalyze
analyzethe
theperformance
performanceofofthe therepeated
repeatedlocal
localsearch
searchininmore
moredetail.
detail.We
Weuse
useboth
boththe
the
145 O–Mopsi instances and the larger Dots dataset, which contains 4226 computer-generated
145 O–Mopsi instances and the larger Dots dataset, which contains 4226 computer-generated TSP TSP
instances.Here
instances. Hereproblem
problemsize
sizevaries
variesfrom
from44toto31.
31.
4.1.Effect
4.1. Effectofofthe
therepetitions
repetitions
Figure15
Figure 15 reports
reports how
how many
manyofofthetheinstances areare
instances solved when
solved we increase
when the number
we increase of repeats.
the number of
In the case of O–Mopsi, most instances (144) are solved already with 15 repeats, and
repeats. In the case of O–Mopsi, most instances (144) are solved already with 15 repeats, and all all with 25 repeats.
The results
with 25 repeats.with The
the Dots instances
results are Dots
with the similar: 15–25 iterations
instances are reasonably
are similar: good, are
15–25 iterations afterreasonably
which only
Appl.instances
100 Sci. 2019, 9,were
x FORnot
PEER REVIEW
solved optimally. After 150 repeats, only one instance remains unsolved.15 of 26
good, after which only 100 instances were not solved optimally. After 150 repeats, only one instance
remains unsolved.
Figure 15. Instances solved as a function of the number of repetitions.

Figure 16 demonstrates the effect of repetition for a single (typical) instance as the average path
length. We portray all improving steps from a random start in the first run. There is a similar kind of
length. We portray all improving steps from a random start in the first run. There is a similar kind
improving steps from different random starts in the consecutive runs. However, we only show more
of improving steps from different random starts in the consecutive runs. However, we only show
improved steps than the previous run. Thus, we can see there are very few better steps in higher runs,
more improved steps than the previous run. Thus, we can see there are very few better steps in
and three or four repeated runs are usually enough to achieve the optimum result. The gap values
higher runs, and three or four repeated runs are usually enough to achieve the optimum result. The
gap values are also shown to demonstrate the closeness of our outcomes to the optimum result. In
the first run, we usually achieve a result that is already very close to the optimum. The repetitions
push the result to the global optimum, but they act merely as a fine tuner for the results. If we can
settle the sub-optimal result very close to the global optimum, then a local search with only a few
length. We portray all improving steps from a random start in the first run. There is a similar kind
of improving steps from different random starts in the consecutive runs. However, we only show
Appl. Sci. 2019, 9, 3985 15 of 24
more improved steps than the previous run. Thus, we can see there are very few better steps in
higher runs, and three or four repeated runs are usually enough to achieve the optimum result. The
gapalso
are values are to
shown also shown to demonstrate
demonstrate the closenessthe closeness
of our outcomesof our outcomes
to the optimum to the optimum
result. In the result. In
first run,
theusually
we first run, we usually
achieve a resultachieve a resultvery
that is already that close
is already
to the very close The
optimum. to the optimum.
repetitions Thethe
push repetitions
result to
push
the the result
global optimum,to thebutglobal optimum,
they act merely asbuta they act merely
fine tuner for theas a fine Iftuner
results. for settle
we can the results. If we can
the sub-optimal
settle very
result the sub-optimal result optimum,
close to the global very closethen
to the global
a local optimum,
search then
with only a local
a few search
repeats withbeonly
would a few
sufficient
repeats
for thesewould be sufficient for these datasets.
datasets.
Figure 16. Typical

Figure 16. Typical example
example of
of the
the effect
effectof
ofrepetition.
repetition.
Next,
Next, we we evaluate
evaluate the
the total
total work
work based
based onon the
the number
number ofof iterations
iterations and
and the
the number
number of of repeats.
repeats.
We plot four different curves with random start repeated 1, 5, 15, and 25
We plot four different curves with random start repeated 1, 5, 15, and 25 times for the O–Mopsitimes for the O–Mopsi
instances.
instances. Figure
Figure1717 shows
showshow closeclose
how the path
the length becomes
path length to the optimum
becomes during 10,000
to the optimum iterations.
during 10,000
A similar plot
iterations. is also drawn
A similar plot is for
alsothe Dots instances
drawn withinstances
for the Dots 1, 5, 15, 25, 50,1,
with 100,
5, and 15050,
15, 25, repeats. All the
100, and 150
Appl. Sci.coincide
curves 2019, 9, x FOR
on PEERother
each REVIEW so they appear as only one line in Figure 17. 16 of 26
repeats. All the curves coincide on each other so they appear as only one line in Figure 17.
Figure 17. Total amount of work.

