2nd International Conference on Engineering Technology and their Applications 2019-IICET2019- Islamic University,

Alnajaf-Iraq

Effect of Loading on Optimum Weight of Planer

Trusses Using Genetic Algorithms
Farid H. Arna'ot
Computer Technical Engineering Department,
College of Technical Engineering, the Islamic University
Najaf, Iraq
[email protected]

Abstract ‒ Selection a proper layout of a structure can be modified, and can treat discrete variables. However, GA,
considered as a challenge decision that facing the structural similar to other algorithms, stuck in local optimums and
designer, this layout should withstand the applied loads based on requires many function evaluations, which may lead to quick
different provisions, such as the allowable stress, the convergence with similar individuals. In 1975, Holland
serviceability limit state and etc. The challenges will be increases
introduced GA in his book “Adaptation in natural and
when the cost and the duration of the construction are of vital
importance. artificial systems” [2]. He proposed the GAs as a heuristic
This paper demonstrates the use of an optimization tool based method based on “Survival of the fittest”. The solution in his
on a genetic algorithm in order to examine the effect of different technique mimics the nature. The technique initially created a
loading values on a trusses response. Three design categories had random solution then represented the development principles
been adopted in the evaluation; the size, the shape, and the of the best solutions through many generations until the
topology that satisfies the relevant constraints given in the AISC- optimal solution converges.
ASD-89 standard. For this reason, a traditional program under Whereas the structural design of a truss is resolute in
the name of GS-USA frame was employed. The program is selection the topology, an optimization task focuses on
judged based on the analysis of a benchmark truss, and by
minimizing the cost of the construction through optimizing
comparing the optimized a three well known trusses found in the
literature that adopted different optimization techniques. The the truss weight. Controlling the weight means decreases the
evaluation results showed that the genetic algorithm introduced material usage and decreases the self-weight.
a good optimization technique compared to other techniques. GA technique is adopting the topology, size, and shape of
Although the GS-USA program was efficient in optimizing the the truss at the same time. In spite of, the shape optimization
weight, and the tensile stresses, however, the program stucked in (optimal position of nodes) shows a good effect on the truss
compression stress. optimization, the topology and the size (the former seeking to
Analysis results showed that the weight, volume, and the find good nodes’ connectivity, the latter seeking to find
members' stresses can be efficiently optimized for the suggested optimum member size) are the best options in weight
planer truss compared to Howe and Baltimore Trusses of the
optimization [3]. Chan and Wong [3] examined planar frame
same span and bays, where the weight increments were slightly
with increasing the load, and 57% of materials were saved, for 40 stories of three bays using optimality criteria design
moreover the truss members were effectively stressed up to 74% method, they found that the topology optimization is, in
of the allowable stress. general, more significant than the sizing one. It saves about
35% of the materials compared with sizing optimization only.
Keywords‒Trusses, optimization, Genetic algorithms, GS-USA From economical prospective, topology optimization is the
most rewarding technique in structural optimization [4,5].
I. INTRODUCTION Many attempts had been adopted the GAs in order to
Optimization is that branch of engineering science that optimize a planers trusses. Dep [4] examined the power of
searches for optimal solutions under given circumstances. GAs and its mechanism through a small fraction (one in over
Optimization of the structural members seems to have a deep 800 million) of the search space of a welded joint beams, he
history. In 1638, Galileo Galilee studied the earliest found that using GA with reproduction, crossover, and
optimization problems, he designed a cantilever beam with mutation operators is able to converge the design parameters
low weight, high bending resistance, and constant shear [1]. in a fast manner. Rajeev and Krishnamoorthy [6] agreed that
This branch of engineering had been studied in details during the GAs is the best technique to improve the structural design.
the last few decades, supported by the availability of cheap The researchers used the GAs for optimizing the weight of
and powerful personal computers. There are many methods trusses using discrete design variables. Almost all the design
were proposed to obtain the optimal results in engineering variables in this study were discrete, that make the GAs is the
such as Genetic Algorithms (GAs), simulated annealing, best choices from mathematical point of view. The study
particle swarm optimization, ant colony optimization, and concluded that the analysis process of GAs is slower than the
Fuzzy optimization. Among these methods a GA is broadly other traditional optimization technique; this is may be due to
applied because of its simplex application, can be easily high number of functions to be solved that leads to increase

