Anda di halaman 1dari 9

Multiobjective Optimization of Bioactive Compound

Extraction Process via Evolutionary Strategies


Timothy Ganesan1(), Irraivan Elamvazuthi2,
Pandian Vasant3, and Ku Zilati Ku Shaari1
1
Department of Chemical Engineering,
Universiti Teknologi Petronas, 31750 Tronoh, Perak, Malaysia
tim.ganesan@gmail.com
2
Department of Electrical and Electronics Engineering,
Universiti Teknologi Petronas, 31750 Tronoh, Perak, Malaysia
3
Department of Fundamental and Applied Sciences,
Universiti Teknologi Petronas, 31750 Tronoh, Perak, Malaysia

Abstract. Systematic and simultaneous optimization of a collection of objectives


is called multiobjective or multicriteria optimization. These sorts of optimization
procedures are becoming commonplace in fields involving engineering design,
process and system optimization. In this work, the multiobjective (MO) optimization of the bioactive compound extraction process was carried out. Using the
Normal Boundary Intersection (NBI) approach the MO optimization problem is
transformed into a weighted form called the beta-subproblem. This subproblem
is then solved using two evolutionary strategies (differential evolution (DE) and
genetic algorithm (GA)). Using these evolutionary strategies, the solutions to
the extraction process which form the efficient Pareto frontier was generated.
The Hypervolume Indicator (HVI) was applied to the solutions to rank the
strategies based on the solution quality. Critical analyses and comparative studies were then carried out on the strategies employed in this work and that from
the previous work.
Keywords: Multi-objective optimization Extraction process Normal Boundary Intersection approach Differential evolution (DE) Genetic algorithm
(GA) Hypervolume Indicator (HVI)

Introduction

The problem formulation of the extraction process was done by Shashi et al, (2010) [1].
This formulation involves the modeling of the objective functions and the identification
of the range of the extraction process parameters. The primary target was yield optimization of specific extracted chemical products from the Gardenia Jasminoides Ellis fruit.
The process yields three bioactive compounds; crocin, geniposide and total phenolic
compounds. Identifying a series of optimal process yields that generate an efficient Pareto frontier is critical. Gauging solution quality in MO optimization can be very difficult
and tricky. Ideas involving solution properties like diversity and convergence have become popular in recent times [2], [3]. The utility of these ideas have proved useful for
Springer International Publishing Switzerland 2015
N.T. Nguyen et al. (Eds.): ACIIDS 2015, Part II, LNAI 9012, pp. 1321, 2015.
DOI: 10.1007/978-3-319-15705-4_2