Figure 17. Total amount of work.
4.2. Processing Time

The time complexity of the method can be analyzed as follows. At each iteration, the algorithm
tries to find improvement by a random local change. The effect of this change is tested in constant
Appl. Sci. 2019, 9, 3985 16 of 24
4.2. Processing Time

The time complexity of the method can be analyzed as follows. At each iteration, the algorithm
tries to find improvement by a random local change. The effect of this change is tested in constant time.
Only when improvement is found, the solution is updated accordingly, which takes O(N) time. Let us
assume probability p for finding an improvement. The time complexity is, therefore, O(IR + pNIR),
where I and R are the number of iterations and repetitions, respectively. The value for p is higher
at the beginning, when the configuration is close to random and gradually decreases to zero when
approaching to a saturation point (the solution that cannot be improved by any local operation). In our
observations, p is small (0.002 on average for O–Mopsi and even lesser for Dots), so the time complexity
reduces to O(IR), in practice. This means that the speed of the algorithm is almost independent on the
problem size (see Figure 18).
In contrast, optimal solvers such as Concorde [17] have exponential time complexity. We are
unable to find a complexity analysis for Concorde in literature, however, we estimate it empirically by
analyzing the running time required by Concorde to solve TSPLIB [44] instances. From the website
(http://www.math.uwaterloo.ca/tsp/concorde/benchmarks/bench99.html), we choose 105 instances with
increasing problem sizes (N between 14 and 13509) and analyze the trend. According to this, Concorde
has a complexity of (1.5eˆ(0.004N)). This implies that for small instances (up to several thousands) the
algorithm is expected to be fast. However, the analysis also reveals that some points highly deviate
from the expected line. This implies that some instances are more difficult to solve (with a higher
number of nodes in the search tree) and require more time, such as the instance (d1291), with 1291
targets, which
Appl. Sci. 2019, 9,has almost
x FOR PEER 8 hours of running time where another instance (rl1304) with 1304 targets
REVIEW 17 of 26
has only 22 minutes of running time. In comparison to that, Random Mix finishes more predictably.
Figure 18. Run time as a function of problem size.
Figure
Considering the repeats, we 18. Run time
summarized theasresult
a function of problem
in Table 3, whichsize.
shows that a single run takes
0.8 ms and 25 repeats take 16 ms on an average, regardless of the dataset.
Considering the repeats, we summarized the result in Table 3, which shows that a single run
takes 0.8 ms and 25 repeats take 16 msTable
on an3.average,
Running regardless
time. of the dataset.
Single
Table 3. Run
Running time.25 Repeats
O–Mopsi 0.8 ms 16 ms
Dots 0.7Single
ms run 25 repeats
16 ms
O–Mopsi 0.8 ms 16 ms
4.3. Parameter Values
Dots 0.7 ms 16 ms
We investigate different values for the I and R parameters using grid-search and recorded how
many operations are needed to solve different sized problems. The results are summarized in Figure 19.
In4.3.
thisParameter Values
context, we want to mention, as there are not a sufficient number of samples with more than 25
targetsWein the O–Mopsidifferent
investigate dataset and 30 targets
values for the Iinand
the RDots dataset, using
parameters we ignore those samples.
grid-search From the
and recorded how
many operations are needed to solve different sized problems. The results are summarized in
Figure 19. In this context, we want to mention, as there are not a sufficient number of samples with
more than 25 targets in the O–Mopsi dataset and 30 targets in the Dots dataset, we ignore those
samples. From the chart, we can conclude that proper parameter values to use for the O–Mopsi
dataset are I = 213 and R = 25. The dots dataset requires higher values of I = 215 and R = 26. The
4.3. Parameter Values
We investigate different values for the I and R parameters using grid-search and recorded how
many operations are needed to solve different sized problems. The results are summarized in
Appl. Sci. 2019, 9, 3985
Figure 19. In this context, we want to mention, as there are not a sufficient number of samples17with of 24
more than 25 targets in the O–Mopsi dataset and 30 targets in the Dots dataset, we ignore those
samples.
chart, we From the chart,
can conclude weproper
that can conclude
parameter that proper
values parameter
to use values to dataset
for the O–Mopsi use forare
theI =
O–Mopsi
213 and
dataset
R = 2 . The
5 are dots
I = 2dataset
13 and Rrequires
= 2 . The
5 dotsvalues
higher dataset
of Irequires
= 2 andhigher
15 R = 2 values
6 of I = 2 and
15
. The relationship R = 2I 6and
between . TheR
relationship between I and R seems
seems to stabilize for larger instances. to stabilize for larger instances.
Figure 19. Required number

Figure 19.REVIEWnumber of
Required of operations
operations increases
increases with
with problem
problem size.
size.
Appl. Sci. 2019, 9, x FOR PEER 18 of 26
In Figure 20, we fix the product IR to 218 (O–Mopsi) and 221 (Dots) and vary their individual
In Figure 20, we fix the product IR to 218 (O–Mopsi) and 221 (Dots) and vary their individual
values. Typically,
values. Typically,different same-product
different same-productcombinations
combinations areare
equally effective
equally at solving
effective the instances.
at solving the
However,
instances. However, using too few repetitions, the algorithm typically gets stuck at awhich
using too few repetitions, the algorithm typically gets stuck at a local optimum, local is not
the global
optimum,one. On the
which other
is not the hand,
global when
one. Onthe
thenumber of iterations
other hand, when thebecomes
number of too small, itbecomes
iterations is likely that
sometoo small,
local it is likely can
operations that still
someimprove
local operations can(not
the result still saturated).
improve the result (not saturated).
Figure 20. Gap variation with varied iterations and repetitions.