978-1-7281-4105-3/19/$31.00 ©2019 IEEE 37

Authorized licensed use limited to: Western Sydney University. Downloaded on July 25,2020 at 19:06:17 UTC from IEEE Xplore. Restrictions apply.
 2nd International Conference on Engineering Technology and their Applications 2019 -IICET2019- Islamic University,
Alnajaf-Iraq
the number of analysis in addition to the discrete variables. span, 4.5m height, divided into 6 equal bays of 4 m length for
Lin and Hajela [7] used continuous variables, and analyzed each. Four different loads were applied to the truss at each
10, 25, 72-bars trusses. The study found that GAs required node in the lower chord as shown in Fig. 2a. The loads were,
less computational effort when the discrete variables are 50, 100, 200, 300 kN.
large. Detailed information on the development of the GA has Baltimore truss is analyzed as the 2nd layout; this truss
been address in Stople critical review [8]. consists of forty-nine members, of the same span, height, and
In the present work, the weight, displacement, tensile and the number of bays of Howe truss model. The main
compressive stresses were evaluated for three types of planer differences between the two layouts is that the lower bays
trusses that subjected to different load values. The same bays were extra braced by short vertical members, and also the two
number and span were used in the evaluation process. A GAs edge bays had more bracing than the Howe truss, as shown in
program under a traditional name GS-USA frame was used, Fig 2a and 2b. The Baltimore truss was loaded with same
the program was designed by Rajan, (1991-2001) [9]. In this loading cases as in Howe truss.
program, the design variant of cross-sectional properties,
nodal data, element data, nodal displacement, and modal data
were optionally constraints variables. Those variables are
compliant with the provisions of AISC-ASD [10] for the
stress, displacement, Euler buckling, cross-sectional
dimensions and nodal locations. The objective function was
weight, mass, cost, displacement, support reaction).

II. ANALYSIS CATEGORIES

(a) Howe Truss layout
A. Analysis Procedure
The present study includes two main stages; the program is
assessing at the first stage, and then applied in the second
stage. In the assessment stage, fifteen-bar truss that tagged as
benchmark truss were analyzed using the GS-USA program.
The results of weight, design variables and constraints were
compared to the available data from the literature.
At the next stage, the program was applied for optimizing
three well-known trusses (Howe, Baltimore, and Arch (b) Baltimore truss
trusses). In order to overcome the effect of curvature in the
arch truss, the truss were modified utilize the principles of
simple geometry shape (triangle), keeping the same functional
dimensions (length, number of bays and the span) and
members sections.
B. Trusses Geometry and Sections
Figure 1 shows the geometry of the benchmark truss. The (c) Arch truss
fifteen-bar truss was used as a benchmark in the structural
optimization problems, where the position of the nodes is
rarely altered. The truss is supported by two vertical supports
with a distance of 9.144 meters and two loads of 445.374 kN
located at the end and mid span distance 9.144 and 18.288
meters from the lower support as shown in Fig. 1.

(d) Suggested modification for Arch truss

Fig. 2 :Trusses layout

The configuration layout of the Arch truss shown in Fig.

2c. The curved members were straightened in order to get the
Fig. 1 :Benchmark truss layout best restraint to the whole truss. For comparison purpose, the
suggested truss has the same length, divided into same bays,
As mentioned previously, three trusses were analyzed.
with 6 m height; the height of the truss is not affecting the
Howe truss that consists of twenty-seven members, 24 m clear

