Solving System of Non-Linear Equations using

Genetic Algorithm
Gopesh Joshi M. Bala Krishna*
University School of Information and University School of Information and
Communication Technology Communication Technology
G G S Indraprastha University G G S Indraprastha University
New Delhi, INDIA New Delhi, INDIA
E-Mail: [email protected] E-Mail: [email protected]

Abstract— Solving system of non-linear equation is a general and represented as follows:

purpose problem that can be applied to real-time applications.
Although there are several methods like Newton-Raphson F(x) : f1(x1, x2, x3, xm) = 0
method, secant method, bisection method, fixed-point iteration
f2(x1, x2, x3, xm) = 0
etc. evolutionary and soft computing methods optimize the space
and time complexity to solve nonlinear systems. In this article, we
f3(x1, x2, x3, xm) = 0 (1)
present the Genetic Algorithm approach to solve nonlinear system ……………………..
of equations for travelling salesman problem, tank reactor system fn(x1, x2, x3, xm) = 0
and neurophysiology application. The roots of nonlinear system
using GA is estimated using population size, degree of mutation, Where x is a vector {x1, x2, x3,…, xm}consisting of m variables.
crossover rate and coefficient size. Nonlinear parameters are If m is equal to n, then the system is known as square system. A
analyzed and simulated for multi-dimensional NP hard travelling simple square non-linear system with two variables (x, y),
salesman problem, two dimensional tank reactor and functions (f1, f2) and constants (c1 to c4) is given as follows:
neurophysiology application.
f1(x, y) : x2 + c1*y + c2 = 0 (2)
Keywords—Nonlinear equations, Genetic Algorithm, Crossover
f2(x, y) : y2 + c3*x - c4 = 0 (3)
rate, Degree of Mutation
As the size of system increases, the complexity in Newton and
I. INTRODUCTION
secant method [6] [7] increases exponentially. Hence, GA
approach would be a better choice for nonlinear resource
System of non-linear equations varies with respect to variables, constrained applications.
degree of variables and coefficients in the system. Solving a
system of linear equations is much simpler as compared to A.1. Types of Non-linear System
non-linear equations. System of non-linear equations is applied
in medical science, engineering applications, financial market Non-linear system is classified as follows:
analysis, network traffic analysis, etc. Multi-objective
• Univariate Function System (UFS)
optimization is the process of minimizing or maximizing the
• Multivariate Function System (MFS)
vector F(x), which is a collection of objective functions.
Non-linear system mapped to multi-objective function is solved
In UFS only one variable with varying degree is considered in
using Newton method, Secant method, Quasi-Newton method,
the design. Consider the following UFS model with one
etc. Genetic Algorithm (GA) [1] [2] [3], a meta-heuristic
variable x:
optimization technique with multi-objective genetic variant
functions solves the roots of nonlinear system. Dynamic
x2+2x+5 = 0 (4)
features of artificial immune system [4] solve the time varying
x +5x2+2x-10 = 0
3
(5)
non-linear constraints in multi-objective functions.
Since, the degree of polynomial is two and three multiple
A. System of nonlinear equation
solutions are possible for the above equations. In MFS multiple
variables are considered, and multiplicity of roots are derived.
An efficient genetic algorithm [5] for solving nonlinear
Consider the following MFS model with two variables x and y:
equations use a pair of symmetric and harmonic individuals to
achieve best fitness value and converge at optimal solution.
x2+y2 = 0 (6)
System of non-linear equations F(x) is vector of n equations [5]
x2+xy+2y2-10 = 0 (7)

