Research Article  Open Access
Omar Abu Arqub, Zaer AboHammour, Shaher Momani, Nabil Shawagfeh, "Solving Singular TwoPoint Boundary Value Problems Using Continuous Genetic Algorithm", Abstract and Applied Analysis, vol. 2012, Article ID 205391, 25 pages, 2012. https://doi.org/10.1155/2012/205391
Solving Singular TwoPoint Boundary Value Problems Using Continuous Genetic Algorithm
Abstract
In this paper, the continuous genetic algorithm is applied for the solution of singular twopoint boundary value problems, where smooth solution curves are used throughout the evolution of the algorithm to obtain the required nodal values. The proposed technique might be considered as a variation of the finite difference method in the sense that each of the derivatives is replaced by an appropriate difference quotient approximation. This novel approach possesses main advantages; it can be applied without any limitation on the nature of the problem, the type of singularity, and the number of mesh points. Numerical examples are included to demonstrate the accuracy, applicability, and generality of the presented technique. The results reveal that the algorithm is very effective, straightforward, and simple.
1. Introduction
Singular boundary value problems (BVPs) for ordinary differential equations arise very frequently in many branches of applied mathematics and physics such as gas dynamics, nuclear physics, chemical reactions, atomic structures, atomic calculations, and study of positive radial solutions of nonlinear elliptic equations (e.g., see [1–3]). In most cases, singular twopoint BVPs do not always have solutions which we can obtain using analytical methods. In fact, many of real physical phenomena encountered, are almost impossible to solve by this technique, these problems must be attacked by various approximate and numerical methods.
The purpose of this letter is to introduce the continuous genetic algorithm (CGA), previously developed by the second author, as an alternative to existing methods in solving singular twopoint BVPs of the form subject to the boundary conditions where is an open or halfopen interval with endpoints and , , , , are real finite constants, and is linear or nonlinear function of and .
When applied to the singular twopoint BVP, standard numerical methods designed for regular BVP suffer from a loss of accuracy or may even fail to converge [4], because of the singularity. However, the finite difference method can be used to solve linear singular twopoint BVPs, but it can be difficult to solve nonlinear singular twopoint BVPs. Furthermore, the finite difference method requires some major modifications that include the use of some rootfinding technique while solving nonlinear singular twopoint BVPs.
Special numerical methods have been proposed to handle the singular problem. To mention a few, in [5], the author has discussed existence and uniqueness of solutions of singular BVP , including the approximation of solutions via finite difference method. In [6], the author has discussed existence and uniqueness of solutions of singular equation and presented variable mesh methods for numerically solving such problems. The homotopy analysis method has been applied to solve the singular equation as described in [7]. Furthermore, the higher order finite difference and cubic spline methods are carried out in [8, 9] for the singular BVP . In [10] also, the authors have provided the fourorder accurate cubic spline method to further investigate the singular equation . Also, the reproducing kernel method for solving the singular BVP is proposed in [11]. Recently, the modified Adomian decomposition method for solving the singular equation is presented in [12].
The reader is kindly requested to go through the survey paper [13] in order to know more details about singular twopoint BVPs. In that paper, the authors introduced various numerical techniques including finite difference, splines, finite element, collocation, variational iteration, and other special approximation methods used in literature followed by own critical comments as remarks for solving linear and nonlinear singular problems. However, in most of the present references, the problems discussed are mostly special cases of the general form (1.1) and (1.2), and there are few valid methods of solving (1.1) and (1.2). Hence, one has to go for nonstandard methods.
CGA (The term “continuous” is used to emphasize that the continuous nature of the optimization problem and the continuity of the resulting solution curves) depends on the evolutions of curves in onedimensional space. The algorithm begins with a population of randomly generated candidates and evolves towards better solution by applying genetic operators which are reproduction, crossover, and mutation. This novel approach is a relatively new class of optimization technique, which generates a growing interest in the mathematics and engineering community. CGA is well suited for a broad range of problems encountered in science and engineering [14–22].
CGA was developed by the second author [14] as an efficient method for the solution of optimization problems in which the parameters to be optimized are correlated with each other or the smoothness of the solution curve must be achieved. It has been successfully applied in the motion planning of robot manipulators, which is a highly nonlinear, coupled problem [15, 16], in the numerical solution of regular twopoint BVPs [17], in the solution of optimal control problems [18], in the solution of collisionfree path planning problem for robot manipulators [19], in the numerical solution of Laplace equation [20], and in the numerical solution of nonlinear regular system of secondorder BVPs [21]. Their novel development has opened the doors for wide applications of the algorithm in the fields of mathematics and engineering. It has been also applied in the solution of fuzzy differential equations [22]. The reader is asked to refer to [14–22] in order to know more details about CGA, including their justification for use, conditions on smoothness of the functions used in the algorithm, several advantages of CGA over conventional GA (discrete version) when it is applied to problems with coupled parameters and/or smooth solution curves, and so forth.
The work presented in this paper is motivated by the needs for a new numerical technique for the solution of singular twopoint BVPs with the following characteristics.(1)It does not require any modification while switching from the linear to the nonlinear case; as a result, it is of versatile nature.(2)This approach does not resort to more advanced mathematical tools; that is, the algorithm should be simple to understand, implement, and should be thus easily accepted in the mathematical and engineering application's fields.(3)The algorithm is of global nature in terms of the solutions obtained as well as its ability to solve other mathematical and engineering problems.However, being a variant of the finite difference scheme with truncation error of the order , the method provides solutions with moderate accuracy.
The organization of the remainder of this paper is as follows: in the next section, we formulate the singular twopoint BVPs. Section 3 covers the description of CGA in detail. Numerical results and discussion are given in Section 4. Finally, concluding remarks are presented in Section 5.