83

Authorized licensed use limited to: Western Sydney University. Downloaded on July 25,2020 at 19:06:17 UTC from IEEE Xplore. Restrictions apply.
 2nd International Conference on Engineering Technology and their Applications 2019 -IICET2019- Islamic University,
Alnajaf-Iraq
function of truss, because the loads were applied to the upper E. Design Codes and Limitations
chord of the truss as shown in Fig. 2d. The constraints variables gi consists of two parameters yi
C. Materials' Mechanical Properties and ya, the former can be found by analyze the structure using
stiffness matrix method. ya are the allowable limits of the
Same materials were considered for all trusses used in the response. yi should be less than ya. The AISC had developed
analyses. No provision on the limits of highest and lowest design specifications for structural steel with two different
values. The properties of materials used as follows; design approaches: “Specification for Structural Steel
Young’s modulus (E) = 200 x 109 N/m2. Buildings ̶ Allowable Stress Design (ASD) and Plastic
Mass density (ρ) = 7850 kg/m3. Design” and “Load and Resistance Factor Design (LRFD)
Coefficient of thermal expansion (α) = 12 x 10-6 m/m.oC. Specification for Structural Steel Buildings”. The GS-USA
Yield stress (σy) = 250 x 106 N/m2. program designed to follow the limitations of the allowable
Allowable tensile strength (fy) = 410 x 106 N/m2. stress design specification (ASD). The constraints imposed on
D. Optimization Parameters the structure according to AISC-ASD 89 are:
The current paper study the optimization of the structural 1) Tensile Stress; the allowable tensile stresses are 0.60Fy
member lower weight f(x) that have different cross-sectional on the gross area and 0.50Fu on the effective net area, where
areas (xi) under the limitation of the specification of American Fy and Fu are specified minimum yield and tensile strength
Institute of Steel Construction (AISC-ASD 89) [10] (σa and respectively.
δ). i.e. find the xi where,
2) Compressive Stress; the allowable compression stress
xi = {Ft, Wt, Fw, Wh, Rs, Rh, Ct}
depends on whether buckling will be elastic or inelastic. This
where the parameters in the curly bracket represent the
dependents on the slenderness ratio KL/r, where K is the
dimensions of the cross sections as a string of the single
effective length factor, L is the member length and r is the
chromosome xi used in the examples, as shown in Fig. 3. It is
radius of gyration.
worth to mention here that the member cross-sections were
used in the same configuration for all trusses, where, the When KL/r < Cc, the allowable compression stress Fa (kips)
symmetrical I-section used for upper and lower chord, as on the gross section should be computed from
shown in Fig. 3a. The circular solid section was used for 1 - (KL/r)2 / 2C2c
vertical members, as shown in Fig. 3b. The hollow circular Fa = × Fy (3)
section was used for inclined bracing members Fig. 3c. The KL 2 KL 3
5 3( r ) ( )
r
weight of the truss f (x) was functioned, as, f(x) = Weight of + ̶
3 8Cc 8C3c
truss = Density × Volume
n where
Minimize f(x) = ∑ ρi × A × li (1) 2 × π2 × E
i
i=1 Cc =√ (4)
Fy
where: ρ is the steel density, A is the cross-sectional area of
the different type of members, and l is the length of the
member. When KL/r > Cc, the allowable compression stress Fa is;
The objective function is obedient to the constraint 12π2 × E
variables gi, where, Fa = (5)
KL 2
|y | 23 ( )
r
gi = i ̶ 1 ≤ 0 (2)
ya 3) Vertical Deflections; the Specification limits the
The subscripts i and a, represent the actual and the deflection to 1⁄360 of the span L for simply supported beam
allowable response of element (member or node), and y is the and 1/150 of the span L for a cantilever.
response form, that may take different forms, such as stress σ,
displacement δ, and Euler local buckling (pcr), hence, III. RESULTS AND DISCUSSION
y = {σ, δ, pcr}
A. Application of the GS-USA
GS-USA program has been applied with optimizing the
size, the shape and the topology simultaneously; most
commonly, the topology of the truss (i.e. the inner
connectivity of the members) was fixed. Therefore, here we
are trying to use different topology to the trusses.
Study of optimizing the sizes of the Howe, Baltimore, and
Arch s trusses under different applied loads was performed,
the results show:
Fig. 3 :dimensions of sections used in the current work
1) Benchmark Truss; Table I summarized the results
optimized parameters. The optimizing results were compared

