Technology and Innovation for Sustainable Development Conference (TISD2008)

Faculty of Engineering, Khon Kaen University, Thailand

28-29 January 2008

Application of genetic algorithms for sugar cane harvesting decision

Panwipa Taechasook1* Kanchana Sethanan1 Sujin Bureerat2
1
Department of Industrial Engineering, Faculty of Engineering, Khon Kaen University, Khon Kaen 40002
2
Department of Mechanical Engineering, Faculty of Engineering, Khon Kaen University, Khon Kaen 40002
E-mail: panwipat@yahoo.com*

Abstract
Inbound logistics interconnection of sugar cane
and sugar industry consists of planning sugar cane
cultivation, harvesting, transportation to the mill, and
management of factory front yard. At present, the
amount of sugar cane transported to the mill relies on
production capacity of the factory such that sugar
cane is continuously supplied to sugar production
process at its fullest capacity. However, optimum
benefit of the industry also relies on the commercial
cane sugar (CCS) value. This research was aimed at
prioritizing harvesting from the plantations of sugar
cane available for the factory, for example the
factory’s own plantations or farmers’ plantations. The
objective was to obtain optimum value of total sugar
yield and the quantity of sugar cane is in accordance
with the factory’s capacity. A mathematical model
was created for the study to solve small-scale
problems, and a genetic algorithm was constructed for
sugar cane harvesting decision for medium- and large-
scale problems. Assessment of the efficiency of the
constructed algorithm was done by comparing the
solutions obtained from the genetic algorithm with
those obtained from the mathematical model. Total
sugar yield values were determined by designing four
experiments, each with 10 iterations. Two factors
were used in the experiment, i.e. the questionable
number of harvesting days and the number of
plantations to be prioritized. These two factors were
divided into two levels. Following this, the
experimental results were statistically analyzed using
the factorial design and SAS version 6.12 program.
The result revealed the average differential percentage
of total sugar yield from the genetic algorithm and
mathematical model of 10.99 percent. It was also
found that the two factors used in the experiment had
an effect on the percentage of differences of the total
sugar yield values at a significant level.
Keywords: Sugar cane harvest, Genetic algorithm,
Integer programming, Inbound logistics
1. Introduction
Sugar cane and sugar industry is an important
industry of Thailand since there are a great number of
sugar cane farmers spreading all over the different
regions of Thailand. Thai sugar mills are known to
produce sugar for domestic consumption as well as
for export, creating a lot of incomes for all who are
involved in this industry.
Inbound logistics of sugar cane and sugar
industry consists of planning sugar cane cultivation,
harvesting, transportation to the mill, and
management of factory front yard. In cultivation
planning, major factors involved are for examples,
cane breeds, weather, etc., which have an effect on the
commercial cane sugar. The results indicate the sugar
cane product and value of commercial cane sugar.
At present, harvest and transportation of sugar
cane to the factory is the function of productivity.
This is because the production capacity of a factory
each day is limited. Any remaining sugar cane at the
end of the day need to be pressed the following day
and cannot be retained for a longer period, for that
will reduce its commercial cane sugar value
(Semenzato, 1995), reducing the farmers’ sugar cane
price and hence a less amount of cane juice on the
part of the mill.
Thus, the study of sugar cane harvesting
prioritization aiming at obtaining the maximum total
sugar yield, which is related to the daily capacity of
the factory was important. Sugar cane harvested
appropriately both in quantity as well as quality will
benefit all stakeholders. This research explains the
problems of sugar cane harvesting allocation, which
provides harvesting patterns a factory can choose, i.e.
from the factory’s own plantations, or buying from
farmers’ plantations. The method used in the problem
solution was to develop a mathematical model and
apply the genetic algorithm, a technique for finding a
solution to a large-scale, complex problem with
increasing variables and conditions. The mathematical
model was found to be time consuming and difficult
to find the answer. In addition, the efficiency of
genetic algorithm was assessed through comparison
of answers derived from them and from the
mathematical model.
2. Literature review
2.1 Management of sugar cane harvest
Problems related to sugar cane harvest are being
interested by many sugar cane-growing countries, for
example Australia, Cuba, South Africa, Thailand, etc.
There has been research work on sugar cane
harvesting in many aspects, most of which were based

