Annals of Nuclear Energy 139 (2020) 107214

Contents lists available at ScienceDirect

Annals of Nuclear Energy

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

Application of Cuckoo Search algorithm to Loading Pattern Optimization

problems
Anderson Alvarenga de Moura Meneses a,⇑, Patrick Vasconcelos da Silva a, Fernando Nogueira Nast a,
Lenilson Moreira Araujo a, Roberto Schirru b
a
Federal University of Western Pará, Laboratory of Computational Intelligence, Av. Vera Paz, s/n, Salé, CEP 68.035-110 Santarém, PA, Brazil
b
Federal University of Rio de Janeiro, COPPE, Nuclear Engineering Program, Brazil

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

Article history: The Loading Pattern Optimization (LPO) is related to important goals in a Nuclear Power Plant (NPP)
Received 6 September 2019 operation such as the extension of the cycle according to safety margins. The LPO is a combinatorial prob-
Received in revised form 11 November 2019 lem of relevance and interest for Nuclear Engineering. Optimization metaheuristics have been efficient in
Accepted 15 November 2019
solving the LPO. The recent metaheuristic Cuckoo Search (CS) is based on the brood parasitism of some
cuckoo species, combined with the behavior of the Lévy flight of some birds. In the present work the
results of the application of CS to the LPO using IAEA-3D and BIBLIS-2D benchmarks are presented, as
Keywords:
well as the application of CS in the optimization of 7th cycle of Angra 1 NPP, in Brazil. The results are com-
Loading Pattern Optimization
Nuclear Power Plant
pared to the metaheuristics Artificial Bee Colony and Population-Based Incremental Learning. Statistical
Optimization metaheuristics analyses show that CS is the most robust algorithm for the set of instances selected for tests.
Pressurized water reactor Ó 2019 Elsevier Ltd. All rights reserved.
Cuckoo Search
Nuclear fuel

1. Introduction Particle Swarm Optimization (PSO; Meneses et al., 2009), Artificial

Bee Colony (ABC; de Oliveira and Schirru, 2011), and Cross-Entropy
The Loading Pattern Optimization (LPO) is related to important (CE; Meneses and Schirru, 2015).
goals in a Nuclear Power Plant (NPP) operation such as the exten- In the present work, CS was applied to the benchmarks IAEA-3D
sion of the operation cycle by determining the position of the Fuel (Argonne National Laboratory, 1977) and BIBLIS-2D (Poursalehi
Assemblies (FAs) in the reactor core, within adequate safety mar- et al., 2013), using the PARCS v3.0 code (U.S. NRC Core Neutronics
gins (Levine, 1986). The LPO is a prominent problem of economic Simulator; Downar et al., 2009a,b; Joo et al., 1998) as well as to the
interest, as well as a complex combinatorial problem, classified 7th cycle of Angra 1 NPP, located in Brazil, using the RECNOD code
as NP-hard (Papadimitriou and Steiglitz, 1982), with a considerable (Chapot, 2000).
number of feasible solutions, sub-optimal solutions, disconnected The remaining of the present article is organized as follows:
feasible regions, high dimensionality, complex and time- Section 2 contains a review of previous works; in Section 3 the the-
consuming evaluation of candidate solutions using Reactor Physics oretical background is presented; Section 4 describes the method-
codes (Stevens et al., 1995). ology used to solve the LPO problems; Section 5 presents the
Computational intelligence methods such as Optimization results; the discussion is presented in Section 6; and finally,
Metaheuristics (OMH) are efficient in solving combinatorial prob- Section 7 is devoted to the conclusions.
lems such as the LPO. Several OMHs have been successful when
applied to LPO, including: Simulated Annealing (SA; Parks, 1990;
Stevens et al., 1995), Genetic Algorithm (GA; Chapot et al., 1999; 2. Related works
Poon and Parks, 1992), Tabu Search (TS; Hill and Parks, 2015; Lin
et al., 1998), Population-Based Incremental Learning (PBIL; The CS algorithm was first presented by Yang and Deb (2009).
Caldas and Schirru, 2008; Machado, 2005), Ant Colony Optimiza- The CS is based on the strategy of reproduction of some cuckoo
tion (ACO; de Lima et al., 2008; Machado and Schirru, 2002), species combined with the behavior of the Lévy Flight of some
birds, and is a swarm intelligence algorithm (Engelbrecht, 2007).
Ouaarab et al. (2014) applied a CS to the Traveling Salesman
⇑ Corresponding author. Problem (TSP). In such work the Improved Discrete CS (DCS) uses
E-mail address: [email protected] (A.A.M. Meneses). a fraction of intelligent cuckoos to explore other areas of the search

https://doi.org/10.1016/j.anucene.2019.107214
0306-4549/Ó 2019 Elsevier Ltd. All rights reserved.
2 A.A.M. Meneses et al. / Annals of Nuclear Energy 139 (2020) 107214