83

Authorized licensed use limited to: Western Sydney University. Downloaded on July 25,2020 at 19:06:17 UTC from IEEE Xplore. Restrictions apply.
 2nd International Conference on Engineering Technology and their Applications 2019 -IICET2019- Islamic University,
Alnajaf-Iraq
to the collected data found in [11], these parameters are
presented in Table II. The most common material used in
2) Howe Truss; summary of optimization results are
these data was aluminum alloy with E = 68.95 GPa, ρ = 2768
introduced in Table III. The results show that in case of
kg/m3 and the element stresses were limited to 172.37 MPa in
duplicating the load magnitudes the mass of truss increased
both tension and compression. The buckling was ignored. The
about 105%, 94%, in turns, increasing the load from 200kN to
displacements were limited to 50.8 mm both horizontally and
300kN (50% of load magnitude) the weight was increased by
vertically, knowing that their algorithms were compliant with
55%. In case of increasing the truss weight shows uniformly
Eurocode 3 (EN 1993-1-1).
proportional to the load increments.
TABLE I. SUMMARY OF RESULTS FOR THE
BENCHMARK TRUSS
TABLE III. SUMMARY OF RESULT OF OPTIMIZED HOWE
11% < heaver truss TRUSS
Total mass 3277 kg &
Cases of loading
32% > than the lighter truss Constraints
50kN 100kN 200kN 300kN
Total weight 32.15 kN -
Total Mass (kg) 2081 4277 8304 12909
Total volume 1.184 m3 -
Total weight (kN) 20.42 41.96 81.46 126.64
51‡ % of limited
δmax -62.042mm Total volume m3 0.265 0.544 1.057 1.644
displacement
δmax (mm) -14.3 -12.2 -14 -11.8
σc,max 69.99×106 N/m2 28† % of limited stress
% to allowable disp. 21.4% 18.3% 21% 17.7%
σt,max 77.9×106 N/m2 64† % of limited stress
σc,max 70×106 70×106 70×106 70×106
pcr* 193 ×103 N -
‡ % to allowable Stress 28.4% 28.4% 28.4% 28.4%
Max. deflection allowed = l/150 =121
† σt,max (N/m2) 204×106 118×106 189×106 95 ×106
allowable stress = 0.6 Fy = 246
π2 × E × I
% to allowable Stress 83% 48% 77% 39%
⋆ Min. Euler buckling = , where K = 1 in case of pinned- pcr (kN) 130 222 60.9 540
K2 L2
pinned support

TABLE II. RESULTS COLLECTED IN REF. [11] 3) Baltimore Truss; the summarized results of this layout
Researcher Optimization Displacement Mass
are introduced in Table IV. The results show that in case of
technique (mm) (kg) duplicating the load magnitudes the mass of truss increasing
Deb [12] Genetic algorithm - 2222.2 in percentage of 99%, 100%, however, increasing the
Hajela and Lee
Genetic algorithm - 2242.0 magnitude from 200kN to 300kN (50% of load magnitude)
[13] the weight was increased 50%. The relationship between the
Li, Huang, and Particle swarm
Liu [14] optimizer
- 2295.6 load and mass shows a perfect proportional, but the weight of
Kripakaran, et Hybrid search the truss was very high compared with the previous truss.
- 2301.1
al., [15] method. 4) Suggested Arch Truss; The summarized results of the
Galante, [16] Genetic algorithm - 2322.1
modified Arch Truss layout are introduced in Table V. results
Hultman [11] Genetic algorithm 50.72 2327.1
Guerlement, et showed that in case of duplicating the load magnitudes the
Genetic algorithm 33.9 3695.0 mass of truss increasing in percentage of 155%, 57%,
al. [17]
however, increasing the magnitude from 200kN to 300kN
Comparison of the results of benchmark truss in Table I (50% of load magnitude) the weight was increased 52%. The
and previous data from Table II shows that GS-USA frame behavior of the program shows instability proportional
program presents acceptable results compared with heavier between loading and mass increasing, this case clearly can be
truss (about 11% less than the heaver truss and 32% more seen in tensile stress, where the element showed high stress in
than lighter truss), also the displacement was 51% of the 50kN loading compared with the stress under higher loading.
AISC-ASD-89 limitation but more than the displacement of
the other trusses. TABLE IV. SUMMARY OF RESULT OF OPTIMIZED
BALTIMORE TRUSS
Cases of loading
As mentioned before the material used in the listed Constraints
50kN 100kN 200kN 300kN
researches was of aluminum alloy, however, in the present Total Mass (kg) 3684 7336 14682 22027
study the material that was simulated was steel. Looking at Total weight (kN) 36.14 71.97 144 216
the material properties, steel has 3.05 times higher modulus of Total volume m3 0.469 0.934 1.870 2.806
δmax (mm) -8.2 -8.3 -8.2 -8.1
elasticity, and 2.84 times higher density compare to the % to allowable disp. 12.3% 12.3% 12.3% 12.3%
aluminum, giving it a small advantage in that aspect.
σc,max 70×106 70×106 70×106 70×106
Knowing that the in the previous studies the algorithms were % to allowable Stress 28.4% 28.4% 28.4% 28.4%
not checking the buckling (If buckling is not taken into
σt,max (N/m2) 133×106 151×106 126×106 107 ×106
account, the calculated trusses are not suitable for real
% to allowable Stress 54% 61% 51% 43%
constructions since they are likely to collapse), also all pcr (kN) 2.629 7.15 49 180
mentioned research were followed the Eurocode, whereas the
GS-USA frame designed to obey AISC provisions, which is
showing less conservative than the Eurocode.