172
on the development of a mathematical model and
solution-finding technique. One example is Higgins
(1999) who developed an integer mathematical model
for scheduling harvesting periods, cultivating cycles
and ploughing of sugar cane plantations, with an aim
to obtain maximum return. The tabu search technique,
as apposed to the tabu list was applied for solution
finding, and the linear mathematic patterns developed
by Higgins et al (1998) were then used in
approximating the initial solutions for Tabu search.
Jiao et al. (2005) developed a Second-order
polynomial mathematical model for allocation of
appropriate harvesting and time of harvesting from
the relation of commercial cane sugar with the
plantation area, with an aim to obtain the maximum
commercial cane sugar value. Salassi et al. (2002)
developed an integrated integer mathematical model
to determine quality traits, i.e. stalk weight and
product per stalk in order to find the suitable
harvesting day, the major purpose of which was,
again, to obtain maximum sugar product.
2.2 Genetic algorithm
Wanida and Supachai (2003) explained that the
genetic algorithm is an algorithm that is used to
determine the most satisfying solution from the
sample group of answers based on selection
mechanism and natural reproduction. Primarily, a
number of probable answers were randomly selected.
The most satisfying solution was selected from the
best ones among the surviving group or among the
previously selected breed group by pairing a good
breed group and locating the position in which there
would be a change of pattern between the pair so that
the new pair of solutions would emerge, which made
it possible to develop a result nearest to the real target.
Although this is based on a simple principle but is
effective for a large-scale problem without having to
determine all probable variables and then select the
best answer. Santi and Duangkaew (2005) discussed
the function of genetic algorithm as follows. Firstly,
the initial population had to be created by randomly
creating values of each bit of each chromosome. Then
the suitable value of each chromosome was analyzed
followed by constructing a mating pool or parent
chromosomes that could survive as the prototype.
Finally, genetic implementation was carried out by
sampling the mating pools to produce a new
generation of population. The operation is as follows.
1. Reproduction: the process of selecting
highly appropriate chromosomes as the initial solution
for the next generation.
2. Crossover: Crisscrossing by exchanging
parts of parent chromosomes based on probability of
crossover: Pc in order to develop offspring
chromosomes.
3. Mutation: the process of mutating some
values of chromosomes to obtain new values at the
random position based on probability of determined
mutation: Pm.
Athit (2002) designed a flow chart of
173
procedures of genetic algorithm and linked to the real-
world system as shown in Figure 1.
Creating initial S = System answer
population
E = Appropriate value
Population S
evaluation (for each
chromosome)
E
Reproduction Objective System
Functions
Crossover and
mutation E
Population S
evaluation
Population replacement
Figure 1. General procedures of GA in real-world
system
3. Mathematical model in prioritizing sugar
cane harvesting
The developed mathematical model explained
the problem found in prioritizing sugar cane
harvesting. The aim was to obtain total sugar yield.
The interesting point was the factory’s selectable
pattern. The details of this mathematical model are as
follows:
3.1 Assumptions
In mathematical modeling, assumptions of sugar
cane harvesting problems were made in line with
various problems found. The factory was able to
select the harvesting patterns as in the following
details.
3.1.1 Sugar cane prioritized for harvesting could
be supplied directly to the sugar mill, with minimum
sustaining time, hence not lessening commercial cane
sugar value.
3.1.2 There was a uniform commercial cane
sugar value of sugar cane in each plantation.
3.1.3 Sugar cane prioritized for harvesting is
ready for harvesting; there is no flooding problem or
plantation condition making it unable to harvest, or
any problem related to shortage of harvesting labor.
3.1.4 The total quantity of sugar cane grown is
less than the total production capacity. This enabled
us to stipulate the quantity of harvest of each
plantation such that the total sugar yield was attained.
3.1.5 Transportation cycles could be assigned
unlimitedly according to the determined quantity of
harvesting, provided that the quantity of sugar cane
reaped from each plantation was equal to one another
and not exceeding the payload.
3.1.6 Each loading of sugar cane on a truck had
 to be from the same plantation. Sugar cane from any
other plantation should not be mixed.
3.2 Parameters
Following are the parameters used in the
mathematical model of sugar cane harvesting
prioritization.
i harvesting day index; i = 1, 2,…, t
j plantation index; j = 1, 2,…, n
k truck load index; k = 1, 2,…, m
Cij CCS value of sugar cane in Plantation j
on Day i
MC i Production capacity required by factory from
the plantation to be selected on Day i
TQ j Maximum harvested quantity per day of
Plantation j
qt k Payload k per one truck
Nmijk Maximum number of trucks having k capacity
on Day i and Plantation j
3.3 Decision variables
Following are the decision variables used in the
mathematical model of sugar cane harvesting
prioritization.
Qij Sugar cane quantity to be harvested from
Plantation j on Day i
nijk Number of trucks having
Day i and Plantation j
3.4 Objective Function
The objective function of the mathematical
model of sugar cane prioritization was
t n
the total quantity of sugar cane of all harvesting days
in Plantation j which did not exceed the total sugar
cane quantity of that plantation. Equation (4) was the
sum from multiplying results of number of trucks and
truck payload, which was less or equal to sugar cane
quantity harvested on Day i from Plantation j .
Equations (5) and (6) mean the number of trucks
used, which did not exceed the total number of trucks
and was greater than or equal to 0 respectively.
Equation (7) shows the amount of sugar cane
harvested on Day i from Plantation j which was
greater or equal to 0.
4. Genetic algorithm of sugar cane
harvesting prioritization
Solutions of sugar cane harvesting prioritization
derived from the mathematical model were limited to
small-scale problems with a small number of
variables and equality or inequality constraints. In
reality, however, there were many variables and
equality or inequality constraints, and so solution
finding would be time-consuming and complex.
Hence, solving large-scale problems should be based
on a heuristic approach. This research study applied
the genetic algorithm approach, which is one solution-
finding technique. This genetic algorithm application
was based on the following procedures.
k capacity on First step: Produce initial chromosomes, with
each having a number of bit and each bit having a
binary value of 0 or 1. The initial binary of
chromosomes was derived randomly.
Second step: Find fitness value of chromosomes
from the objective function. The objective function of