space, using the local search heuristic 2-opt (Johnson and neutron diffusion theory. In the present work both reactors were
McGeoch, 1997). DCS and its improved version were applied to modeled considering the symmetry of ¼ in PARCS code, and during
41 instances of the problem and compared to the Genetic Simu- the OMH search new candidate solutions (LPs) are formed with
lated Annealing-Ant Colony System with PSO Techniques (GSA- octant symmetry. FAs on symmetry lines are not permuted with
ACS-PSOT; Chen and Chien, 2011), as well as to the Discrete PSO FAs off symmetry lines as described by Meneses et al. (2018).
(DPSO). Improved DCS obtained superior results in comparison to Despite being a two-dimensional reactor, BIBLIS-2D presents a
GSA-ACS-PSOT and DPSO for the TSP. much higher level of complexity than the IAEA-3D regarding the
CS has also been applied to a nuclear reload problem by LPO, because of the number of candidate LP solutions. The number
Yarizadeh-Beneh et al. (2016), in order to optimize the reload of of candidate solutions for the BIBLIS-2D reactor is approximately
the first cycle of the reactor located in the Bushehr NPP in Iran, 4:0  1016 whereas for the IAEA-3D reactor the total number is
using Random Keys (RK; Bean, 1994). For thermal-neutronic calcu- 288,288. More information about the geometry, boundary condi-
lations, CS was coupled to PARCS reactor physics code, WIMSD-5B, tions, and nuclear parameters of those two benchmarks is given
and COBRA-5B, obtaining results that outperform those obtained in by Meneses et al. (2018).
the original reload by a designer, extending the cycle in up to
7 days. Yarizadeh-Beneh et al. (2016) applied CS to the LPO of a
3.1.2. 7th cycle of Angra 1 Nuclear Power Plant (Brazil)
real-world WWER NPP, and calculations for beginning of cycle
The Angra 1 NPP is located in the Rio de Janeiro state, Brazil,
(BOC) and end of cycle (EOC) states. In the present work the Critical
with a 626 MW PWR designed by Westinghouse and operated by
Boron Concentration is calculated at the equilibrium of Xenon in
Eletronuclear. Table 1 shows the data for the 7th cycle of Angra 1
the RECNOD code (7th cycle of Angra 1 NPP), and at the BOC in
NPP.
the IAEA-3D and BIBLIS-2D LPO problems.
The symmetry of 1/8 Angra 1 PWR yields 21 FAs for permuta-
For the application of CS to the Angra 1 – 7th cycle LPO, Silva
tion. In our implementations, the ten quartet elements (on symme-
et al. (2017) performed a preliminary investigation. In the present
try lines) are not permuted with the ten octet elements (off-
work, the CS was also applied to two well-known benchmark prob-
symmetry lines). Considering that the central FA is not permuted,
lems (IAEA-3D and BIBLIS-2D), being compared favorably with the
results of two competitive metaheuristics for the LPO, namely the the total number of possible solutions is 10!  10! ffi 1:3  1013 .
ABC and PBIL algorithms.
The results of the present work are compared to the results 3.2. Cuckoo Search algorithm
described by Meneses et al. (2018), who reported the application
of the PSO, CE, PBIL, and ABC algorithms to LPO problems based The Cuckoo Search (CS) algorithm is a nature-inspired OMHs
on the data of the benchmarks BIBLIS-2D (Poursalehi et al., 2013) within the swarm intelligence paradigm, and was initially pro-
and IAEA-3D (Argonne National Laboratory, 1977), both of them posed by Yang and Deb (2009) for the solution of multimodal prob-
modelled with the Reactor Physics code PARCS as well as the Angra lems. CS is based on the breeding strategy of some cuckoo species,
1
7th Cycle LPO problem, which was modelled with the RECNOD the so-called brood parasitism. In this strategy the cuckoo places
code (Chapot et al., 1999). According to Meneses et al. (2018) the its eggs in nests of birds of other species, so that the hosts raise
best algorithms were ABC and PBIL respectively for IAEA-3D and the offspring after the egg hatching.
for BIBLIS-2D. For the problem of Angra 1 NPP modeled with the The evolution of some cuckoo species has occurred in such a
reactor physics code RECNOD the algorithms PBIL, ABC, and CE way that some females are experts at mimicking the color and pat-
achieved the best results. tern of host bird eggs, thus reducing the chances of their eggs being
rejected (Payne, 2005). Occasionally, host birds discover these par-
asitic eggs and discard them or build a new nest elsewhere.
3. Theoretical background
In addition, the algorithm is improved by using the Lévy flight
instead of simple random isotropic search (Brown et al., 2007).
3.1. Loading Pattern Optimization (LPO)
The Lévy flight represents a random search model characterized

The LPO aims to determine LPs that maximize energy produc-

tion satisfying safety requirements (Levine, 1986). The LPO is an
NP-hard problem, which means that its complexity grows non- Table 1
polynomially, with a discontinuous and multimodal search space Burnup and kinf data for the 7th cycle of Angra 1 NPP simulated by RECNOD code.
(de Lima, 2005). FA Burnup (MWD/TU) kinf
The substitution of burned FAs is necessary when it is not pos-
1 9603 1.069
sible to keep the reactor running at nominal power, or more pre- 2 13,045 0.906
cisely after a period called the operating cycle. After each cycle 3 7882 1.087
the FAs present different neutronic characteristics. For example, 4 13,006 0.906
the FAs are burned in different ways depending on the exposure 5 0 1.187
6 13,012 0.906
time and position in the reactor core, with the accumulation of fis-
7 14,650 1.037
sion products such as Xenon and Samarium. Consequently, only 8 8622 1.079
some FAs are reused in the next cycle, and approximately 1/3 or 9 13,181 0.903
1/4 of the FAs are replaced by new ones (Hill and Parks, 2015; 10 0 1.193
11 14,068 1.026
Meneses et al., 2010). Thus, forming an LP by fresh and burned
12 13,115 0.906
FAs characterizes the LPO as a combinatorial problem. 13 13,135 0.904
The benchmark LPO problems used for tests in the present work 14 0 1.188
are described below. 15 0 1.194
16 11,404 1.050
17 7873 1.099
3.1.1. IAEA-3D and BIBLIS-2D benchmarks 18 0 1.191
IAEA-3D and BIBLIS-2D are theoretical problems taken as 19 0 1.188
20 13,285 0.907
reference for the validation of numerical methods applied to the
 A.A.M. Meneses et al. / Annals of Nuclear Energy 139 (2020) 107214 3