14

T. Ganesan et al.

developing metrics for evaluating specific aspects of the generated solutions. Hence,
these metrics provide the decision maker with some useful information regarding the
techniques effectiveness [4]. However, due to the local nature of these metrics, an absolute ranking of the solutions is not attainable. One effective approach that can be utilized
for the overall ranking of solution sets is the Hypervolume Indicator (HVI) [5] which is
based on the idea of Pareto dominance. This metric measures the Hypervolume (multidimensional) enclosed by a Pareto front approximation with respect to a reference set
(see [6], [7], and [8]). This metric ensures strict compliance to monotonicity related to
Pareto dominance [9], [10]. This makes the ranking of solution sets and hence algorithms
feasible for any given MO problem. The techniques introduced in this work is directed to
generating a series of solutions (with the associated weights) which efficiently approximates the Pareto frontier.
This problem was attempted using the particle swarm optimization (PSO) technique (within a weighted sum framework) in Shashi et al., (2010) [1]. In that work,
more emphasis was given on the modeling works as compared to the optimization
procedures. Although an individual solution optimal solution is attained, the approximate Pareto frontier of the solutions was not constructed and rigorous solution evaluation (for ranking purposes) was not conducted. In this work, the bioactive
compound extraction process optimization problem was tackled using Differential
Evolution (DE) [11] and Genetic Algorithm (GA) [12]. This was carried out within
the basis of the Normal Boundary Intersection (NBI) framework [13]. The ranking of
the techniques of the Pareto frontiers produced by the algorithms were carried out
using the HVI metric [5], [6]. Comparative analyses were then conducted on the individual best solutions and the frontiers obtained in this work against those obtained in
Shashi et al., (2010) [1].
Genetic Algorithms (GA) were the earliest form of evolutionary algorithms
introduced by Holland, (1992) [12]. These algorithms contain the fundamental components that make up an evolutionary algorithm such as the cross-over operator,
mutation operator and fitness evaluation mechanisms (which aids the algorithm to
successively improve the populations fitness during execution). Differential Evolution (DE) is a population-based evolutionary algorithm that has been derived from
Genetic Algorithms (GA) [12]. DE was developed in the nineties by Storn and Price
[11]. DE has been used extensively to solve problems which are non-differentiable,
non-continuous, non-linear, noisy, flat, multidimensional, have many local minima,
constraints or high degree of stochasticity. Lately, DE has been applied to a variety of
areas including optimization problems in chemical and process engineering
[14],[15],[16]. This paper is organized as follows: Section 2 of this paper presents an
overview of the evolutionary strategies. In Section 3, the process formulation is described followed by Section 4 which discusses the computational results. Finally, this
paper ends with the concluding remarks in Section 5.

Evolutionary Strategies

Evolutionary intelligence originate from the idea presented in Hollands [12] genetic
algorithm. The central theme of all evolutionary algorithms is that survival of fittest
(natural selection) which acts on a population of organisms under environmental

Multiobjective Optimization of Bioactive Compound Extraction Process

15

stress. Thus, a fitness distribution is formed in the mentioned population of organisms


(individuals). In the case of optimization, let there be an objective function to be
optimized. Then, various random candidate solutions (individuals) to the problem can
be gauged by applying a fitness function to each of these solutions (where the higher
the fitness function, the better the solution quality with respect to the objective
function). Each evaluated individual with respect to the best fitness values are then
chosen for producing the next generation of potential solution vectors through the
process of cross-over and mutation. The cross-over operator crosses the parent
individuals from the previous generation (fittest individuals) and produces the next
generation of individuals. The mutation operator on the other hand, perturbs the gene
pool of the population to generate new optimization capabilities in the new
individual offspring. Thus, each individual competes in their population as they
achieve higher fitness values. The repeated execution of this cycle is meta-heuristic
since the solution vectors in the population are subsequently improved as the
iteration proceeds. In this work, Genetic algorithm (GA) and Differential Evolution
(DE) algorithms are used.
2.1

Differential Evolution (DE)

DE is a class of evolutionary techniques first introduced in 1995 by Storn and Price


[11]. This class of techniques pioneered the development of perturbative evolutionary
algorithms. DE begins by randomly initializing a population (P) in the first generation
of at least four individuals. These individuals are real-coded vectors with some size,
N. A single principal parent (xpi ) and three auxiliary parents denoted (xai) is randomly
selected from the population, P. In DE, every individual, I in the population, P would
become a principle parent, xpi at one generation or the other. Therefore each individual has a chance to breed with the auxiliary parents, xai. The three auxiliary parents
then be mutated (via differential mutation) to generate a mutated vector, Vi.
Vi = xa1 + F(xa2 - xa3)

(1)

where F [ 0 ,1] is the real-valued mutation amplification factor. Next Vi is then


recombined (or exponentially crossed-over) with xpi to generate a child trial vector,
xchildi. The cross-over problability (CR) is set by the user. In DE, knock-out competition is used to select the survivors which will then be inserted into the next populations gene pool. This fitness selection procedure required the principle parent, xpi
and the child trial vector, xchildi to engage in direct competition:
x child ( gen ) f ( xichild ) better than f ( xip )
x i ( gen + 1 ) = i p
otherwise
x i ( gen )

(2)

The parameters initialized for the DE algorithm is given in Table 1. The algorithm of
the DE method is shown in Execution Scheme 1.