Figure 20. Gap variation with varied iterations and repetitions.
4.4. The Productivity of the Operators

4.4. The Productivity of the Operators
We Wenextnext
analyze the
analyze productivity
the productivity of the operators,
of the operators,asas follows.
follows. WeWecountcount
howhow
manymany
timestimes
each each
operator found an improvement in the solution. For this case, we also evaluate
operator found an improvement in the solution. For this case, we also evaluate our operators by our operators by
closed-loop
closed-loopinstances (TSPLIB
instances (TSPLIBinstances)
instances)[44].
[44]. We
We should
should mention herethat
mention here thatclosed-loop
closed-loop instances do
instances
not contain singlesingle
do not contain links links
to swap. Therefore,
to swap. Therefore,thethe
link swap
link swapoperation
operationis
is not
not possible forthem.
possible for them.We We can
can only use the relocate and 2-opt operators to solve TSPLIB problems. In this
only use the relocate and 2-opt operators to solve TSPLIB problems. In this context, we want to context, we want to
mention
mention thatthat
we we choose
choose 12 12 TSPLIB
TSPLIB instancestototest
instances testwith
with problem
problem size
size of
of 52–3795.
52–3795.These
Theseinstances
instances are
are significantly larger than O–Mopsi and Dots instances, and therefore, we used a much higher
significantly larger than O–Mopsi and Dots instances, and therefore, we used a much higher number
number of iterations (108) and 25 repeats. The total distribution of the improvement achieved by
of iterations (108 ) and 25 repeats. The total distribution of the improvement achieved by operators is
operators is shown in Figure 21 both for the open-loop (O–Mopsi and Dot games) and closed-loop
shown in Figure 21 both for the open-loop (O–Mopsi and Dot games) and closed-loop (TSPLIB) [44]
(TSPLIB) [44] cases. From the results, we can make the following observations.
cases. FromFirst,the
weresults,
can seewethat
canthe
make theand
2-opt following
relocateobservations.
are almost equally productive in all cases.
However, the situation changes dramatically from the closed-loop to the open-loop case. While
2-opt and relocate both have about a 50% share of all improvements in the closed-loop case
(TSPLIB), they are inferior to the new link swap operator in case of open-loop test sets (O–Mopsi
and Dots). The link swap operator dominates by a 50% share, while 2-opt and relocate have roughly
a 25% share each. The relocate operator becomes also slightly more productive than 2-opt.
can only use the relocate and 2-opt operators to solve TSPLIB problems. In this context, we want to
mention that we choose 12 TSPLIB instances to test with problem size of 52–3795. These instances
are significantly larger than O–Mopsi and Dots instances, and therefore, we used a much higher
number of iterations (108) and 25 repeats. The total distribution of the improvement achieved by
Appl. Sci. 2019,
operators is 9,shown
3985 in Figure 21 both for the open-loop (O–Mopsi and Dot games) and closed-loop
18 of 24
(TSPLIB) [44] cases. From the results, we can make the following observations.
First, we can see that the 2-opt and relocate are almost equally productive in all cases.
First, we can see that the 2-opt and relocate are almost equally productive in all cases. However,
However, the situation changes dramatically from the closed-loop to the open-loop case. While
the situation changes dramatically from the closed-loop to the open-loop case. While 2-opt and relocate
2-opt and relocate both have about a 50% share of all improvements in the closed-loop case
both have about a 50% share of all improvements in the closed-loop case (TSPLIB), they are inferior to
(TSPLIB), they are inferior to the new link swap operator in case of open-loop test sets (O–Mopsi
the new link swap operator in case of open-loop test sets (O–Mopsi and Dots). The link swap operator
and Dots). The link swap operator dominates by a 50% share, while 2-opt and relocate have roughly
dominates by a 50% share, while 2-opt and relocate have roughly a 25% share each. The relocate
a 25% share each. The relocate operator becomes also slightly more productive than 2-opt.
operator becomes also slightly more productive than 2-opt.
Figure 21. Share of the improvements by all three operators both with the open-loop (O–Mopsi and
Figure
Dot 21. Share
games) of the improvements
and closed-loop (TSPLIB) by all three
problem operators
instances. bothTSBPLIB,
With with the we
open-loop
use 108 (O–Mopsi and
iterations and
Appl. Sci.
Dot 2019, 9, x FOR
games) andPEER REVIEW(TSPLIB) problem instances. With TSBPLIB, we use 10 8 iterations and
closed-loop 19 of 26
25 repeats.
25 repeats.
One question
One questionisiswhether
whetherallall operators
operators workwork equally
equally efficiently
efficiently at beginning
at the the beginningof theof the search,
search, when
when there is more room for improvements, as they do at the end
there is more room for improvements, as they do at the end of the search when the improvements of the search when the
improvements
are harder to find.are We
harder tofour
select find.typical
We select
O-Mopsi four game
typical O-Mopsiand
examples gamefindexamples and find
the productivity the
of the
productivity
operators of thethe
through operators through
total search the (Figure
spaces total search spaces
22). We plot(Figure
them in22). We 23.
Figure plotSimilarly,
them in Figure
Figure 23.
24
Similarly, Figure 24 shows the productivity of operators along the search spaces
shows the productivity of operators along the search spaces of four typical dot games. We plot those of four typical dot
games.in
games We plot those
Figure games
25. The in Figure
diagrams show25.thatThe diagrams
there show that
are no visible theretoare
patterns beno visible patterns
observed. to be
All operators
observed. All operators remain productive at all stages of the search. Link swap is the
remain productive at all stages of the search. Link swap is the most active and relocate is slightly more most active and
relocate is slightly more productive at the end of the
productive at the end of the search with the O–Mopsi instances. search with the O–Mopsi instances.
Figure 22. Productivity of the operators during the search with for 4 different O–Mopsi instances.
Appl. Sci. 2019, 9, 3985 19 of 24