∑∑ C
the problem involving equality or inequality
Maximize ij * Qij (1)
constraints had to be changed to penalty function
i =1 j =1
f ' ( x) instead of the former objective function f ( x) ,
Equation (1) was the result of multiplication of
commercial cane sugar and amount, or the total sugar as can be seen from Equations (8) and (9),
f ( x)
yield, with an aim to obtain optimum values. f ' ( x) = + (1 − μ c ( x)) ⋅ K D (8)
3.5 Equality or Inequality Constraints 1 + f ( x)
The following equality or inequality constraints where
were used in the mathematical model of sugar cane ⎧0; if g i ( x) ≥ d i
harvesting prioritization. ⎪
⎪ g ( x) − ε i
n
μ c ( x) = ⎨1 − i
∑Q
; if ε < g ( x) < d
ij ≤ MC i ; ∀i (2) ⎪ di i i i
j =1 ⎪1;
t
⎩ if g ( x ) ≤ ε
i i

∑
i =1
Qij ≤ TQ j ; ∀j (3) (9)
where ε i และ d i was the probable scope value of each
m inequality constraint having i , K D as the penalty
Qij ≤ ∑ qt
k =1
k * nijk ; ∀i∀j (4) coefficient.
Third step: Produce a new set of chromosomes
nijk ≤ Nmijk ; ∀i∀j∀k (5) following three sub-procedures; reproduction, cross-
nijk ≥ 0 และ nijk integer ; ∀i∀j∀k (6) over, and mutation.
Step 3.1: In reproduction, sets of chromosomes
Qij ≥ 0 ; ∀i∀j (7) were selected in order to transfer good chromosomes
Equation (2) means sugar cane quantity from all to the production of the next set. Those having low
plantations on harvesting day i which did not exceed objective function values had a higher chance to be
the factory’s capacity of that day. Equation (2) means selected than chromosomes with high objective