978-1-4799-3080-7/14/$31.00 2014
c IEEE 1302
Here the two variables x and y are real numbers within the range schemes. A novel GA approach [11] solves the optimization
[xmin, xmax] and [ymin, ymax] problem by using the diversity operator in multi-parent
crossover and achieves better solution.
A.2. Mapping to Multi-Objective Optimization
This article is organized as follows. Section II presents the
Due to similarities between the system of non-linear equations literature survey of linear and non-linear multi-objective
and Multi-Objective Optimization (MOP) [2], the functions f1 dynamic systems. Section III presents the procedure to find the
to fn are mapped as follows: roots of non-linear system using GA approach. Section IV gives
F(x) = min (abs (f1(x))) the simulation results for travelling salesman problem, tank
min (abs (f2(x))) reactor system and neurophysiology applications. Simulation
min (abs (f3(x))) (8) results are based on comparative analysis of insert operator and
…………………... inverse operator for each of the above nonlinear system.
min (abs (fn[x))) Section V concludes our achievements and gives the direction
Here each function fi is a combination of exponential, for future extension.
trigonometric, logarithmic and complex functions.
II. RELATED WORK
B. Genetic algorithm
Clever algorithms [12] explain various stochastic, evolutionary,
Genetic Algorithm (GA) is an evolutionary based soft swarm intelligence, immune systems, and neural network
computing technique [8] that uses the scheme “survival of the algorithms that can be used in linear and non-linear
fittest.” Figure 1 illustrates the sequence of events performed in multi-objective dynamic systems. Various approaches to solve
Genetic Algorithm. In this approach, the solution is evolved nonlinear equations are given as follows:
multiple times so that the final result will be either global
maximum or global minimum. Simulated Annealing: This method uses guided random search
to find the global optima with several local optima. Practical
Encoding Initialize Population approach of simulated annealing is the process of slow cooling
used in hot metals. This results in appropriate ordering of metal
crystals and lower its defects in the final production. Thus, the
Evaluate Fitness state of metal changes from high energy to low energy (ground
state) and lead to desired solution.
Hill Climbing: This method is a local search technique that uses
Selection
the greedy approach to find the local optima solution. Initially,
the algorithm selects two points, the current point and the next
point. If the next point converges towards the local solution as
Crossover compared to the current point, then the current point is updated,
else the search process for a new point is performed.
Newton method: This classical method uses Jacobian matrix to
Mutation solve the system of nonlinear equations. Initially, one value of
the root starts the iterative process and finds the zeros in
differential equations based on multiplicative inverse of
No Jacobian matrix.
End Trust region dogleg method: The trust region method is an
extension of Newton-Raphson method. Variations of trust
Ye region methods are 2D subspace minimization, dogleg method,
and nearly exact solution. The merit function in trust region
Decoding Best Individual dogleg method decides the quality of solution and uses the
Powell dogleg procedure [2] to calculate the step dk.
Figure 1. Flow of modules in Genetic Algorithm Ant Colony Optimization: This method is inspired by the
behavior of ants for the search of food source. The ants start
The main operations in GA are selection, crossover and searching for food in multiple directions by laying a chemical
mutation [8] [9]. Initial step performs encoding, initialization, substance known as pheromone. Ants identify the food source
and the final step performs the decoding solution. Primary based on the intensity of pheromone count laid by multiple ants.
component of GA is the evaluation criteria of fitness function. Evolutionary algorithms [3] for solving large scale nonlinear
Optimal inverse system of feedback control in non-linear constrained problems use smoothing and balancing techniques
systems [10] is solved using effective genetic programming