Figure
Figure 23.23.Screenshots
Screenshotsofofthe
the 4 different
4 different O–Mopsi
O–Mopsi instances
instances (in(in
thethe same
same order).
order).
Figure 24. Productivity of the operators during the search with for 4 different Dots instances.
Appl. Sci. 2019, 9, 3985 20 of 24
Figure 25. Screenshots of 4 different dots instances (in the same order).
Figure 25. Screenshots of 4 different dots instances (in the same order).
5. Stochastic Variants
5. Stochastic Variants
We consider repeat to overcome the local optimum values. However, there are several other
methods consider
We repeat
to overcome to optimum
local overcomevalues
the local
usingoptimum values.
tabu search [45],However, there areannealing
simple simulated several other
and
methods to overcome local optimum values using tabu search [45], simple simulated annealing and
combined variants [46–50], genetic algorithm [51–54]. Besides these, Quintero–Araujo et al. [55]
combined variantslocal
combined iterated [46–50],
searchgenetic algorithm
and Monte [51–54]. Besides
Carlo simulation these,
on vehicle Quintero–Araujo
routing problems. We et al. [55]
compared
combined iterated local search and Monte Carlo simulation on vehicle routing problems. We
our results with some of these methods; additionally, we consider two stochastic variants of the local
compared
search: our results with some of these methods; additionally, we consider two stochastic variants
of the local search:
•• Tabu search [45]
[45]
•• Simulated annealing [47]
Tabu search is a metaheuristic first introduced by Glover [45]. The idea is to allow uphill
Tabu search is a metaheuristic first introduced by Glover [45]. The idea is to allow uphill moves
moves that worsen the solution to avoid being stuck in a local optimum. It forces the search to new
that worsen the solution to avoid being stuck in a local optimum. It forces the search to new directions.
A tabu list of previously considered solutions, or moves, is maintained to prevent the search from
going into cycles around the local optima.
We incorporate the tabu search with our local search as follows. When an improvement is found
in the solution, we mark the added links as tabu. These links are not allowed to be changed in the next
0.2·N iterations. We use the best improvement as the search strategy. Table 4 shows results of tabu
search for O-Mopsi and Dots games.
Simulated annealing (SA) is another approach to make the search stochastic. Although there
are a large number of adaptive simulated annealing variants, we consider a simpler one, the noising
method of Charon and Hudry [47]. The idea is to add random noise to the cost function to allow
uphill moves and therefore avoid local optima. The noise is a random number between r ∈ [0, ± rmax ],
and acts as a multiplier for the noisy cost function: f noise = r·f. The noise starts from rmax = 1 and
decreases at every iteration by a constant 2/T, where T is the number of iterations. The noise reaches r
= 0 halfway, and the search then reduces back to normal local search by using normal cost function
(f ) for the remaining iterations. We apply a random mixed local search with 25 repeats. Table 4 also
shows simulated annealing results for O-Mopsi and Dots instances.
Table 4. Comparison of Tabu search and simulated annealing (SA) result with random mixed local
search result with the same iteration and repetition for O–Mopsi and dot instances.
Repeated (25 times)