174
function values. Selection was done by random using
a roulette wheel.
Step 3.2: Crossover means the mixing of values
between parents that had low values of objective
functions, resulting in one pair of offspring. There are
a number of crossover approaches. This research
applied a two-point crossover as shown in Figure 2
with probability value Pc.
Step 3.3: In mutation, new data were produced
for chromosome sets by mutating chromosomes as
shown in Figure 3. A chance that mutation would
occur was very small, with a probability of Pm.
Step 3.4: The chromosomes that had undergone
the first three steps were taken to find objective
functions as in Step 2 to obtain data for the next
production.
Parent 1 :
Parent 2 :
Offspring 1 :
Offspring 2 :
Figure 2. Two-point crossover
Fourth step Repeat all of the 4 sub-steps in Step
3 until the pre-determined cycles of chromosome
production were completed.
Fifth step After Step 4, the chromosomes with
lowest objective functions were selected as the
solutions. These solutions were in binary scale and
needed to be changed to the ten-scale system so that
the answers showed the right quantity of sugar cane to
be harvested on the set harvesting day and from the
determined plantation.
Parent :
Randomly selected position
Offspring :
Figure 3. Mutation
5. Examples of Computation of Genetic
Algorithms of Sugar cane Harvesting
Prioritization
Suppose the relevant parameters for prioritizing
sugar cane harvest had the following details: Six
interesting days to harvest and transport sugar cane,
10 plantations for computation with each plantation
capacity for transportation of not over 40 units,
commercial cane sugar value of each plantation
having an increasing trend during the first period,
being stable during the mid-period, and a declining
trend toward the end period (Jiao et al, 2005). In
addition, the total quantity of sugar cane from all
175
plantations under consideration was less than the total
quantity the factory required for the total production
days. The steps in computing by genetic algorithm
were as follows.
First step Produce initial chromosomes by
sampling 50 chromosomes. Each was changed from
the real number to a binary system, with the total bit
of 8. The quantities to be harvested each day from
each plantation were set from 0 to 40.
Second step Change the objective function from
the problem of maximum value finding to the lowest
one because the genetic algorithm applied in this
research was based on lowest value finding. Thus, the
value of objective function had to be changed to a
lowest value according to Equation (10).
t n
Maximize ∑∑ C
i =1 j =1
ij * Qij =
t n
Minimize −( ∑∑ C
i =1 j =1
ij * Qij ) (10)
Then the objective function was changed to penalty
function based on Equation (8).
Third step Produce a new set of chromosomes
by reproduction, crossover (Pc = 1.0), and mutation
(Pm=0.01) in order to find the objective function of
the new chromosome set.
Fourth step Repeat Step 5.3 until the 300 cycles
of computation was completed.
Fifth step Select the objective function with
lowest value as the answer to the sugar cane
harvesting prioritization. This provided the order of
priorities, the quantity to be harvested from each
plantation for each harvesting day, and the value of
sugar product derived from harvesting according to
the priorities, which would be the answers of
maximum values.
6. Experimental Design
The objective of the experiment in this research
study was to obtain maximum total sugar yield.
Solution was found from the mathematical model and
from the application of genetic algorithm. Different
parameters were established from the creation of
questions with the parameter of the quantity of sugar
cane from all plantations being lower than the total
sugar cane required by the factory. The transportation
quantity was also limited for each day and the
commercial cane sugar values of sugar cane of each
plantation were based on the trend invented by Jiao et
al (2005). The trend varied with the number of days of
cultivation. There were 3 periods. First, the
commercial cane sugar values were increasing.
Second, the commercial cane sugar values were
stable, and finally, the values declined.
In this experimental design to select sugar cane
plantation for harvesting with an aim to optimize
sugar productivity, the researchers have taken into
account related factors that had impact on total sugar
yield. These were the number of sugar cane
plantations (N) and the length of harvesting period purpose of inbound logistics for sugar and sugar cane
from the beginning to the end (D). Table 2. Responses of experimental design from each
Factorial design 22 was used in the design in factor (Percentage)
which the factor values at each level could be differences of total sugar yield (%)
Iteration
stipulated, as in Table 1. Each combination of the Number
6 days 10 days
10 plan- 15 plan- 10 plan- 15 plan-
experiment was subject to 10 iterations. The value of tations tations tations tations
each iteration was the differential value of maximum 1
10.12 11.51 12.65 13.24
yields between the optimal solution from 2
10.17 11.79 8.24 13.67
mathematical model and solution obtained from 3
8.97 12.59 11.49 13.04
genetic algorithm application. Additionally, multiple 4
comparison was investigated using the Duncann’s 6.60 10.82 11.33 14.47
5
Multiple Range Test in order to observe any 8.13 12.94 7.32 13.36
6
statistically significant differences among the factors 10.65 11.91 5.77 14.12
at each level. 7
9.43 12.44 9.25 12.51
8
8.71 11.94 10.87 15.90
Table 1. Experimental design 9
7.32 11.81 8.66 13.22
N D 10
Combination 6.47 12.66 8.83 14.74
Number L H L H Average
(10 plantations) (15 plantations) (6 days) (10 days) 8.66 12.04 9.44 13.82
1 √ √
Table 3. Results of Factorial design statistic analysis
2 √ √
from SAS program
3 √ √ Source DF Anova SS Mean F Pr > F
Square Value
4 √ √ Day 1 0.001652 0.001652 8.13 0.0072
Plantation 1 0.015098 0.015098 74.26 0.0001
Day* 1 0.000252 0.000252 1.24 0.2732
Plantation
7. Results of Experiment
The factorial design 22 enabled us to determine
the percentage of differences of total sugar yield Table 4. Results of Multiple comparison of the
between the optimum solution and the solution number of harvesting days by Duncann’s Multiple
obtained from genetic algorithm, as shown in Table 2. Range Test
Duncan Grouping Mean N Level
It can be seen from Table 2 that the average A 0.116346 20 2
percentage of value difference between the optimum B 0.103491 20 1
solution and the solution from genetic algorithm was
10.99, with the lowest average being 8.66% and the
Table 5. Results of Multiple comparison of the
highest average being 13.82%. Following this, the
number of plantations by Duncann’s Multiple Range
percentages of differences among the solutions were
Test
analyzed statistically using SAS Version 6.12 Duncan Grouping Mean N Level
program. The results, which are shown in Table 3, A 0.129347 20 2
reveal that the two factors of the number of harvesting B 0.090490 20 1