04

Authorized licensed use limited to: Western Sydney University. Downloaded on July 25,2020 at 19:06:17 UTC from IEEE Xplore. Restrictions apply.
 2nd International Conference on Engineering Technology and their Applications 2019 -IICET2019- Islamic University,
Alnajaf-Iraq
Figure 4 introduces the relation between the applied loads
and the objective function (Mass). The relation can be
assessed in three stages; the first is the mass under low load,
where it’s clearly shows that there are no big differences in
the trusses weight. The differences developed when the loads
are duplicated, this is the second stage, in the former two
trusses the masses are increasing steadily, however, the
suggested truss shows 1.5 times the first truss weight, which
reduces its deference with the Howe truss, this is not
surprisingly, the organization of the elements in former two
trusses provide a good priority in weight distribution. The big
changes appear in the third stage, where all the trusses
weights were increased in the same percentage when the loads
Fig. 5 Relation between load and maximum tensile stress of optimized
were increased. trusses

TABLE V. SUMMARY OF RESULT OF OPTIMIZED SUGGESTED

TRUSS
Cases of loading
Constraints
50kN 100kN 200kN 300kN
Total Mass (kg) 1346 3441.38 5414.03 8228.1
Total weight (kN) 13.208 33.75 53.11 80.71
Total volume m3 0.171 0.438 0.689 1.048
δmax (mm) -12.3 -11.1 -11.1 -11.4
% to allowable disp. 18.4% 16.6% 16.6% 17.1%
σc,max 70×106 70×106 70×106 70×106
% to allowable Stress 28.4% 28.4% 28.4% 28.4%
σt,max (N/m2) 204×106 197×106 169×106 183 ×106
% to allowable Stress 83% 80% 69% 74%
pcr (kN) 11.7 111.7 180.0 540 Fig. 6 Relationship between load and maximum Displacement of optimized
trusses

The load- displacement relationship was investigated and

introduced in Fig. 6. In this figure Howe truss layout showed
good behavior related to displacement. The suggested truss
showed the same behavior, but in higher values. The
unexpected behavior developed in Baltimore truss where the
displacement changed harmonically with load incrementing.
For all trusses, the GS-USA stuck in the compressive
stress, this is an expected matter for all optimization methods.
B. Convergence History
All examined trusses show good convergence history, in all
loading values, where, the final solution was obtained in a
Fig. 4 Relation between load and mass of optimized trusses
smooth developed iteration. Table VI shows the maximum
number of iteration for each truss under the different load
Figure 5 illustrated the tensile stress-load relationship. The value. From this table, it is shown that the suggested truss got
figure shows that the tensile stress acted on the suggested the answer in convergent number, means that there was no
truss was closer to the allowable stress, where the element effect of varying load on smoothens of the result. The effect
stressed is 74% of the allowable stress. The Baltimore showed of loading amount clearly can be seen for the Howe Truss,
just 43% of allowable stress, and Howe truss showed 38%. In where, in case of increasing load value the iterations were
another word, the suggested Arch truss was the best in decreased, however, increasing the loading value were
configuring the tensile stress. increased the iterations for Baltimore Truss.