by steps that obey a power law distribution and has its length v Nð0; r2v Þ ð6Þ
defined by the Lévy distribution (Eq.1), which has infinite mean
with
and infinite variance (Yang and Deb, 2009). Researches show that
8 91=b
the behavior of some predators in the search for preys also follows <Cð1 þ bÞ sin pb=
the characteristics of Lévy flights (Yang and Deb, 2013). This model ru ¼ 2
and rv ¼ 1 ð7Þ
is commonly represented by small steps followed by big ones : C1þb b 21
b
2 ;
2
(Brown et al., 2007; Shlesinger et al., 1995).
A symmetrical Lévy stable process {x} (Mantegna, 1994) is a where C is the standard Gamma distribution. The CS algorithm is
stochastic process which has a probability density function L(x) shown in Fig. 1.
in integral form that can be expressed as a series expansion. It
can be shown that the leading term of the L(x) series is propor- 3.2.1. Random Keys (RK)
tional to x -1-b and In the CS for the LPO problem feasible combinatorial candidate
solutions are obtained with RK. The RK model was proposed by
LðxÞ  jxj
1
b ; 0 < b  2; ð1Þ Bean (1994) and consists of transforming vectors of real numbers
into possible solutions of combinatorial problems. In Fig. 2 the
where b is the Lévy index. It is worth to notice that Cauchy distribu-
RK steps are described for a hypothetical five-dimensional problem
tion and Gaussian distribution are special cases of the Lévy distribu-
(for an LP with five FAs): (a) a real vector is obtained by the CS
tion respectively with b = 1 and b = 2. In addition, Gutowski (2002)
algorithm; (b) an integer vector is associated with the real vector;
discusses the restrictions in b values.
(c) the real vector is sorted; (d) an associated permutation of the
According to Yang (2014), a simple description of the CS meta-
integer vector is obtained; (e) finally the permutation will repre-
phor is that each candidate solution is represented by an egg laid
sent an LP (candidate solution) to be evaluated.
by one cuckoo in a randomly selected nest. Modelling more com-
Fig. 3 depicts the RK method for generating a combinatorial
plicated versions is also possible. In that simple case, in a search
solution for the LPO: (a) one candidate solution with 20 dimen-
space with d dimensions, at iteration t the quality of n eggs (that
sions is generated by the CS algorithm; (b) since in our implemen-
is, n candidate solution vectors xti where i 2 {1, . . ., n} where each
tations symmetry lines FAs are not permuted with FAs off
vector has d components) is then evaluated with a fitness function.
symmetry lines, the 20 keys are split in two groups (the order of
The main idea is that the best candidate solutions are kept and the
the numbers is also represented, as in Fig. 2 – step b); (c) each real
worst ones are replaced.
group is sorted and then the integer numbers change their posi-
The CS uses a balanced combination of local search and global
tions; (d) the reference LP is taken as a base and each FA is repre-
search. The local search (or local random walk according to Yang,
sented by an integer number; (e) finally, symmetry lines FAs are
2014) is given by
shuffled following the first group order, as well as FAs off symme-


xtþ1 ¼ xti þ a1 s  Hðpa
2Þ  xtj
xtk ð2Þ try lines according to the second group order.
i

where xti is the ith candidate solution at iteration t; a1 is the scaling 4. Methodology
factor and in the present work empirically defined as 1, according to
the results of preliminary tests; s is the step size, which is a random 4.1. LPO with the Cuckoo Search algorithm
number over the range ½0; 1; Hð Þ is the Heaviside step function; the
commutation parameter pa 2 (0, 1) is the probability of an egg being The IAEA-3D and BIBLIS-2D benchmarks were implemented
discovered by a host bird;  is a random real number obtained from with the PARCS reactor physics code using the same configurations
a uniform distribution over the range ½0; 1; and xtj and xtk are two presented in Meneses et al. (2018), that is, a combination of the
different solutions randomly selected from current solutions, repre- Nodal Expansion Method (Finnemann et al., 1977) and Analytical
sented with the indices j and k 2 {1, . . ., n}. The Hadamard multipli- Nodal Methods (Smith, 1979) with 2  2 spatial discretization in
cation operator  represents the entrywise product between each FA. Critical Boron search is performed without considering
matrices or vectors with the same dimensions. thermal-hydraulic feedback. The optimization parameter used as
The global search is given by a safety constraint for the IAEA-3D and BIBLIS-2D benchmarks is
the peak power factor F q , defined as the quotient between the
xtþ1
i ¼ xti þ a2  LðbÞ ð3Þ maximum power of each FA and the average power of the whole
core.
where a2 in the present work is defined as 0.01 so that the algo-
All the benchmarks were simulated with a null initial critical
rithm does not make very large jumps, as recommended by Yang
Boron concentration. Under such conditions, the original LPs of
and Deb (2010). LðbÞ is a number drawn from the Lévy distribution
the benchmarks IAEA-3D and BIBLIS-2D yield respectively the
(Eq. (1)), with b = 1.5 in the present work as also implemented by
eigenvalues keff = 1.029096 and keff = 1.025368, which are in accor-
Yang and Deb (2010). Computationally LðbÞ can be determined in
dance to the reference values. Under the same conditions, when
several ways, but the most direct way is the use of the so-called
the critical Boron search is performed, the values obtained are
Mantegna Algorithm for a stable and symmetric Lévy distribution
respectively CB = 344.70 ppm and CB = 325.71 ppm. For a compar-
(Mantegna, 1994; Yang, 2010; Yang and Deb, 2013), which can be
ison between CS and the ABC and PBIL algorithms described by
summarized as
Meneses et al. (2018), the CB values yielded by the PARCS code
u were multiplied by 0.01 so that CB ranges from 0 to less than
LðbÞ  ðxtbest
xti Þ ð4Þ
jv j b
1
2000 ppm, which are typical values of a real PWR, for which the
absolute Boron Worth is approximately 7–10 pcm/ppm (see Oka
where xtbest is the candidate solution with the best fitness evalua- et al., 2014, p. 196 and p. 197).
tion, determined by the CS algorithm until the iteration t. In addi- The nuclear parameters produced by the RECNOD code are the
tion, u and m are obtained from the normal distribution, that is, Boron Concentration (C B ) in the Xenon equilibrium, so that, accord-
ing to Chapot (2000), 4 ppm correspond to 1 Effective Full Power
u  Nð0; r2u Þ ð5Þ
Day (EFPD), used to estimate the duration of the cycle (optimiza-
and tion criterion) and the Maximum Normalized Power of FA (P rm )
4 A.A.M. Meneses et al. / Annals of Nuclear Energy 139 (2020) 107214

Fig. 1. Pseudocode of the CS algorithm.