2014 International Conference on Advances in Computing,Communications and Informatics (ICACCI) 1303

to estimate the errors and adapt step-wise refinement to a. Evaluate the fitness function fi for each
converge towards the precise local solution. chromosome in the population.
Neural Network: This method uses the working of neurons and b. Calculate the probability (that is the size of slot
trains the network based on supervised and unsupervised wheel) as follows:
learning methods [13]. The trained data is fed to the system Pi = fi / ™fi
until the desired output is achieved. GA using neural network c. Evaluate the cumulative probability as:
approach [13] helps in understanding the process with Ci=™ Pi+Ci-1
additional constraints and training sets. The generated d. Generate a random number r between 0 and 1
population is fed as input to the system, and the best generated e. If r is less than Ci then select first individual, else
sample is achieved at the output.
if r is greater than Ci but less than or equal to Ci+1
Linear and Non-Linear System: Linear system satisfies two
then select the i+1 individual.
principle conditions namely, superposition and homogeneity.
6. Execute the crossover operator using random pick of
The law of superposition for two inputs x and y is defined as:
two chromosomes from the parent pool. The generated
random number r between 0 and 1 is compared with the
f( x + y ) = f(x) + f(y) (9)
crossover rate. If r > crossover rate, the parent genes
are copied to the child genes, else the crossover
The law of homogeneity for two input x with scalar m is defined
operator is executed iteratively. The three principle
as:
crossover methods used in GA are one-point, two-point
f(xm )=m*f(x) (10)
and uniform methods.
7. The mutation rate (user input) is kept low since the
The system that doesn’t satisfy these conditions is termed as chance of variation in inheritance will be low. Each
non-linear system. mutant is randomly selected, and the product of
population size, chromosome size and mutation rate
III. PROCEDURE ADOPTED FOR FINDING THE ROOTS OF [14] is evaluated. The inverse mutation is calculated.
NONLINEAR SYSTEM 8. Steps 5-7 are iterated until the number of generations is
maintained, and the best solution is achieved.
GA approach to solve the nonlinear behavior of classical NP The weed optimization approach [15] for solving the nonlinear
hard travelling salesman combinatorial optimization problem is equation uses two-phase root finder technique to estimate the
based on the availability of chromosome length and set of real and complex roots of the system.
genetic operators. This section motivates to solve the complex
behavior of nonlinear system [1] using GA approach. Nonlinear IV. SIMULATIONS
TSP system is analyzed using the crossover and mutation
operators. Nonlinear highly complex tank reactor system with
varying degrees is tested using GA method, traditional Newton A. Example 1: Travelling Salesman Problem
method and Trust region dogleg method. Finally, the nonlinear Travelling Salesman Problem (TSP) is an NP-hard problem in
neurophysiology application is tested using GA approach with which the salesman moves around the set of cities such that total
mutation operator. The global optimization GA method refers tour cost is minimum. The proposed GA approach considers the
to overall system optimization. The sequence of steps for chromosome attribute as the set of cities, such that the tour cost
finding the roots of nonlinear system using GA approach is from one city to next city is the product of chromosome genes.
given as follows: Standard TSP approach uses symmetric rule and the sum of
objective function [16] is given as:
1. Encode the roots of equation into chromosomes. n −1
2. Instead of minimizing each equation, the sum of F = ¦ xi * xi +1 + xn−1 * x1 (11)
non-linear equation is evaluated as: i =1
F = f1 + f2 + f3 +…. + fn Where n is the number of cities, and the degree of nonlinear
If each objective function is minimum then the sum of system is configured to two. The crossover operator can be
objective functions would be minimum. implemented as follows:
3. Evaluate the fitness of whole system based on initial • 1-point Crossover
solution F. • 2-point Crossover
4. Select the individuals that generate next solution. • Uniform Crossover
5. In roulette wheel selection method [14], the individuals One-point crossover method selects a single random gene
with high fitness probability are selected and the rest are from the chromosome set and this will be the same for
eliminated. This is done by the imaginary wheel randomly selected parents. The exchange of genes in
containing the search space and functions as follows: crossover is similar to the genes exchange in meiosis process

1304 2014 International Conference on Advances in Computing,Communications and Informatics (ICACCI)

of biology. This results in two child chromosomes that will Table 2. Average analysis of TSP using nonlinear GA
replace their parent. Second-point crossover method selects
two points randomly such that the middle section of genes Number of Tour Time
will exchange the chromosome set. Uniform crossover Executions Cost (sec)
Run 1 32794 15.05
method exchange is even for one parent and odd for other, so
Run 2 33758 15.89
that the child genes are in alternate order of their parent
Run 3 32380 21.40
chromosomes. This method indicates improvement in GA as
Run 4 32660 16.68
compared to one-point and two-point approaches. The GA
Run 5 32170 24.40
parameters used in this model is given in table 1. Table 2
Average 32752.4 18.68
shows the average analysis of TSP based on tour cost and
traversal time.

Table 1. TSP Parameters used in non-linear GA

GA Parameters Value
Population Size 100
Crossover rate 0.8
Mutation rate 0.001
Number of Generations 400

Figure 4. Comparisons of crossover operator for TSP using GA

B. Example 2: Tank reactor system for two-dimensional

non-linear system
Tank reactor [17] system is a series of two non-adiabatic stirred
tanks. Both reactors at steady state are represented using two
dimensional exothermic irreversible equations. Representative
models of tanks reactor [18] [19] [20] is given as follows:

Figure 2. Comparisons of mutation operator for TSP using GA

§ 10Φ1 ·
f 1 = (1 − R ) [ D 10(1+ β 1) − Φ1]exp ¨ ¸ − Φ1 (12)
© 1 + 10Φ1 / γ ¹
§ 10Φ 2 ·
f 1 = (1 − R )[ D 10 − β 1Φ1 − (1 + β 2) Φ 2]exp ¨ ¸+
© 1 + 10Φ 2 / γ ¹
Φ1 − (1 + β 2)Φ 2
(13)
The constant ȕ1, ȕ2, Ȗ and D are set to 2, 2, 1000 and 22
respectively. Two variables ߔ 1 and ߔ 2 indicate temperature in
the range [0, 1]. Table 3 illustrates the parameters used in
non-linear GA.