TABLE VI. NUMBER OF ITERATIONS

Iteration
Truss Applied load
50 kN 100 kN 200 kN 300 kN
Suggested 66 79 48 60
Howe 99 85 56 49
Baltimore 56 59 88 71

04

Authorized licensed use limited to: Western Sydney University. Downloaded on July 25,2020 at 19:06:17 UTC from IEEE Xplore. Restrictions apply.
 2nd International Conference on Engineering Technology and their Applications 2019 -IICET2019- Islamic University,
Alnajaf-Iraq
[15] P. Kripakaran, A. Gupta, J.R. Baugh, and W. John, "A novel
optimization approach for minimum cost design of trusses,"
IV. CONCLUSION Computers & Structures, vol. 85, pp. 1782‒1794, 2007.
[16] M. Galante, "Genetic algorithm as an approach to optimize real
Planer steel trusses were optimized, in order to investigate trusses," International Journal for Numerical Methods in Engineering,
the effect of truss topology on the weight and stress vol. 39, pp. 361‒382, 1996.
[17] G. Guerlement, R. Targowski, W. Gutkowski, J. Zawidzka and J.
distribution under different load values. From the calculations Zawidzki, "Discrete minimum weight design of steel structures using
and analyses conducted in this research, it can be concluded EC3 code," Structural and Multidisciplinary Optimization, vol. 22, pp.
that GAs is a good technique for optimizing the trusses that 322-327, 2001.
have discrete variables, however, like all optimization
methods, the current technique suffers from premature
convergence, and this appeared in compressive stress
analysis. Moreover, it was clear that the responses of the
trusses related to the applied load are based on the first
guesses of the section dimensions, and the topology. The most
important finding is that the truss with proper topology can
significantly distribute the stress, as shown that the GS-USA
frame program is able to optimize the element sizes of the
given trusses, giving material savings exceeding 57%,
moreover, the truss members stressed up to 74% of its
allowable stress.

REFERENCE
[1] I. Elishakoff, and M. Ohsaki, Optimization and Anti-optimization of
Structures under Uncertainty, I. Elishakoff and M. Ohsaki, Eds.
London, UK: Imperial college press, 2010.
[2] D. E. Goldberg, Genetic Algorithms in Search, Optimization and
Machine Learning, 2nd ed. United States of America: Addison-wesley
publishing company, Inc, 1989.
[3] C. M. Chan and K. M. Wong, "Structural topology and element sizing
design optimisation of tall steel frameworks using a hybrid OC–GA
method," Structural and Multidisciplinary Optimization, vol. 35, pp.
473‒ 488, 2008.
[4] K. Deb, "Optimal design of a class of welded structures via genetic
algorithms," in Proc. 31st Structures, Structural Dynamics and
Materials Conference, 1990, pp. 444-453.
[5] G. I. N. Rozvany and N. Olhoff, Eds, topology optimization of
structures and composite continua,, Budapes, Hungary, Nato Science
Series II, 2000 vol. 7.
[6] S. Rajeev and C. S. Krishnamoorthy, "Discrete optimization of
structures using genetic algorithms," Journal of Structural Engineering,
vol. 118, pp. 1233‒1250, 1992.
[7] C. Y. Lin and P. Hajela, "Genetic search strategies in large scale
optimization," in Proc. 34th Structures, Structural Dynamics and
Materials Conference, 1993, pp. 2437-2447.
[8] M. Stolpe, "Truss Optimization with Discrete Design Variables: a
Critical Review," Structural and Multidisciplinary Optimization, vol.
53, pp. 349–374, 2016.
[9] S. D. Rajan. (1991-2001) Graphics-based System for Understanding
Structural Analysis. [Online]. http://structures.asu.edu/rajan/gs-usa-
frame-program/
[10] Specification for Structural Steel Buildings‒ Allowable Stress Design
(ASD), American institute of steel construction, Inc. (AISC), 1989.
[11] M. Hultman, "Weight optimization of steel trusses by a genetic
algorithm -Size, shape and topology optimization according to
Eurocode," M. Eng. thesis, Lund Institute of Technology, Sweden,
Feb. 2010.
[12] K. Deb and S. Gulati, "Design of truss-structures for minimum weight
using genetic algorithms," Finite Elements in Analysis and Design,
vol. 37, pp. 447‒465, 2001.
[13] P. Hajela and E. Lee, "Genetic algorithms in truss topological
optimization," International Journal of Solids and Structures, vol. 32,
pp. 3341‒3357, 1995.
[14] L. Li, Z. Huang, and F. Liu, Eds., An improved particle swarm
optimizer for truss structure optimization. Lecture Notes in Computer
Science. Berlin, Heidelberg, Springer, 2006, vol. 4456.

04

Authorized licensed use limited to: Western Sydney University. Downloaded on July 25,2020 at 19:06:17 UTC from IEEE Xplore. Restrictions apply.