where k = 1 ppm is a constant used to maintain the fitness dimen-

sionless. The discontinuity in S0 does not represent a drawback for
the performance of OMHs, which do not depend on derivatives.
Meneses et al. (2009), Meneses and Schirru (2015), and Meneses
et al. (2018) also used Eq. (10) as fitness function for the OMHs
PSO, CE, PBIL, and ABC. Caldas and Schirru (2008) used another dis-
continuous fitness function for the OMH Parameter-Free PBIL
(FPBIL).
Hill and Parks (2015) give examples of objective functions used
in the LPO problem. In this sense, the fitness function could include
other parameters or be otherwise formulated (e.g. see Caldas and
Fig. 2. Random Keys example. Schirru, 2008; Yarizadeh-Beneh et al., 2016). However, given that
in the present work we compare CS to ABC and PBIL, the fitness
function is the same as the one used by Meneses et al. (2018) for
used as a safety parameter (constraint), corresponding to the F q of a fair comparison.
the PARCS code. The CS code and all interfaces used to calculate the fitness func-
In our implementations the objective function for the LPO tions calling the executable files of reactor physics codes were
according to the parameters generated by the reactor physical implemented in C language with the Microsoft Visual Studio Com-
codes is to minimize munity Edition 2017 IDE on a Windows 10 operating system. The
Windows 7 operating system was used to run the tests. The com-
1 puters used were two HPÒ desktop computers with an IntelÒ Core
ð8Þ
CB i7-3440 (3.40 GHz) processor and 8.0 GB of RAM.
The parameters used in CS are shown in Table 2. Fifty indepen-
subject to
dent tests (i.e., with different seeds) were performed.
Sconstr  S0 ð9Þ
where Sconstr is a security constraint variable (F q for IAEA-3D and 4.2. Statistical analyses
BIBLIS-2D, as well as P rm for Angra 1
7th Cycle) and S0 is a real
number. Prm = 1.395 for Angra 1
7th Cycle assures the compliance Since OMHs are stochastic algorithms and some variability is
with technical safety requirements according to Chapot (2000), and expected in the results, the same algorithm can perform differently
was also used in other works (Chapot et al., 1999; Meneses et al., depending on the problem. Wolpert and Macready (1997) pre-
2009; Meneses and Schirru, 2015). Fq = 2.200 for IAEA-3D and sented No Free Lunch (NFL) theorems that address the average per-
Fq = 1.400 for BIBLIS-2D were arbitrated by Meneses et al. (2018) formance of general-purpose algorithms concerning all possible
according to preliminary tests and also used in the present work. problems, which was proven identical for static and time-
The aggregated fitness function (considering that the values of dependent problems, among other considerations. In other words,
Sconstr are always greater than reciprocal of the Boron concentra- in such conditions there is no ‘‘always-the-best” algorithm. How-
tion) is ever, as pointed out by Joyce and Herrmann (2018), there are still
successful OMHs, used in specific sets of problems, which possess
(
k C1B ; ifScontr  S0 certain structures and characteristics, for which some algorithms
Fitness ¼ ð10Þ are better suited and therefore such algorithms perform better
Sconstr ; otherwise
than others on average in those problems. So, the better an OMH’s
 A.A.M. Meneses et al. / Annals of Nuclear Energy 139 (2020) 107214 5

Fig. 3. RK example for configurating a combinatorial candidate solution (LP) for the 7th cycle of Angra 1 NPP.

Table 2 than a cut-off level indicates that there is significant difference

Parameters of the CS algorithm, in each problem. between the groups. The ANOVA only verifies if there is an overall
RECNOD BIBLIS IAEA
statistical difference between the groups, thus a pairwise compar-
ison between the OMHs is needed (that is, a post-hoc test) and in
Nests 100 100 25
our work the Tukey’s studentized range (Honestly Significant Dif-
pa 25% 25% 25%
Iterations 500 500 20 ference, HSD) test was used.
Evaluations 100,000 100,000 1000 In the Shapiro-Wilk test, the null hypothesis is that the sample
comes from a normal distribution. A p-value less than a cut-off
value indicates that the sample does not come from a normal dis-
overall performance in a set of problems, the more robust such tribution and thus the parametric ANOVA cannot be used, that is,
OMH is in that set of problems. the assumption (b) is not fulfilled. The null hypothesis of Levene’s
In order to compare if there are significant differences between test is that the variances are homogeneous. If a p-value is less than
the algorithms, statistical tests were performed. If certain assump- a cut-off level then the assumption (d) is not fulfilled and ANOVA
tions are fulfilled, such statistical differences can be assessed by cannot be used either (see also García et al., 2009).
the parametric test Analysis of Variance (ANOVA; see Rutherford, Since such population parameters assumptions do not hold in
2011). Such assumptions are: (a) samples randomly obtained; (b) many cases for comparing algorithms, some authors discuss the
normal distribution of the results; (c) results obtained application of nonparametric procedures for machine learning
independently; and (d) homogeneity of variances. The use of a algorithms (e.g. Demšar, 2006; García et al., 2010) and OMHs
pseudorandom number generator guarantees the fulfilment of (e.g. García et al., 2009).
assumption (a). Regarding the assumption (c), according to Thus, the Kruskal-Wallis test (see Sheskin, 2003), which is the
García et al. (2009) the algorithms’ runs are independent events nonparametric equivalent of ANOVA, was performed in all cases,
because the occurrence of one run does not modify the probability whether the assumptions concerning population parameters were
of the another run since those events have different initial seeds for fulfilled (IAEA-3D) or not (BIBLIS-2D and Angra 1
7th Cycle), as
the pseudorandom number generation. In the present work the described in section 5. The Kruskal-Wallis test statistic H can be
assumption (b) was assessed with the Shapiro-Wilk test and approximated by a v2 distribution and if the p-value is less than
the assumption (d), with the Levene’s test. a cut-off level then there is statistical difference among the mean
In the one-way ANOVA the F-statistic is the ratio of between- ranks of the groups. Such as ANOVA, the Kruskal-Wallis test veri-
groups to within-groups variations. The null hypothesis is that fies if there is an overall statistically significant difference and
there is no significant difference between the groups. A p-value less therefore a pairwise post-hoc test is also needed. Thus the Dunn’s
6 A.A.M. Meneses et al. / Annals of Nuclear Energy 139 (2020) 107214