Table 3. Tank Reactor Parameters used in non-linear GA

Figure 3. Time comparisons of mutation operator for TSP using GA
GA Parameters Value
Figure 2 illustrates better results for insert mutation operator as Population Size 100
compared to inverse mutation operator. Figure 3 indicates that Crossover rate 0.8
average execution time for insert mutation operator is less than Mutation rate 0.01
the inverse mutation operator. Simulations in figure 4 indicate Number of Generations 400
the crossover comparisons of nonlinear TSP. Clearly, the
uniform crossover method is best suited for TSP problem as Newton method and Trust Region Dogleg are used to simulate
compared to one-point and two-point crossover methods. the two-dimensional non-linear tank reactor system.

2014 International Conference on Advances in Computing,Communications and Informatics (ICACCI) 1305

B.1. Newton method
Figure 6 illustrates that the inverse mutation is more unstable
The simplest method of solving nonlinear system using Newton than the insert mutation. Hence, insert mutation operator is the
method is based on first order derivative resulting in Jacobian better choice to solve nonlinear systems. Figure 7 illustrates the
matrix J. This method represent each function as LHS-RHS average time comparisons of inverse mutation and insert
matrix f such that the dimension is n*m, where n is number of mutation. Clearly, inverse mutation takes more time to solve
equations (dimensions of matrix are n: rows and m: column). the nonlinear system as compared to insert mutation.
The derivative of each equation using independent variable is
evaluated as follows:
xnew = xold − ( J −1 * f ) (14)
In the next step the inverse of Jacobian matrix is evaluated,
where the xold vector consists of previous independent
variables and xnew vector consists of the current independent
variables.

B.2. Trust Region Dogleg

Trust region dogleg is a subclass of trust region algorithms that

overcomes the complexities in Newton method. The Jacobian Figure 6. Comparison of mutation operator for tank reactor system
matrix is singular, and the computational step size is cost using GA
effective in Newton method. Trust region method a modified
version of Newton method solves two-dimensional nonlinear
system where the step size is computed by Cauchy step and
Gauss-Newton method. This method is more robust and
achieves the desired results.

Table 4. Comparative analysis of Tank Reactor system using Newton,

GA and Trust Region approaches

Method Recycle Ratio (R)

0.935 0.945 0.955 0.965 0.975 0.985
Newton 25.56 21.70 17.84 13.98 10.14 6.29
GA 0.766 1.43 1.364 1.72 2.93 1.51
Trust region 0.301 0.28 1.001 0.23 0.17 0.52 Figure 7. Time comparisons of mutation operator for tank reactor
dogleg system using GA

Table 4 illustrates the comparisons of recycle ratio in tank C. Example 3: Neurophysiology application for
reactor using Newton, GA and trust region dogleg methods. two-dimensional non-linear system
Figure 5 illustrates the comparisons of nonlinear tank reactor We consider the risk equations given in [7] [21] [22] for
system. neurophysiology application. This non-linear system with six
variable is given as follows:
x12 + x32 = 1
x22 + x42 = 1
x5*x33 + x6*x43 = c1
x5 *x13 + x6*x23 = c2 (15)
x5 *x1* x32 + x6*x2*x42 = c3
x5 *x12 *x3 + x6*x22 *x4 = c4
Here c1, c2, c3 and c4 are constants and set to zero. The degree of
non-linear system is set to 1024 and the risk equations represent
the objective function. Threshold to depict the safe state of the
Figure 5. Comparative analysis of Newton, GA and Trust region system is [-2000, +2000] where the variable xi is defined within
approaches used in nonlinear tank reactor system
the range [-10, 10].