Appl. Sci. 2019, 9, 3985 21 of 24
Appl. Sci. 2019, 9, x FOR PEER REVIEW Gap (avg.) Not solved
22 of 26
Table 4. Comparison of Tabu

Random searchlocal
mixed and search
simulated annealing (SA) result
0% with random mixed0local
search result with the same iteration and repetition for O–Mopsi and dot instances.
O–Mopsi Tabu search 0% 0
Table 4. Comparison of Tabu search and simulated annealing (SA) result with
Repeated random mixed local
(25 times)
Simulated
search result with the annealing
same iteration (SA) for O–Mopsi and 0%
and repetition dot instances. 0
Gap (avg.) Not Solved
RandomRandom
mixed mixed
local search
local search 0.001%
0% 0 times) 1
Repeated (25
Dots Tabu search 0.0003% 3
Simulated annealing (SA) 0% (avg.)
Gap 0 Not solved
SA mixed local search
Random 0.007%
0.001% 1 10
Dots Random mixedTabu
localsearch
search 0%
0.0003% 3 0
SA 0.007% 10
Simulated annealing (SA) 0% 0

Finding the productivity of the operators as in Figure 21 for random mixed local search, we consider
finding out the sameRandom
in case of tabulocal
mixed search and simulated annealing.
search 0.001% From a random mixed 1 local
search, we found that link swap was the most effective operation. Figure 26 illustrates that in tabu
Dots link swap is the least effective.
search, Tabu searchHowever, link swap is again0.0003%
the most effective for the 3simulated
annealing, as shown in Figure 27. SA We mark the added links as forbidden
0.007% for tabu search, so10 link swap
becomes too restrictive in this case. This makes link swap often non-working.
Figure 26. Share of improvement by all three methods for Tabu search.
Finding the productivity of the operators as in Figure 21 for random mixed local search, we
consider finding out the same in case of tabu search and simulated annealing. From a random
mixed local search, we found that link swap was the most effective operation. Figure 26 illustrates
that in tabu search, link swap is the least effective. However, link swap is again the most effective
for the simulated annealing, as shown in Figure 27. We mark the added links as forbidden for tabu
search, so link swap becomes too restrictive in this case. This makes link swap often non-working.
Finding the productivity of the operators as in Figure 21 for random mixed local search, we
consider finding out the same in case of tabu search and simulated annealing. From a random
mixed local search, we found that link swap was the most effective operation. Figure 26 illustrates
that in tabu search, link swap is the least effective. However, link swap is again the most effective
for the simulated annealing, as shown in Figure 27. We mark the added links as forbidden for tabu
search, so link swap becomes too restrictive in this case. This makes link swap often non-working.
Figure 27. Share of improvement by all three methods for simulated annealing (SA).
We compare the performance of random mixed local search, tabu search, and simulated annealing
for O-Mopsi and dots games in Table 5. We study the gap value and execution time for every method.
Results show that neither tabu nor simulated annealing improved the performance of the random
mixed local search algorithm.
Appl. Sci. 2019, 9, 3985 22 of 24
Table 5. Summary of overall results (mean).
Results Random mix Tabu SA