16

T. Ganesan et al.
Table 1. DE Parameter Setting
Initialized Parameters

Values

Individual Size, N

Population Size, P

Mutation amplification factor, F

0.3

Cross-over Probability, CR

0.667

Execution Scheme 1: Differential Evolution (DE)


Step 1: Initialize input parameters
Step 2: Set random population vectors, xGi.
Step 3: Select one random principal parents, xpi
Step 4: Select three random auxilary parents, xai
Step 5: Implement differential mutation to generate a mutated vector, Vi
Step 6: Merge Vi with xpi to produce a child trial vector, xchildi
Step 7: Perform fitness evaluation for the next generation
Step 8: IF the fitness condition is fulfilled and t= Tmax , stop and print solutions
ELSE repeat step 3

2.2

Genetic Algorithm (GA)

GA is one of the first evolutionary search and optimization techniques [12]. This population-based approach uses an N-point crossover operator to create new offspring for
successive generations. To avoid algorithmic stagnation at some local minima, mutation operators are usually employed to diversify the search. In this work a bit flip-type
mutation operator was utilized. The GA scheme applied in this work is provided in
Execution Scheme 2. The parameter settings initialized prior to the execution of the
GA used in this work are shown in Table 2. The flow of the GA algorithm is shown in
Execution Scheme 2:
Table 2. Genetic Algorithm Parameter Setting

Parameters

Values

Define individual string length


Define amount of individuals in the
population
Mutation Probability
Recombination Probability
Cross-over type
Mutation type
Selection type

5 bits
6
0.3333

N-point
N-bit flip
Tournament

Multiobjective Optimization of Bioactive Compound Extraction Process

17

Execution Scheme 2: Genetic Algorithm


Step 1: Randomly initialize chromosomes for n individuals in the population.
Step 2: Designate fitness terms for all individuals.
Step 3: Using recombination create offspring for the next generation.
Step 4: Mutate offspring for this generation.
Step 5: Via tournament selection, select parents to create the next generation.
Step 6: Choose the next population of n individuals.
Step 7: Set a new population to current population.
Step 8: Evaluate the fitness of each offspring in the generation.
Step 9: IF the stopping criterion is met, halt program and print solutions,
ELSE go to Step 3.

Process Formulation

The formulation of the extraction process was developed in Shashi et al., (2010) [1].
Through this work the model describing the yields (of specific chemical products)
which are extracted from the Gardenia Jasminoides Ellis fruit was attained. Equipped
with this, the complete MO optimization model for the extraction process of bioactive
compounds from gardenia with respect to the constraints was successfully developed.
The MO optimization model was established to maximize the yields which consists of
three bioactive compound; crocin (f1), geniposide (f2) and total phenolic compounds
(f3). This process extraction MO system is given as follows:
Maximize

Yields (f1, f2, f3)

subject to process constraints.

(3)

The objective functions (yields of each of the bioactive compound in the units of
mg/g of dry powder) are modeled with respect to the constrained decision variables.
This model is given as follows:

f1 = 3.84 + 0.06978158 X 1 + 0.0208399 X 2 + 0.0417941X 3 0.00121613 X 12


+ 0.000094209 X 22 0.00031697 X 32 + 0.000575415 X 1 X 2 0.000910268 X 2 X 3 (4)
+ 0.000770328 X 1 X 3
f 2 = 46 .2468 + 0 .68221 X 1 + 0 .428799 X 2 + 1.0 X 3 0.016162 X 12 0.0034969 X 22 (5)
0 .0116497 X 32 + 0 .0121561 X 1 X 2 0 .0095057 X 2 X 3 + 0.00889058 X 1 X 3
f 3 = -6.8 + 0.41185 X 1 + 0.339428 X 2 + 0.34154 X 3 0.00538109 X 12 0.00211038 X 22 (6)
0.00419323 X 32 + 0.00164593 X 1 X 2 0.00125152 X 2 X 3 + 0.000783732 X 1 X 3

