Performance of four meta-heuristics to solve a forestry production planning problem
to 50 individuals, the mutation rate equal to 1% and the
crossing rate equal to 80%. For CSA, three different values
were considered for hypermutation (H), cloning (C),
selection (S) and substitution (Sub) rates, being equal to
0.20, 0.50 and 0.80. Three values for initial population
(InP) size were evaluated, being 20, 50 and
80 individuals. In
order to test the CSA parameters, it was
decided to follow the
methodology proposed by Araújo Júnior et al. (2018).
combination was 72.2% for SA, 10.8% for VNS and 4.4%
for
CSA.
The viability of the solutions is related to the fact
that it is necessary to produce an annual quantity of wood
that does not exceed the established minimum and maximum
limits. This was observed for the best solutions (Table 1). It
is important to note that harves- ting a larger total amount of
wood (over 16 years) did not reflect higher NPV. This is
due to the fact that the model seeks to optimize the financial
return rather than the maximization of production. Thus, the
importance of sequencing of the harvest is emphasized so
that there is a better yield in terms of NPV with lower
volume of harvested wood.
For the Simulated Annealing (SA) was evaluated:
three initial temperatures (IT), being 108, 106 and 104; three
temperature reduction rates (), being 0.01, 0.05 and 0.10;
and three numbers of neighbors evaluated in each iteration
(V), with quantities equal to 1, 10 and 30.
In the Variable Neighborhood Search (VNS), two
structures were analyzed: structure A with percentages of
changes of 1%, 2%, 3% and 4% and structure B with
percentages of 10%, 20%, 30% and 40%; each structu- re
was executed considering five different amounts of
neighbors evaluated (1, 10, 30, 50 and 100).
These results demonstrate the superiority of SA
metaheuristics in relation to the search for viable solu-
tions. Such performance had already been highlighted in
other studies, such as those of Ezquerro et al. (2016), in
which the authors mention that SA is one of the most
cited in
the forest literature, and Rodrigues et al. (2004a), where the
authors state that SA is one of the metaheuris-
tics that are less
affected by changes in their parameters.
Data processing considering the different configu-
rations was carried out in the application Metaheuristics for
Forest Planning (MeP), developed in the Laboratory of
Operational Research and Forest Modeling (LPM) of the
Federal University of Minas Gerais.
The metaheuristics AG is widely applied to solve
forestry planning problems (Rodrigues et al., 2004b; Go- mide
et al., 2009; Silva et al., 2009; Binoti et al., 2014 and
Matos,
2017). However, its use has been diminished as
new
algorithms appear that are more efficient during the
search
process of solutions. The algorithm was sensitive to the
parameter definition, as observed by Gomide et al. (2009).
With the results by the algorithms in each repeti- tion,
the amount of viable solutions obtained was counted.
After that,
the fitness values related to the NPV and the solutions that
obeyed the demand for the required wood demand were
evaluated.
Another aspect is related to the number of ope-
rations that the algorithm needs to perform during the
same
generation, especially when adopting a larger num-
ber of
individuals for the initial population. Thus, too much time
is spent without evolution of populations to generate better
individuals, as discussed by Rodrigues et al. (2004b). The
fact that the algorithm found no fea- sible solution in all
cases reveals that it is necessary to evaluate new ways of
improving the GA search process as suggested in Gaspar-
Cunha et al. (2013).
The entire linear programming model was exe-
cuted in the Lingo software considering a stop processing
time
equal to fifteen minutes.
The best solutions presented by each algo- rithm
and their respective configurations were: R$ 31,954,028.00
for the VNS, with 100 neighbors and structure 1; R$
31,872,534.00 for CSA, with initial po- pulation size equal
to 20, substitution rate 0.2, selection rate 0.2, cloning rate
equal to 0.8, and hypermutation rate equal to 0.2; R$
30,782,473.00 for SA, with initial
temperature of 10e8,
temperature decay rate of 0,01 and
evaluation of 30
neighbors. The GA did not present any viable solution in all
evaluations. The solution obtained
by the branch and bound
algorithm presented value equal
to R$ 31,274,496.00.
The CSA and VNS metaheuristics presented bet- ter
solutions in terms of fitness value than the solutions found
by the SA. These algorithms were initially applied to forest
problems in the works of Araújo Júnior et al. (2017, 2018)
presenting efficacy above that found for the traditional
algorithms (SA and AG).
The values of the hypermutation and substitution
rates
were the same ones found by Araújo et al. (2018).
However,
the other parameters did not present the same values. This
shows the need to define a set of parameters
for each situation
analyzed, which suggests the use of adaptive mechanisms,
so that these values are changed during the execution of the
algorithm.
The CSA presented feasible solutions in only three
parameter combinations. VNS presented feasible solutions
in four of the 10 combinations evaluated and the SA
presented feasible solutions in 26 of the 27 com- binations
tested. The mean of viable replicates for each
Cad. Ciênc. Agrá., v. 12, p. 01–05, 2020. e-ISSN: 2447-6218 / ISSN: 1984-6738