Gap Time Gap Time Gap Time
O–Mopsi 0% <1 s 0% 1.3 s 0% <1 s
Dots 0.001% <1 s 0.0003% 1.3 s 0.007% <1 s
6. Discussion
We have studied the open-loop variant of TSP to find out which operators work best. Two new
operators were proposed: link swap and 3–permute. The link swap operator was shown to be the
most productive. Even though 2-opt works best as a single operator, the best results are obtained using
a mixture of all three operators (2-opt, relocate, link swap). The new operator, link swap, provides,
50% of all the improvements, while 2-opt and relocate have roughly a 25% share each. To sum up,
the link swap operator is the most efficient operator in the mix.
For our problem instances up to size 31, the proposed combination of the local search (random
mixed local search) finds the optimum solution in all O–Mopsi instances, and in all except one, Dots
problem instances. Iterations and repetitions are the most significant parameters of the proposed
method. Additionally, a suitable number of iterations and repetitions are essential with respect to
the problem size, which is 213 and 25 for the O–Mopsi dataset, and 215 and 26 for the Dots dataset.
For O–Mopsi instances, processing times are 0.8 ms (single) and 16 ms (repeats). For Dots instances,
processing times are 0.7 ms (single) and 16 ms (repeats). The overall complexity of the algorithm
mainly depends on the number of iterations and repeats. Furthermore, considering the multi-threaded
platform, different repeats can work simultaneously to decrease the processing time by a factor of a
number of threads.
Tabu search and simulated annealing were also considered but they did not provide further
improvements in our tests. The productivity of the operators was consistent with those of the local
search. We conclude that the local search is sufficient for our application where the optimum result is
needed in real-time, and the occasional sub-optimal result can be tolerated.
The limitation of the new operator, link swap, is that it is suitable only for the open-loop case
and does not apply to TSPLIB instances. Therefore, only 2-opt and relocate can be used for these
instances. Without link swap, relocate also becomes weaker in the mixing algorithm, which might
weaken the result.
Author Contributions: Conceptualization, P.F., R.M.-I., and L.S.; methodology, L.S., R.M.-I.; writing—original
draft preparation, L.S.; writing—review and editing, P.F and R.M.-I.; supervision, P.F.
Funding: This research received no external funding
Conflicts of Interest: The authors declare no conflict of interest.
References
1. Garey, M.R.; Johnson, D.S. Computers and Intractability: A Guide to the Theory of Np-Completeness; W.H. Freeman
& Co.: New York, NY, USA, 1979; ISBN 0716710455.
2. Papadimitriou, C.H. The Euclidean travelling salesman problem is NP-complete. Theor. Comput. Sci. 1977, 4,
237–244. [CrossRef]
3. Fränti, P.; Mariescu-Istodor, R.; Sengupta, L. O-Mopsi: Mobile Orienteering Game for Sightseeing, Exercising,
and Education. ACM Trans. Multimed. Comput. Commun. Appl. 2017, 13, 56. [CrossRef]
4. Vansteenwegen, P.; Souffriau, W.; Van Oudheusden, D. The orienteering problem: A survey. Eur. J. Oper. Res.
2011, 209, 1–10. [CrossRef]
5. Chieng, H.H.; Wahid, N. A Performance Comparison of Genetic Algorithm’s Mutation Operators in n-Cities
Open Loop Travelling Salesman Problem. In Recent Advances on Soft Computing and Data Mining. Advances in
Intelligent Systems and Computing; Herawan, T., Ghazali, R., Deris, M., Eds.; Springer: Berlin/Heidelberg,
Germany, 2014; Volume 287.
Appl. Sci. 2019, 9, 3985 23 of 24
6. Gavalas, D.; Konstantopoulos, C.; Mastakas, K.; Pantziou, G. A survey on algorithmic approaches for solving
tourist trip design problems. J. Heuristics 2014, 20, 291–328. [CrossRef]
7. Golden, B.L.; Levy, L.; Vohra, R. The Orienteering Problem. Nav. Res. Logist. 1987, 34, 307–318. [CrossRef]
8. Perez, D.; Togelius, J.; Samothrakis, S.; Rohlfshagen, P.; Lucas, S.M. Automated Map Generation for the
Physical Traveling Salesman Problem. IEEE Trans. Evol. Comput. 2014, 18, 708–720. [CrossRef]
9. Sengupta, L.; Mariescu-Istodor, R.; Fränti, P. Planning your route: Where to start? Comput. Brain Behav. 2018,
1, 252–265. [CrossRef]
10. Sengupta, L.; Fränti, P. Predicting difficulty of TSP instances using MST. In Proceedings of the IEEE
International Conference on Industrial Informatics (INDIN), Helsinki, Finland, June 2019; pp. 848–852.
11. Dantzig, G.B.; Fulkerson, D.R.; Johnson, S.M. Solution of a Large Scale Traveling Salesman Problem; Technical
Report P-510; RAND Corporation: Santa Monica, CA, USA, 1954.
12. Held, M.; Karp, R.M. The traveling salesman problem and minimum spanning trees: Part II. Math. Program.
1971, 1, 6–25. [CrossRef]
13. Padberg, M.; Rinaldi, G. A branch-and-cut algorithm for the resolution of large-scale symmetric traveling
salesman problems. SIAM Rev. 1991, 33, 60–100. [CrossRef]
14. Grötschel, M.; Holland, O. Solution of large-scale symmetric travelling salesman problems. Math. Program.
1991, 51, 141–202. [CrossRef]
15. Applegate, D.; Bixby, R.; Chvatal, V. On the solution of traveling salesman problems. Documenta Mathematica
Journal der Deutschen Mathematiker-Vereinigung. Int. Congr. Math. 1988, Extra Volume III, 645–656.
16. Laporte, G. The traveling salesman problem: An overview of exact and approximate algorithms. Eur. J. Oper.
Res. 1992, 59, 231–247. [CrossRef]
17. Applegate, D.L.; Bixby, R.E.; Chvatal, V.; Cook, W.J. The Traveling Salesman Problem: A Computational Study;
Princeton University Press: Princeton, NJ, USA, 2011.
18. Johnson, D.S.; Papadimitriou, C.H.; Yannakakis, M. How easy is local search? J. Comput. Syst. Sci. 1988, 37,
79–100. [CrossRef]
19. Clarke, G.; Wright, J.W. Scheduling of Vehicles from a Central Depot to a Number of Delivery Points.
Oper. Res. 1964, 12, 568–581. [CrossRef]
20. Christofides, N. Worst-Case Analysis of a New Heuristic for the Travelling Salesman Problem (Technical Report388);
Graduate School of Industrial Administration, Carnegie Mellon University: Pittsburgh, PA, USA, 1976.