days and the number of plantations had significant

effect on the differences in the values of total sugar
yield between the optimum solution and the solution
from genetic algorithm at 95% reliability. It can be
seen that p-values of the two factors were lower than
0.05. The multiple comparison, as depicted in Tables
4 and 5 reveal that the different levels of the two
factors had a significant difference of average
responses at a reliability level of 95%. The low level
of each factor had a response value lower than the
high level. This means that the two factors, the
number of harvesting days and the number of
plantations bearing low values made the solution from
genetic algorithm nearer to the most appropriate
solution.
8. Conclusion
Harvesting sugar cane with high commercial
sugar cane value and transportation of the required
quantity of sugar cane to factories are the major
industry. Prioritization of sugar cane harvesting
according to the quantity required by the factory is an
important tool that efficiently assists sugar cane
harvesting decision. This research study presented a
mathematical model, with an aim to optimize total
sugar yield, to explain problems in the patterns of
sugar cane plantation selectable by the factory. This
enabled us to determine the order and quantity of
harvesting which was the most appropriate solution
for small-scale problems. Genetic algorithm was
applied in prioritizing harvesting in the case of large-
scale and complex problems.
The experiment conducted through the use of
genetic algorithm and comparison of total sugar yield
values and the most appropriate solution from
mathematical model revealed that the average value
of differences between the solution from genetic
algorithm and the optimum solution was 10.99%. The
differences between solutions from the two

176
approaches became higher when the problem was
bigger both in terms of harvesting days and number of
plantations taken into account in prioritization. Hence,
in order to obtain solution values from genetic
algorithm that were close to the optimum solution
values, genetic algorithm could be improved by such
methods as follows. (1) Creating heuristic to find
initial solutions prior to entering the genetic algorithm
process instead of sampling initial solution values,
which would be faster. (2) Increasing computation
cycles. (3) Increasing solutions or chromosomes in
each computation cycles in order to improve the
solution values from each set of crossover and to
accelerate solution findings. Such improvement could
be parallel among all sets so that the efficiency in
solution finding in genetic algorithm application is
increased in prioritizing sugar cane harvest.

References
[1] Semenzato, R. 1995. A Simulation Study of Sugar
Cane Harvesting. Agricultural Systems. 47: 427-
437.
[2] Higgins, A.J. 1999. Optimising cane supply
decisions within a sugar mill region. Journal of
Scheduling, 2: 229-244.
[3] Higgins, A.J., Muchow, R.C., Rudd, A.V. and
Ford, A.W. 1998. Optimizing harvest date in sugar
production: A case study for the Mossman mill
region in Australia. Field Crops Research. 57:
153-162.
[4] Jiao, Z., Higgins, A.J. and Prestwidge, D.B. 2005.
An integrated statistical and optimization
approach to increasing sugar production within a
mill region. Computers and electronics in
agriculture. 48: 170-181.
[5] Salassi, M.E., Breaux, J.B. and Naquin, C.J. 2002.
Modelling within-season sugar cane growth for
optimal harvest system selection. Agricultural
Systems. 73:261-278.
[6] Wanida Rattanamanee and Supachai Patumnakul.
2003. Satisfactory Solution Finding by Genetic
Algorithm. KKU Engineering Journal. 3: 319-326.
[7] Santi Suansri and Duangkaew Sawamiphak. 2005.
Use of Genetic Algorithm in Automatic
Composition of Web Service. Science and
Technology Journal. Thammasart University. 3: 1-
9.
[8] Athit Srikaew. 2002. Genetic Algorithm Section 1.
Journal of Suranaree Technological University. 1:
61-68.

177