X 1 [19.5, 80.5] , X 2 [27.1, 72.9] , X 3 [7.1, 52.9]

(7)

18

T. Ganesan et al.

Computational Results and Discussion

All techniques employed in this work were developed using the C++ programming
language on a personal computer (PC) with an Intel dual core processor (running at 2
GHz). The solution sets which construct the Pareto frontier was generated using the
two evolutionary techniques (GA and DE). The HVI was utilized to evaluate and absolutely rank these solutions in terms of dominance. In these evaluations, the nadir point
which is the most non-dominated point is utilized as a reference point for the computation of the hypervolume. The nadir point used in this work is (r1, r2, r3) = (0, 0, 0).
Along with the construction of the entire Pareto frontier, the individual solutions were
classified into the best, median and worst solutions were identified. These solutions
generated by both the algorithms with their respective rankings are shown in Table 3.
Table. 3. Individual Solutions Generated by the DE and GA Algorithms
Technique

Differential Evolution (DE)

Genetic
Algorithm
(DE)

Description

Best

Median

Worst

Objective Function
(f1,f2,f3)

(8.5265,
109.414,24.7106)

(8.2968,
106.016,
24.3831)

(7.4067, 92.6152,
20.5962)

Decision Variable
(x1,x2,x3)

(54.9127,
72.8872,
40.4613)

(42.5321,
72.599, 35.4178)

(23.9009, 72.5093,
27.9649)

HVI

23052.89406

21447.08821

14128.51344

Weights
(w1,w2,w3)

(0.2,0.7,0.1)

(0.1, 0.2, 0.7)

(0.5, 0.2, 0.3)

Objective Function
(f1,f2,f3)

(8.0301, 95.1255
19.8195)

(7.7289,
91.1258, 18.606)

(7.6231, 88.7974,
17.7027)

Decision Variable
(x1,x2,x3)

(29.1553,
72.7404,
9.2592)

(23.8437,
72.6443, 9.9997)

(21.8074, 72.7049,
8.37861)

13104.1788

11983.0976

HVI

Particle
Swarm
Optimization by
Shashi et
al.,
(2010)[1]

15139.5241

Objective Function
(f1,f2,f3)

(8.52, 108.761,
24.67)

(7.7289,
91.1258, 18.606)

(7.6231, 88.7974,
17.7027)

Decision Variable
(x1,x2,x3)

(50.11, 72.23,
28.72)

HVI

22860.3006

Weights
(w1,w2,w3)

(0.33, 0.33, 0.33)

Multiobjective Optimization
O
of Bioactive Compound Extraction Process

19

The Pareto frontier was co


onstructed approximately using 28 solutions from variious
weights. The weights weree varied at an interval of 0.1 for the range of 0 to 1. T
The
Pareto frontiers generated using
u
the DE and GA techniques are presented in Figurre 1
respectively:

Fig. 1. Pareto fro


ontier spread obtained using the GA and DE methods

Using the HVI, the dominaance ranking of the solutions produced by each of the allgorithms employed in this wo
ork is possible. This way the HVI of the entire Pareto frrontier was evaluated. The exeecution time for constructing the entire frontier by the DE
technique is higher than the
t GA technique by 71.719% (where DE takes 4.44414
seconds and the GA takes 18.2913 seconds). The DE technique outperforms the GA
approach in terms of domiinant frontier generation. It can be observed that the DE
approach produces a moree dominant frontier as compared to the GA approachh by
54.439%. In terms of best in
ndividual solution by dominance ranking, the DE approoach
outperforms both the GA and
a the PSO[1] techniques by 52.2696% and 0.8425% respectively. Thus, the DE ap
pproach has produced a new individual optima as compaared
to the previous work [1].
It can be observed in Table
T
3 that the DE and PSO[1] techniques share a sim
milar
feature during the optimization of the individual objectives. Both methods tendd to
emphasize on maximizing the
t second objective f2 as compared to the other objectivves.
In most computational optiimization techniques, the factor of execution time usuaally
conflicts with the notion off solution quality. This is because better solution quality is
usually achieved when the technique spends more time searching the objective sppace
thoroughly. However, in th
his case, the DE technique seems to produce highly doominant solutions at the frontieer and it simultaneously manages to do this with minim
mal
execution time as compared
d to the GA approach. This is may be attributed to the llow
algorithmic complexity of the DE method as compared to the GA and most otther
metaheuristic approaches. In
I addition, it can be observed in Figure 1 that the solutiions
constructing the Pareto fro
ontier generated by the DE technique is very narrow and
specific in certain regions of
o the objective space as compared to the one producedd by
the GA approach. Thus, th
he reason the DE technique is more effective is becausse it
manages to limit the search
h to the region in the objective space with the highest fitness and not get deviated into other less fruitful search trajectories. This propeerty
ore capability in finding more optimal solutions in the obgives the DE technique mo
jective space. The HVI is a reliable tool for ranking the solution spreads for a M
MO

20

T. Ganesan et al.

optimization problem. However, this metric is very dependent on the selection of the
nadir point. If the correlation between the nadir point and the hypervolume value is
irregular, then the HVI should be tested with multiple values of the nadir point to
attain a trustworthy dominance ranking. Both evolutionary techniques employed in
this work performed stable computations. Search stagnation, solution divergence or
halting problems did not occur during the numerical experimentations. The solutions
constructing the Pareto frontier were all within the specified ranges and thus realistically feasible. In this work, a new optimal set of solutions (see Pareto frontier in
Figure 1) has been achieved using the DE approach within the NBI framework. As
compared to the weighted sum approach in Shashi et al., (2010) [1], the NBI framework utilized in this work seem to be more effective. This may be due to the geometrical aspect in the NBI framework which equips the metaheuristic with enhanced
search capabilities as compared to the weighted sum approach which merely acts as a
conventional scalarization system. Since DE is an enhanced evolutionary-type algorithm (perturbative improvement), the diversification and the rigorousness of the
search is high as compared with the GA and PSO [1] methods.

Conclusions and Future Research Directions

Using the DE technique, a new local maximum for the individual objectives was
achieved. In addition, a more dominant approximation of the Pareto frontier was constructed by using the DE method. It was observed that among evolutionary strategies,
perturbative techniques such as DE are computationally inexpensive and effective in
solving industrial MO optimization problems. When gauged with the HVI metric, the
DE approach produced the most dominant approximate of the Pareto frontier as compared to the GA and the PSO [1] methods.
In the future, more thorough investigations using an optimization framework
should be applied to this problem [17]. In addition, other algorithmic enhancement
mechanisms such as (chaos-based improvements [18]) could be incorporated into the
DE approach. For future works, other meta-heuristic algorithms such as hybrid approaches and PSO variant approaches (e.g. Hopfield-PSO [19] and Binary-PSO [20])
could be applied to extraction process problem. Besides the HVI, convergence, diversity and spacing metrics could also be employed to provide more insight regarding the
solution properties.

References
1. Deep, S.K., Katiyar, V.K.: Extraction optimization of bioactive compounds from gardenia
using particle swarm optimization. In: Proceedings of Global Conference on Power Control and Optimization (2010)
2. Li, X., Branke, J., Kieley, M.: On performance metrics and particle swarm methods for
dynamic multiobjective optimization problems. In: IEEE Congress on Evolutionary Computation, pp 576583 (2007)
3. Li, X.: Better spread and convergence: particle swarm multiobjective optimization using
the maximin fitness function. In: Deb, K., Tari, Z. (eds.) GECCO 2004. LNCS, vol. 3102,
pp. 117128. Springer, Heidelberg (2004)