21. Johnson, D.S.; McGeoch, L.A. The traveling salesman problem: A case study in local optimization. Local
Search Comb. Optim. 1997, 1, 215–310.
22. Croes, G.A. A Method for Solving Traveling-Salesman Problems. Oper. Res. 1958, 6, 791–812. [CrossRef]
23. Lin, S.; Kernighan, B.W. An effective heuristic algorithm for the traveling-salesman problem. Oper. Res. 1973,
21, 498–516. [CrossRef]
24. Rego, C.; Glover, F. Local search and metaheuristics. In The Traveling Salesman Problem and Its Variations;
Springer: Boston, MA, USA, 2007; pp. 309–368.
25. Okano, H.; Misono, S.; Iwano, K. New TSP construction heuristics and their relationships to the 2-opt.
J. Heuristics 1999, 5, 71–88. [CrossRef]
26. Johnson, D.S.; McGeoch, L.A. Experimental analysis of heuristics for the STSP. In The Traveling Salesman
Problem and Its Variations; Springer: Boston, MA, USA, 2007; pp. 369–443.
27. Aarts, E.; Aarts, E.H.; Lenstra, J.K. (Eds.) Local Search in Combinatorial Optimization; Princeton University
Press: Princeton, NJ, USA, 2003.
28. Jünger, M.; Reinelt, G.; Rinaldi, G. The traveling salesman problem. Handb. Oper. Res. Manag. Sci. 1995, 7,
225–330.
29. Laporte, G. A concise guide to the traveling salesman problem. J. Oper. Res. Soc. 2010, 61, 35–40. [CrossRef]
30. Ahuja, R.K.; Ergun Ö Orlin, J.B.; Punnen, A.P. A survey of very large-scale neighborhood search techniques.
Discret. Appl. Math. 2002, 123, 75–102. [CrossRef]
31. Rego, C.; Gamboa, D.; Glover, F.; Osterman, C. Traveling salesman problem heuristics: Leading methods,
implementations and latest advances. Eur. J. Oper. Res. 2011, 211, 427–441. [CrossRef]
32. Matai, R.; Singh, S.; Mittal, M.L. Traveling salesman problem: An overview of applications, formulations, and
solution approaches. In Traveling Salesman Problem, Theory and Applications; IntechOpen: London, UK, 2010.
33. Laporte, G. The vehicle routing problem: An overview of exact and approximate algorithms. Eur. J. Oper.
Res. 1992, 59, 345–358. [CrossRef]
Appl. Sci. 2019, 9, 3985 24 of 24
34. Vidal TCrainic, T.G.; Gendreau, M.; Prins, C. Heuristics for multi-attribute vehicle routing problems: A survey
and synthesis. Eur. J. Oper. Res. 2013, 231, 1–21. [CrossRef]
35. Gendreau, M.; Hertz, A.; Laporte, G. New insertion and postoptimization procedures for the traveling
salesman problem. Oper. Res. 1992, 40, 1086–1094. [CrossRef]
36. Mersmann, O.; Bischl, B.; Bossek, J.; Trautmann, H.; Wagner, M.; Neumann, F. Local Search and the Traveling
Salesman Problem: A Feature-Based Characterization of Problem Hardness. In Lecture Notes in Computer
Science, Proceedings of the Learning and Intelligent Optimization, Paris, France, 16–20 January 2012; Springer:
Berlin/Heidelberg, Germany, 2012; Volume 7219, p. 7219.
37. Helsgaun, K. An effective implementation of the Lin–Kernighan traveling salesman heuristic. Eur. J. Oper.
Res. 2000, 126, 106–130. [CrossRef]
38. Pan, Y.; Xia, Y. Solving TSP by dismantling cross paths. In Proceedings of the IEEE International Conference
on Orange Technologies, Xian, China, 20–23 September 2014; pp. 121–124.
39. Martí, R. Multi-start methods. In Handbook of Metaheuristics; Springer: Berlin/Heidelberg, Germany, 2003;
pp. 355–368.
40. O’Neil, M.A.; Burtscher, M. Rethinking the parallelization of random-restart hill climbing: A case study in
optimizing a 2-opt TSP solver for GPU execution. In Proceedings of the 8th Workshop on General Purpose
Processing using GPUs, San Francisco, CA, USA, 7 February 2015; pp. 99–108.
41. Al-Adwan, A.; Sharieh, A.; Mahafzah, B.A. Parallel heuristic local search algorithm on OTIS hyper hexa-cell
and OTIS mesh of trees optoelectronic architectures. Appl. Intell. 2019, 49, 661–688. [CrossRef]
42. Xiang, Y.; Zhou, Y.; Chen, Z. A local search based restart evolutionary algorithm for finding triple product
property triples. Appl. Intell. 2018, 48, 2894–2911. [CrossRef]
43. Lawler, E.L.; Lenstra, J.K.; Rinnooy Kan AH, G.; Shmoys, D.B. The Traveling Salesman Problem; A Guided Tour
of Combinatorial Optimization; Publisher Wiley: Chichester, UK, 1985.
44. Reinelt, G. A traveling salesman problem library. INFORMS J. Comput. 1991, 3, 376–384. [CrossRef]
45. Glover, F. Tabu Search-Part I. ORSA J. Comput. 1989, 1, 190–206. [CrossRef]
46. Kirkpatrick, S.; Gelatt, C.D.; Vecchi, M.P. Optimization by simulated annealing. Science 1983, 220, 671–680.
[CrossRef] [PubMed]
47. Charon, I.; Hurdy, O. Application of the noising method to the travelling salesman problem. Eur. J. Oper. Res.
2000, 125, 266–277. [CrossRef]
48. Chen, S.M.; Chien, C.Y. Solving the traveling salesman problem based on the genetic simulated annealing
ant colony system with particle swarm optimization techniques. Expert Syst. Appl. 2011, 38, 14439–14450.
[CrossRef]
49. Ezugwu, A.E.S.; Adewumi, A.O.; Frîncu, M.E. Simulated annealing based symbiotic organisms search
optimization algorithm for traveling salesman problem. Expert Syst. Appl. 2017, 77, 189–210. [CrossRef]
50. Geng, X.; Chen, Z.; Yang, W.; Shi, D.; Zhao, K. Solving the traveling salesman problem based on an adaptive
simulated annealing algorithm with greedy search. Appl. Soft Comput. 2011, 11, 3680–3689. [CrossRef]
51. Albayrak, M.; Allahverdi, N. Development a new mutation operator to solve the Traveling Salesman Problem
by aid of Genetic Algorithms. Expert Syst. Appl. 2011, 38, 1313–1320. [CrossRef]
52. Nagata, Y.; Soler, D. A new genetic algorithm for the asymmetric traveling salesman problem. Expert Syst.
Appl. 2012, 39, 8947–8953. [CrossRef]
53. Singh, S.; Lodhi, E.A. Study of variation in TSP using genetic algorithm and its operator comparison. Int. J.
Soft Comput. Eng. 2013, 3, 264–267.
54. Vashisht, V.; Choudhury, T. Open loop travelling salesman problem using genetic algorithm. Int. J. Innov.
Res. Comput. Commun. Eng. 2013, 1, 112–116.
55. Quintero-Araujo, C.L.; Gruler, A.; Juan, A.A.; Armas, J.; Ramalhinho, H. Using simheuristics to promote
horizontal collaboration in stochastic city logistics. Prog. Artif. Intell. 2017, 6, 275. [CrossRef]
© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