test was used (Dmitrienko et al., 2007), complementing the appli- The final CB results of the ABC, PBIL, and CS algorithms were
cation of the Kruskal-Wallis test. normal (Shapiro-Wilk test, respectively with p-values 0.4194,
The statistical analyses were performed with SAS/STATÒ 0.1459, and 0.3712). According to Levene’s test, CB variances were
University Edition, using the UNIVARIATE, GLM, and NPAR1WAY homogeneous (p = 0.2761). An ANOVA was then performed and a
procedures, as well as a Dunn’s test macro (Dmitrienko et al., statistically significant difference between the means was found
2007). A cut-off value 0.05 for p was adopted in the statistical tests. (F ffi 7.99; p = 0.0005). The Tukey’s studentized range (HSD) post-
hoc test was also performed (with a = 0.05), indicating that ABC
5. Computational results and CS were the best algorithms for the IAEA-3D LPO problem,
with no statistically significant difference between them (see
The results of the CS are compared to the results of the algo- Table 4). Both ABC and CS outperformed PBIL.
rithms ABC (Karaboga, 2005) and PBIL (Baluja, 1994) obtained in The Kruskal-Wallis test (v2 ffi 16.07; p = 0.0003) and the Dunn’s
previous work (Meneses et al., 2018). ABC and PBIL were selected test (see Table 5) were also performed. Both tests corroborate the
because they were the best algorithms for the instances IAEA-3D, ANOVA’s results.
BIBLIS-2D, and Angra 1 – 7th cycle for 100,000 evaluations. The
best LPs found by CS for each problem are shown in Fig. 4. The 5.2. BIBLIS-2D LPO problem
CS average running times for each instance were: (a) 29.94 min
(0.50 h) for IAEA-3D; (b) 272.27 min (4.54 h) for BIBLIS-2D; and The descriptive statistics of the CB results for the BIBLIS-2D LPO
(c) 212.82 min (3.55 h) for Angra 1
7th cycle. problem is presented in Table 6. In Fig. 6 the performance of the CS,
ABC, and PBIL algorithms are compared over the evaluations.
5.1. IAEA-3D LPO problem The final CB results of the ABC, PBIL, and CS algorithms were
non-normal (Shapiro-Wilk test, respectively with p < 0.0001,
The descriptive statistics of the CB results for the IAEA-3D LPO p = 0.0008, and p < 0.0001) and a Kruskal-Wallis test was then per-
problem is presented in Table 3. In Fig. 5 the performance of the formed indicating a statistically significant difference between the
CS, ABC, and PBIL algorithms are compared over the evaluations. mean ranks (v2 = 17.3; p = 0.0002). The Dunn’s post-hoc test for

Fig. 4. Best LPs found by CS for: (a) IAEA-3D LPO problem (CB = 546.88; Fq = 2.193; FAs with control rods, with number 4, are not swapped; number 1 represents reflector); (b)
BIBLIS-2D LPO problem (CB = 563.49; Fq = 1.382; number 3 represents reflector); (c) Angra 1 PWR
7th cycle (CB = 1406; Prm = 1.395; the central FA is not swapped).
 A.A.M. Meneses et al. / Annals of Nuclear Energy 139 (2020) 107214 7

Table 3 Table 6
Descriptive statistics of CB results (in ppm) for the IAEA-3D LPO problem. Descriptive statistics of CB results (in ppm) for the BIBLIS-2D LPO problem.

ABCa PBILa CSb ABCa PBILa CSb

Average 511.89 502.79 511.92 Average 472.26 524.67 528.44
Standard Dev. 11.56 14.79 12.95 Standard Dev. 121.85 24.16 37.83
Median 511.23 500.81 510.00 Median 515.22 525.69 538.70
Maximum 546.88 543.14 546.88 Maximum 559.70 558.35 563.49
Minimum 485.80 474.09 488.59 Minimum 81.85 439.56 393.60
a a
Meneses et al. (2018). Meneses et al. (2018).
b b
Present work. Present work.

Fig. 5. Comparison between PBIL, ABC and CS for the IAEA-3D LPO problem. The Fig. 6. Comparison between PBIL, ABC, and CS for the BIBLIS-2D LPO problem. The
bars represent a 95% confidence interval for the average (considering the Student’s bars represent a 95% confidence interval for the averages (considering the Student’s
t-distribution for 50 tests). t-distribution for 50 tests).

Table 4 The final CB results of the ABC and PBIL algorithms were normal
Tukey’s Studentized range (HSD) post hoc test results for the IAEA-3D LPO problem (Shapiro-Wilk test, respectively with p-values 0.0893 and 0.2687),
(a = 0.05).
whereas the final results of the CS algorithm were non-normal
Comparison number Group comparison Significant difference = ** (p = 0.0004). Thus, a Kruskal-Wallis test was performed indicating
1 ABC-CS a statistically significant difference between the mean ranks
2 ABC-PBIL ** (v2 = 7.08; p = 0.0290). The Dunn’s post-hoc test for multiple com-
3 CS-PBIL ** parisons was also performed, indicating that CS and ABC were the
**Indicates a statistically significant difference between algorithms. best algorithms for the Angra 1
7th Cycle LPO problem, with no
statistically significant difference between them (see Table 9). CS
outperformed PBIL, although ABC did not outperformed PBIL.
Table 5
Dunn’s post hoc test results for the IAEA-3D LPO problem (a = 0.05).
6. Discussion
Comparison Group Difference in Cutoff at Significant
number comparison average ranks alpha = 0.05 difference = **
As mentioned by Joyce and Herrmann (2018), notwithstanding
1 ABC-CS 1.84 20.8015 the NFL theorems, and as computational experiments point out, for
2 ABC-PBIL 31.04 20.8015 **
specific sets of problems there are algorithms that on average per-
3 CS-PBIL 29.20 20.8015 **
form better than others because of also specific characteristics. For
**Indicates a statistically significant difference between algorithms. the LPO problem some remarkable results have been obtained by
ABC algorithm (de Oliveira and Schirru, 2011; Safarzadeh et al.,
2011; Meneses et al., 2018), as well as by PBIL algorithm (Caldas
multiple comparisons was also performed, indicating that CS and and Schirru, 2008; Meneses et al., 2018).
PBIL were the best algorithms for the BIBLIS-2D LPO problem, with Yarizadeh-Beneh et al. (2016) showed the efficiency of the CS
no statistically significant difference between them (see Table 7). algorithm for a WWER NPP, and in the present work we applied
Both CS and PBIL outperformed ABC. CS to the LPO based on the benchmarks IAEA-3D and BIBLIS-2D,
as well to the Angra 1
7th Cycle problem, and in all of the three
5.3. Angra 1
7th cycle LPO problem cases statistical analyses showed that CS is among the best
algorithms.
The descriptive statistics of the CB results for the Angra 1
7th Yang (2014) discusses an interesting characteristic of the CS
Cycle LPO problem is presented in Table 8. In Fig. 7 the perfor- showing its similarities to Differential Evolution (Storn and Price,
mance of the CS, ABC, and PBIL algorithms are compared over 1997), PSO, and SA in the sense that CS can be seen as an efficient
the evaluations. combination of those algorithms. Yang (2014) also reports that the
8 A.A.M. Meneses et al. / Annals of Nuclear Energy 139 (2020) 107214