Multiobjective Optimization of Bioactive Compound Extraction Process

21

4. Stanikov, R.B., Matusov, J.B.: Multicriteria Optimization and Engineering. Chapman and
Hall, New York (1995)
5. Auger, A., Bader, J., Brockhoff, D., Zitzler, E.: Theory of the hypervolume indicator: optimal -distributions and the choice of the reference point. In: Proceedings of the Tenth
ACM SIGEVO Workshop on Foundations of Genetic Algorithms, pp 87102 (2009)
6. Grosan, C.: Performance metrics for multiobjective optimization evolutionary algorithms.
In: Proceedings of Conference on Applied and Industrial Mathematics (CAIM), Oradea
(2003)
7. Zitzler, E., Thiele, L.: Multiobjective optimization using evolutionary algorithms - a comparative case study. In: Conference on Parallel Problem Solving from Nature (PPSN V),
pp 292301 (1998)
8. Knowles, J., Corne, D.: Properties of an Adaptive Archiving Algorithm for Storing Nondominated Vectors. IEEE Transactions on Evolutionary Computation 7(2), 100116
(2003)
9. Sandgren, E.: Multicriteria design optimization by goal programming. In: Adeli, H. (ed.)
Advances in Design Optimization, pp. 225265. Chapman & Hall, London (1994)
10. Igel, C., Hansen, N., Roth, S.: Covariance Matrix Adaptation for Multi-objective Optimization. Evolutionary Computation 15(1), 128 (2007)
11. Storn, R., Price, K.V.: Differential evolution a simple and efficient adaptive scheme for
global optimization over continuous spaces, ICSI, Technical Report TR-95-012 (1995)
12. Holland, J.H.: Adaptation in Natural and Artificial Systems: An Introductory Analysis
with Applications to Biology, Control and Artificial Intelligence. MIT Press, USA (1992)
13. Das, I., Dennis, J.E.: Normal-boundary intersection: A new method for generating the
Pareto surface in nonlinear multicriteria optimization problems. SIAM Journal of Optimization 8(3), 631657 (1998)
14. Babu, B.V., Munawar, S.A.: Differential evolution for the optimal design of heat exchangers. In: Proceedings of All-India seminar on Chemical Engineering Progress on Resource
Development: A Vision 2010 and Beyond, Bhuvaneshwar (2000)
15. Babu, B.V., Singh, R.P.: Synthesis & optimization of heat integrated distillation systems
using differential evolution. In: Proceedings of All- India seminar on Chemical Engineering Progress on Resource Development: A Vision 2010 and Beyond, Bhuvaneshwar
(2000)
16. Angira, R., Babu, B.V.: Optimization of non-linear chemical processes using modified differential evolution (MDE). In: Proceedings of the 2nd Indian International Conference on
Artificial Intelligence, Pune, India, pp. 911923 (2005)
17. Ganesan, T., Elamvazuthi, I., Shaari, K.Z.K., Vasant, P.: An Algorithmic Framework for
Multiobjective Optimization. The Scientific World Journal 2013 (2013)
18. Ganesan, T., Elamvazuthi, I., Shaari, K.Z.K., Vasant, P.: Multiobjective Optimization of
Green Sand Mould System Using Chaotic Differential Evolution. In: Gavrilova, M.L.,
Tan, C., Abraham, A. (eds.) Transactions on Computational Science XXI. LNCS, vol.
8160, pp. 145163. Springer, Heidelberg (2013)
19. Elamvazuthi, I., Ganesan, T., Vasant, P.: A comparative study of HNN and Hybrid HNNPSO techniques in the optimization of distributed generation (DG) power systems. In:
2011 International Conference on Advanced Computer Science and Information System
(ICACSIS), pp 195200. IEEE (2011)
20. Mirjalili, S., Lewis, A.: S-shaped versus V-shaped transfer functions for binary Particle
Swarm Optimization. Swarm and Evolutionary Computation 9, 114 (2013)

Anda mungkin juga menyukai