Applied Sciences: Which Local Search Operator Works Best For The Open-Loop TSP?

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Applied Sciences: Which Local Search Operator Works Best For The Open-Loop TSP?

Uploaded by

Copyright:

Available Formats

applied

Appl. Sci. 2019, 9, 3985; doi:10.3390/app9193985 www.mdpi.com/journal/applsci

2. Examples of O–Mopsi playing (left), the final tour (middle)

Table 1. Datasets used in this study.

Dataset: Type: Distance: Instances: Sizes:

2.2. Two-Optimization (2-opt)

Figure 5. Two examples of the 2-opt method in O-Mopsi.

2.3. Three-Node Permutation

2.4. Link Swap

2.4. Link Swap

2-opt: In thisFigure 8. Six neighboring

3.2. Search Strategy

IF Length ( NewPath ) < Length ( Path ) THEN

NewBestSolution ( Path, OperationType )

3.3. Results With A Single Operator

Single Operator Double Operators Mix of Three

3.4. Combining the Operators

Figure 12. Operators have complementary search spaces.

Algorithm 3: Finding a solution by mixing local search operators

Figure 15. Instances solved as a function of the number of repetitions.

Figure 16. Typical

Figure 17. Total amount of work.

4.2. Processing Time

4.2. Processing Time

Figure 18. Run time as a function of problem size.

Figure 19. Required number

Figure 20. Gap variation with varied iterations and repetitions.

4.4. The Productivity of the Operators

Appl. Sci. 2019, 9, x FOR PEER REVIEW 20 of 26

Repeated (25 times)

Table 4. Comparison of Tabu

Simulated annealing (SA) 0% 0

Table 5. Summary of overall results (mean).

Results Random mix Tabu SA

You might also like