1306 2014 International Conference on Advances in Computing,Communications and Informatics (ICACCI)

 Table 5. Neurophysiology Parameters used in non-linear GA Analysis of Nonlinear Systems
GA Parameters Value
Population Size 100
Space complexity: Space complexity is estimated from the
Crossover rate 0.8 search space, and the number of possible solutions is the
Mutation rate 0.01 product of generation size (G) and population size (P). Hence,
Number of Generations 400 the search space complexity O(S) = O(G*P). Since the tank
reactor system and neurophysiology application consider real
Table 5 illustrates the parameters used in non-linear GA. Two numbers and evaluate the gene chromosomes, the simulation
ways to implement the mutation operator is (i) insert mutation analysis reveal more search space for respective nonlinear
and (ii) inverse mutation systems as compared to TSP system.

Time complexity: Time complexity of GA method is:

O(G*(O(Fitness)+O(Selection)+O(Crossover)+O(Mutation))).
The complexity of the fitness function for individual size is N.
For TSP system, if number of genes is N (individual size), then
there will be N multiplications and N-1 additions for each
iteration per individual. So, the total computational complexity
of TSP fitness function is O(G*P*(2*N-1))=O(G*P*N). The
selection process takes constant time O(1), the crossover and
mutation process takes O(G*P*N) time. Hence, the total
execution time is O(G*P*N). Similarly, for tank reactor system
and neurophysiology application the time complexity is
O(G*P*N).
Figure 8. Comparisons of mutation operator for neurophysiology
application using GA
V. CONCLUSIONS
In insert mutation, a random value is inserted at any selected
point, and the total number of genes is calculated for a given This article explains the possible solutions to solve nonlinear
population using mutation probability. Inverse mutation does equations using GA approach. GA approach optimizes the
not allow any outside member, instead it reverses the selected space and time complexities in solving the nonlinear system as
chromosome, so that the gene ordering will change. Figure 8 compared to the traditional computing methods. GA approach
illustrates that the sum of objective function is less for insert uses mutation rate, crossover rate and degree of coefficient to
mutation as compared to inverse mutation. Hence, the insert evaluate the roots of nonlinear system. In this article, we use
mutation is better suited for nonlinear neurophysiology system. GA approach to solve the nonlinear equations for
Figure 9 illustrates that the average time comparisons of inverse multi-dimensional NP-hard travelling salesman problem,
mutation and insert mutation. Clearly, inverse mutation takes two-dimensional tank reactor system and two-dimensional
more time to solve the nonlinear system as compared to insert neurophysiology application. This work can be further
mutation. Hence, insert mutation and uniform crossover extended by using neural network approach to solve nonlinear
methods are selected in GA approach to solve the nonlinear systems. Further, fuzzy logic and meta-heuristic methods need
system. to be incorporated in the existing GA methods and solve the
complexity in nonlinear systems.

ACKNOWLEDGEMENT

We thank the faculties of USICT for their kind support. We also

would thank the USICT Lab team and the Library team for
facilitating us with resources and constant support.

REFERENCES

[1] Antonio R, Manuel V and Jesús C, A new methodology to

solve non-linear equation systems using genetic algorithms -
Figure 9.Time comparisons of mutation operator for neurophysiology Application to combined cycle gas turbine simulation, Wiley
application using GA International Journal for Numerical Methods in Engineering,
14 July 2005, 63(10), pp. 1424–1435.