Table 7 And such characteristics of the CS algorithm for global opti-

Dunn’s post hoc test results for the BIBLIS-2D LPO problem (a = 0.05). mization are essential for the search in the LPO as real vectors
Comparison Group Difference in Cutoff at Significant are mapped into combinatorial solutions with the RK. The RK
number comparison average ranks alpha = 0.05 difference = ** approach for efficient global optimizers in continuous spaces have
1 ABC-CS 35.98 20.8015 ** enabled good performances for several algorithms. And as new glo-
2 ABC-PBIL 21.08 20.8015 ** bal optimizers are developed with efficient search in continuous
3 CS-PBIL 14.90 20.8015 spaces, notably the application of RKs makes them also efficient
**Indicates a statistically significant difference between algorithms. in the LPO problem.

7. Conclusion
Table 8
Descriptive statistics of CB results (in ppm) for the Angra 1
7th Cycle LPO problem.
In the present work the CS algorithm for the optimization of the
ABCa PBILa CSb IAEA-3D, BIBLIS-2D, and Angra 1
7th Cycle LPO problems
Average 1307 1303 1324 (Meneses et al., 2018) was implemented. The algorithms selected
Standard Dev. 48 45 61 for comparison were ABC and PBIL.
Median 1311 1304 1328
Statistical analyses showed that: (a) for the IAEA-3D problem,
Maximum 1435 1400 1406
Minimum 1224 1209 1073 CS and ABC were the best algorithms; (b) for the BIBLIS-2D LPO
problem, CS and PBIL were the best algorithms; and (c) for the
a
Meneses et al. (2018).
b
Angra 1
7th Cycle LPO problem, CS and ABC were the best algo-
Present work.
rithms. Thus, in all of the three LPO problems tested, CS was among
the best algorithms, reaching the best overall performance, there-
fore being the most robust of the three algorithms compared.
In future works we will include other instances for algorithms
comparison, implement the Improved DCS for the LPO, and opti-
mize multi-cycle problems.

CRediT authorship contribution statement

Anderson Alvarenga de Moura Meneses: Conceptualization,

Methodology, Formal analysis, Writing - original draft. Patrick
Vasconcelos da Silva: Software, Validation, Investigation, Writing
- original draft. Fernando Nogueira Nast: Software, Investigation,
Visualization. Lenilson Moreira Araujo: Formal analysis, Writing -
original draft. Roberto Schirru: Conceptualization, Methodology,
Writing - review & editing, Supervision.

Declaration of Competing Interest

The authors declare that they have no known competing finan-

Fig. 7. Comparison between PBIL, ABC and CS for the Angra 1
7th cycle LPO cial interests or personal relationships that could have appeared
problem. The bars represent a 95% confidence interval for the averages (considering to influence the work reported in this paper.
the Student’s t-distribution for 50 tests).

Acknowledgements

Table 9 R.S. acknowledges CNPq (Conselho Nacional de Desenvolvimento

Dunn’s post hoc test results for the Angra 1
7th Cycle LPO problem (a = 0.05). Científico e Tecnológico, Brazil) and FAPERJ (Fundação Carlos Cha-
Comparison Group Difference in Cutoff at Significant gas Filho de Amparo à Pesquisa do Estado do Rio de Janeiro, Brazil).
number comparison average ranks alpha = 0.05 difference = ** The authors acknowledge CNEN (Comissão Nacional de Energia
1 ABC-CS 18.66 20.8015 Nuclear – participant of the CAMP agreement with the USA NRC),
2 ABC-PBIL 2.49 20.8015 in the terms of the agreement of mutual cooperation with the Fed-
3 CS-PBIL 21.15 20.8015 ** eral University of Western Pará for using the PARCS v3.0 code. The
**Indicates a statistically significant difference between algorithms.
authors would like to thank the reviewers for their valuable com-
ments and suggestions.

CS algorithm’s global convergence is guaranteed, which is corrob- References

orated by He et al. (2018), with a validation with numerical Argonne National Laboratory, 1977. Benchmark Problem Book (Supplement 2).
analysis. Technical Report ANL-7416.
According to Yang (2014), the efficiency of the CS algorithm is Baluja, S., 1994. Population-Based Incremental Learning: A Method for Integrating
Genetic Search Based Function Optimization and Competitive Learning.
due to the balance of global and local search processes, which
Technical Report. School of Computer Science, Carnegie Mellon University,
are controlled by the parameter pa . The typical value pa = 0.25 also Pittsburgh, Pennsylvania.
used in our tests means that such control enables the global search Bean, J.C., 1994. Genetic algorithms and random keys for sequencing and
to be performed in approximately 3/4 of the search. This fact as optimization. ORSA J. Comput. 6, 154–160. https://doi.org/10.1287/ijoc.6.2.154.
Brown, C.T., Liebovitch, L.S., Glendon, R., 2007. Lévy flights in Dobe Ju/’hoansi
well as the usage of Lévy flights makes the CS a remarkable global foraging patterns. Hum. Ecol. 35, 129–138. https://doi.org/10.1007/s10745-
optimizer. 006-9083-4.
 A.A.M. Meneses et al. / Annals of Nuclear Energy 139 (2020) 107214 9

Caldas, G.H.F., Schirru, R., 2008. Parameterless evolutionary algorithm applied to the Meneses, A.A.M., Araujo, L.M., Nast, F.N., da Silva, P.V., Schirru, R., 2018. Application
nuclear reload problem. Ann. Nucl. Energy 35, 583–590. https://doi.org/ of metaheuristics to Loading Pattern Optimization problems based on the IAEA-
10.1016/j.anucene.2007.08.014. 3D and BIBLIS-2D data. Ann. Nucl. Energy 111, 329–339. https://doi.org/
Chapot, J.L.C., 2000. Otimização automática de recargas de reatores a água 10.1016/j.anucene.2017.09.008.
pressurizada utilizando algoritmos genéticos (D.Sc. thesis) (in Portuguese). Meneses, A.A.M., Gambardella, L.M., Schirru, R., 2010. A new approach for
COPPE/UFRJ, Brazil. heuristics-guided search in the In-Core Fuel Management Optimization. Prog.
Chapot, J.L.C., Da Silva, F.C., Schirru, R., 1999. A new approach to the use of genetic Nucl. Energy 52, 339–351. https://doi.org/10.1016/j.pnucene.2009.07.007.
algorithms to solve the pressurized water reactor’s fuel management Meneses, A.A.M., Machado, M.D., Schirru, R., 2009. Particle swarm optimization
optimization problem. Ann. Nucl. Energy 26, 641–655. https://doi.org/ applied to the nuclear reload problem of a pressurized water reactor. Prog. Nucl.
10.1016/S0306-4549(98)00078-4. Energy 51, 319–326. https://doi.org/10.1016/j.pnucene.2008.07.002.
Chen, S.-M., Chien, C.-Y., 2011. Solving the traveling salesman problem based on the Meneses, A.A.M., Schirru, R., 2015. A cross-entropy method applied to the in-core
genetic simulated annealing ant colony system with particle swarm fuel management optimization of a pressurized water reactor. Prog. Nucl.
optimization techniques. Expert Syst. Appl. 38, 14439–14450. https://doi.org/ Energy 83, 326–335. https://doi.org/10.1016/j.pnucene.2015.04.007.
10.1016/j.eswa.2011.04.163. de Oliveira, I.M.S., Schirru, R., 2011. Swarm intelligence of artificial bees applied to
de Lima, A.M.M., 2005. Recarga de reatores nucleares utilizando redes conectivas de in-core fuel management optimization. Ann. Nucl. Energy 38, 1039–1045.
colônias artificiais. COPPE/UFRJ, Brazil (D.Sc. Thesis) (in Portuguese). https://doi.org/10.1016/j.anucene.2011.01.009.
de Lima, A.M.M., Schirru, R., da Silva, F.C., Medeiros, J.A.C.C., 2008. A nuclear reactor core Oka, Y., Uchikawa, S., Suzuki, K., 2014. Light water reactor design. In: Oka, Y. (Ed.),
fuel reload optimization using artificial ant colony connective networks. Ann. Nucl. Nuclear Reactor Design. Springer. https://doi.org/10.1007/978-4-431-54898-
Energy 35, 1606–1612. https://doi.org/10.1016/j.anucene.2008.03.002. 0_3.
Demšar, J., 2006. Statistical comparisons of classifiers over multiple data sets. J. Ouaarab, A., Ahiod, B., Yang, X.-S., 2014. Improved and discrete cuckoo search for
Mach. Learn. Res. 7, 1–30. solving the travelling salesman problem. In: Yang, X.-.S. (Ed.), Cuckoo Search
Dmitrienko, A., Chuang-Stein, C., D’Agostino, R.B., 2007. Pharmaceutical Statistics and Firefly Algorithm. Springer, pp. 63–84.
Using SAS: A Practical Guide. SAS Institute. Papadimitriou, C., Steiglitz, K., 1982. Combinatorial Optimization: Algorithms and
Downar, T., Xu, Y., Seker, V., 2009a. PARCS v3. 0 US NRC Core Neutronics Simulator Optimization. Prentice-Hall.
Theory Manual. University of Michigan. Ann Arbor, MI. Parks, G.T., 1990. An intelligent stochastic optimization routine for nuclear fuel
Downar, T., Xu, Y., Seker, V., 2009b. PARCS v3. 0 US NRC Core Neutronics Simulator cycle design. Nucl. Technol. 89, 233–246. https://doi.org/10.13182/NT90-
User Manual. University of Michigan. Ann Arbor, MI. A34350.
Engelbrecht, A.P., 2007. Computational Intelligence: An Introduction. John Wiley Payne, R.B., 2005. The Cuckoos. Oxford University Press.
and Sons Inc.. Poon, P.W., Parks, G.T., 1992. Optimizing PWR reload core designs. In: Manner, R.,
Finnemann, H., Bennewitz, F., Wagner, M.R., 1977. Interface current techniques for Manderick, B. (Eds.), Parallel Problems Solving from Nature II, pp. 371–380.
multidimensional reactor calculations. Atomkernenergie 30, 123–128. Poursalehi, N., Zolfaghari, A., Minuchehr, A., 2013. Development of a high order and
García, S., Molina, D., Lozano, M., Herrera, F., 2009. A study on the use of non- multi-dimensional nodal code, ACNEC3D, for reactor core analysis. Ann. Nucl.
parametric tests for analyzing evolutionary algorithms’ behaviour: a case study Energy 55, 211–224. https://doi.org/10.1016/j.anucene.2012.10.022.
on the CEC’2005 Special Session on Real Parameter Optimization. J Heuristics Rutherford, A., 2011. ANOVA and ANCOVA: A GLM Approach. 2nd Ed. John Wiley
15, 617–644. https://doi.org/10.1007/s10732-008-9080-4. and Sons, Inc.
García, S., Fernández, A., Luengo, J., Herrera, F., 2010. Advanced nonparametric tests Safarzadeh, O., Zolfaghari, A., Norouzi, A., Minuchehr, H., 2011. Loading pattern
for multiple comparisons in the design of experiments in computational optimization of PWR reactors using Artificial Bee Colony. Ann. Nucl. Energy 38,