2014 International Conference on Advances in Computing,Communications and Informatics (ICACCI) 1307

[2] Carlos A. Coello C, Gary B. L and David A. Van V, [16] Otman A, Jaafar A and Chakir T, Analyzing the performance
Evolutionary Algorithms for Solving Multi-Objective of mutation operators to solve the travelling salesman problem,
Problems, Second Edition, In Springer Genetic and International Journal of Emerging Sciences, Cornell
Evolutionary Computation Series, David E. Goldberg and John University Library, Computer Science, Neural and
R. Koza (Eds.), Springer Science+Business Media, LLC, 233 Evolutionary Computing, March 2012, 2(1), pp. 1-18.
Spring Street, New York, USA, 2007. [17] Milan K, Hanns H, Vladimír H and JiĜí S, Multiplicity and
[3] Janusz O and Maciej G, On Dedicated Evolutionary stability in a sequence of two nonadiabatic nonisothermal
Algorithms for Large Non-Linear Constrained Optimization CSTR, Elsevier Journal of Chemical Engineering Science,
Problems, In Proceedings of ACM International Genetic and 1980, 35(4), pp. 987-996.
Evolutionary Computation Conference (GECCO), Vancouver, [18] Christodoulos A F, Recent advances in global optimization for
BC, Canada, 12–16 July 2014, pp. 121-122. process synthesis, design, and control: enclosure of all
[4] Zhuhong Z and Shuqu Q, Artificial immune system in dynamic solutions, In Proceedings of Elsevier European Symposium on
environments solving time-varying non-linear constrained Computer Aided Process Engineering, Computers & Chemical
multi-objective problems, Springer Journal of Soft Computing, Engineering, Volume 23, Supplement, 1 June 1999,
July 2011, 15(7), pp. 1333–1349. pp. 963–973.
[5] Hongmin R, Long W, Weihong B and Ioannis K A, Solving [19] Christodoulos A F, Panos P and Claire A, et al, Handbook of
nonlinear equations system via an efficient genetic algorithm test problems in local and global optimization, Springer
with symmetric and harmonious individuals, Elsevier Journal Nonconvex Optimization and Its Applications, vol. 33, 1999.
of Applied Mathematics and Computation, 1 August 2013, [20] Michael J H, Panos M P and Mauricio G C R, Solving systems
219(23), pp. 10967–10973. of nonlinear equations with continuous GRASP, Elsevier
[6] Grosan C and Abraham A, Solving nonlinear equation systems Journal of Nonlinear Analysis: Real World Applications,
using evolutionary algorithms (GECCO 2006 Late Breaking August 2009, 10(4), pp. 2000-2006.
Paper), In Proceedings of ACM International Genetic and [21] Angel Fernando K M, Solution of Simultaneous Non-Linear
Evolutionary Computation Conference (GECCO), Seattle, Equations using Genetic Algorithms, WSEAS Transactions on
WA, USA, 8-12 July 2006, pp. 1-4. Systems, January 2003, 2(1), pp. 44-51.
[7] Grosan C and Abraham A, A new approach for solving [22] Jan V, Pierre V and Ronald C, Homotopies exploiting newton
nonlinear equations systems, IEEE Transactions on Systems, polytopes for solving sparse polynomial systems, SIAM
Man and Cybernetics, Part A: Systems and Humans, Journal on Numerical Analysis, June 1994, 31(3), pp. 915-930.
May 2008, 38(3), pp. 698-714.
[8] Man K F, Tang K S and Kwong S, Genetic algorithms:
concepts and applications, IEEE Transactions on Industrial
Electronics, October 1996, 43(5), pp. 519-534.
[9] Kumara S, David G and Graham K, Chapter 4. Genetic
algorithms, In Springer Search Methodologies: Introductory
Tutorials in Optimization and Decision Support Techniques,
Edmund K. Burke and Graham Kendall (Eds.), 2005,
pp. 97-125.
[10] Kobayashi T, Kimoto T and Imae J, Genetic Programming
Based Output Regulation via Optimal Inversion for Nonlinear
Systems, In Proceedings of IEEE/SICE International
Symposium on System Integration (SII), Kyoto University,
Kyoto, Japan, 20-22 December 2011, pp. 398-403.
[11] Saber M E, Ruhul A S and Daryl L E, A new genetic algorithm
for solving optimization problems, Elsevier Journal of
Engineering Applications of Artificial Intelligence,
January 2014, vol. 27, pp. 57–69.
[12] Jason B, Clever Algorithms: Nature-Inspired Programming
Recipes, Lulu Publishers, Creative Commons (License),
Australia, 2011.
[13] Coehn P, Combining genetic algorithms and neural networks:
the encoding problem: The Encoding Problem, Master of
Science Thesis, The University of Tennessee, Knoxville,
December 1994.
[14] Hermawanto D, Genetic algorithm for solving simple
mathematical equality problem, Cornell University Library,
Computer Science, Neural and Evolutionary Computing,
August 2013, pp. 1-10.
[15] Ebrahim P and Hamed M, Solving nonlinear equations systems
with a new approach based on invasive weed optimization
algorithm and clustering, Elsevier Journal of Swarm and
Evolutionary Computation, June 2012, vol. 4, pp. 33–43.

1308 2014 International Conference on Advances in Computing,Communications and Informatics (ICACCI)