intelligence and data mining: experimental analysis of power. Inform. Sci. 180, 2218–2226. https://doi.org/10.1016/j.anucene.2011.06.008.
2044–2064. https://doi.org/10.1016/j.ins.2009.12.010. Shlesinger, M.F., Zaslavsky, G.M., Frisch, U., 1995. Lévy Flights and Related Topics in
Gutowski, M., 2002. Lévy flights as an underlying mechanism for global Physics. Springer-Verlag.
optimization. Prace Naukowe Politechniki Warszawskiej. Elektronika 139, Sheskin, D.J., 2003. Handbook of Parametric and Nonparametric Statistical
175–184. Procedures. Chapman & Hall/CRC.
He, X.-S., Wang, F., Wang, Y., Yang, X.-S., 2018. Global convergence analysis of Silva, P.V., Nast, F.N., Schirru, R., Meneses, A.A.M. Busca Cuco com Voo de Lévy
cuckoo search using markov theory. In: Yang, X.-.S. (Ed.), Nature-Inspired Aplicado ao Problema de Recarga de Combustíveis em Reatores Nucleares. In:
Algorithms and Applied Optimization. Springer, pp. 53–67. https://doi.org/ Proceedings of the International Nuclear Atlantic Conference 2017, Belo
10.1007/978-3-319-67669-2_3. Horizonte, Brazil, October 22–27, 2017 (in Portuguese).
Hill, N.J., Parks, G.T., 2015. Pressurized water reactor in-core nuclear fuel Smith, K.S., 1979. An Analytic Nodal Method for Solving the Two-group,
management by tabu search. Ann. Nucl. Energy 75, 64–71. https://doi.org/ Multidimensional, Static and Transient Neutron Diffusion Equations.
10.1016/j.anucene.2014.07.051. Massachusetts Institute of Technology (Ph.D. thesis).
Johnson, D.S., McGeoch, L.A., 1997. The traveling salesman problem: a case study. Stevens, J.G., Smith, K.S., Rempe, K.R., Downar, T.J., 1995. Optimization of
In: Aarts, E., Lenstra, J.K. (Eds.), Local Search in Combinatorial Optimization. pressurized water reactor shuffling by simulated annealing with heuristics.
John Wiley and Sons Ltd., pp. 215–310. Nucl. Sci. Eng. 121, 67–88. https://doi.org/10.13182/NSE121-67.
Joo, H.G., Barber, D., Jiang, G., Downar, T., 1998. PARCS: a multi-dimensional two- Storn, R., Price, K., 1997. Differential evolution – a simple and efficient heuristic for
group reactor kinetics code based on the nonlinear analytic nodal method, PU/ global optimization over continuous spaces. J. Global Optim. 11, 341–359.
NE-98-26. https://doi.org/10.1023/A:1008202821328.
Joyce, T., Herrmann, J.M., 2018. A review of no free lunch theorems and their Wolpert, D.H., Macready, W.G., 1997. No free lunch theorems for optimization. IEEE
implication for metaheuristic optimisation. In: Yang, X.-.S. (Ed.), Nature- T. Evolut. Comput. 1, 67–82. https://doi.org/10.1109/4235.585893.
Inspired Algorithms and Applied Optimization. Springer, pp. 27–51. https:// Yang, X.-S., 2010. Nature-inspired Metaheuristic Algorithms. Luniver Press.
doi.org/10.1007/978-3-319-67669-2_2. Yang, X.-S., Deb, S., 2009. Cuckoo search via Lévy flights. In: World Congress on
Karaboga, D., 2005. An idea based on honey bee swarm for numerical optimization. Nature & Biologically Inspired Computing. IEEE, pp. 210–214. https://doi.org/
TR 06, October. Erciyes University, Engineering Faculty, Computer Engineering 10.1109/NABIC.2009.5393690.
Department, Kayseri/Turkiye. Yang, X.-S., Deb, S., 2010. Engineering optimization by cuckoo search. Int. J. Math.
Levine, S., 1986. In-core fuel management of four reactor types. In: Ronen, Y. (Ed.), Model. Numer. Optim. 1, 330–343. https://doi.org/10.1504/
Handbook of Nuclear Reactors Calculations, vol. II. CRC Press. IJMMNO.2010.03543.
Lin, C., Yang, J.-I., Lin, K.-J., Wang, Z.-D., 1998. Pressurized water reactor loading Yang, X.-S., Deb, S., 2013. Multiobjective cuckoo search for design optimization.
pattern design using the simple tabu search. Nucl. Sci. Eng. 129, 61–71. https:// Comput. Oper. Res. 40, 1616–1624. https://doi.org/10.1016/j.cor.2011.09.026.
doi.org/10.13182/NSE98-A1963. Yang, X.-S., 2014. Cuckoo search and firefly algorithm: overview and analysis. In:
Machado, L., Schirru, R., 2002. The Ant-Q algorithm applied to the nuclear reload Yang, X.-S. (Ed.), Cuckoo Search and Firefly Algorithm: Theory and Applications.
problem. Ann. Nucl. Energy 29, 1455–1470. https://doi.org/10.1016/S0306- Springer, pp. 1–26. https://doi.org/10.1007/978-3-319-02141-6_1.
4549(01)00118-9. Yarizadeh-Beneh, M., Mazaheri-Beni, H., Poursalehi, N., 2016. Improving the
Machado, M.D., 2005. Algoritmo evolucionário PBIL multi-objetivo aplicado ao refueling cycle of a WWER-1000 using cuckoo search method and thermal-
problema da recarga de reatores nucleares. COPPE/UFRJ, Brazil (D.Sc. thesis) (in neutronic coupling of PARCS v2. 7, COBRA-EN and WIMSD-5B codes. Nucl. Eng.
Portuguese). Des. 310, 247–257. https://doi.org/10.1016/j.nucengdes.2016.10.006.
Mantegna, R.N., 1994. Fast, accurate algorithm for numerical simulation of Lévy
stable stochastic processes. Phys. Rev. E 49, 4677–4683. https://doi.org/
10.1103/PhysRevE.49.4677.