2. Formulation of the Singular TwoPoint BVPs
In this section, (1.1) and (1.2) are first formulated as an optimization problem based on the minimization of the cumulative residual of all unknown interior nodes. After that, a fitness function is introduced in order to convert the minimization problem into a maximization problem.
To approximate the solution of (1.1) and (1.2), we make the stipulation that the mesh points are equally distributed through the interval . This condition is ensured by setting , , where . Thus, at the interior mesh points, , , the equation to be approximated is given as subject to the boundary conditions
The difference quotients approximation formulas, which closely approximate and , , using an point at the interior mesh points with error up to, where and is the order of the derivative can be easily obtained by using Algorithm (6.1) in [23]. For example, based on that algorithm the point formulas, with truncation error of order , for approximating , are given as while the point formulas, with truncation error of order , for approximating , are given as However, it is clear that the first and last equations in (2.3) and (2.4) approximate the first and second derivatives of at the boundary points and where the solutions are known. Thus, they are neglected, and only we use the remaining formulas.
We mention here that the number is starting from 2 and gradually increases up to . To complete the formulation substituting the approximate formulas of and , in (2.1), discretized form of this equation is obtained. The resulting algebraic equations will be a function of , , , , and . After that, it is necessary to rewrite the discretized equation in the following form: The residual of the general interior node, , denoted by , is defined as The overall individual residual,, is a function of the residuals of all interior nodes. It may be stated as
A mapping of the overall individual residual into a fitness function,, is required in the algorithm in order to convert the minimization problem of into a maximization problem of . A suitable fitness function used in this work is defined as The individual fitness is improved if a decrease in the value of the is achieved. The optimal solution of the problem, nodal values, will be achieved when approaches zero and approaches unity.
3. Description of the CGA
In this section, a general review of the GA is presented. After that, a detailed description of the CGA is given. As will be shown later, the efficiency and performance of CGA depend on several factors, including the design of the CGA operators and the settings of the system parameters.
GA is based on principles inspired from the genetic and evolution mechanisms observed in natural systems. Its basic principle is the maintenance of a population of solutions to the problem that evolves towards the global optimum. It is based on the triangle of genetic reproduction, evaluation, and selection [24]. Genetic reproduction is performed by means of two basic genetic operators: crossover and mutation. Evaluation is performed by means of the fitness function that depends on the specific optimization problem. Selection is the mechanism that chooses parent individuals with probability proportional to their relative fitness for the mating process.
The construction of GA for any problem can be separated in five distinct and yet related tasks [24]. First, the genetic representation of potential problem solutions. Second, a method for creating an initial population of solutions. Third, the design of the genetic operators. Fourth, the definition of the fitness function. Fifth, the setting of the system parameters, including the population size, probabilities with which genetic operators are applied, and so forth. Each of the previous components greatly affects the solution obtained as well as the performance of the GA.
The populationbased nature of GA gives it two major advantages over other optimization techniques. First, it identifies the parallel behavior of GA that is realized by a population of simultaneously moving search individuals or candidate solution [24]. Implementation of GA on parallel machines, which significantly reduces the CPU time required, is a major interesting benefit of their implicit parallel nature. Second, information concerning different regions of solution space is passed actively between the individuals by the crossover procedure. This information exchange makes GA an efficient and robust method for optimization, particularly for the optimization of functions of many variables and nonlinear functions. On the other hand, the populationbased nature of GA also results in two main drawbacks. First, more memory space is occupied; that is, instead of using one search vector for the solution, search vectors are used, which represent the population size. Second, GA normally suffers from computational burden when applied on sequential machines. This means that the time required for solving certain problem using GA will be relatively large. However, the solution time is a major point when we are interested in real time applications. But if offline solutions are required for any reallife problem, then our major concern will be the accuracy of the solution rather than the time required for the solution. For reallife problems, the computational time might be reduced to achieve realtime processes utilizing its parallel nature which can be applied on parallel computers or FPGA [18].
The fact that GA uses only objective function information without the need to incorporate highly domainspecific knowledge points to both the simplicity of the approach from one side and its versatility from the other. This means that once a GA is developed to handle a certain problem, it can easily be modified to handle other types of problems by changing the objective function in the existing algorithm. This is why GA is classified as a generalpurpose search strategy. The stochastic behavior of GA cannot be ignored as a main part that gives them much of their search efficiency. GA employs random processes to explore a response surface for a specific optimization problem. The advantage of this behavior is the ability to escape local minima without supervision [18, 25].
The use of CGA in problems with coupled parameters and/or smooth curves needs some justification [14, 17]. First, the discrete initialization version of the initial population means that neighbouring parameters might have opposite extreme values that make the probability of valuable information in this population very limited, and correspondingly the fitness will be very low. This problem is overcome by the use of continuous curves that eliminate the possibility of highly oscillating values among the neighbouring parameters and result in a valuable initial population. Second, the traditional crossover operator results in a jump in the value of the parameter in which the crossover point lies while keeping the other parameters the same or exchanged between the two parents. This discontinuity results in a very slow converging process. On the other hand, the CGA results in smooth transition in the parameter values during the crossover process. Third, the conventional version of the mutation process changes only the value of the parameter in which the mutation occurs while it is necessary to make some global mutations which affect a group of neighbouring parameters since either the parameters are coupled with each other or curve should be smooth. To summarize, the operators of the CGA are of global nature and applied at the individual level, while the operators of the traditional GA are of local nature and applied at the parameter level. As a result, the operators of the traditional GA result in a stepfunctionlike jump in the parameter values while those of CGA result in smooth transitions.
However, when using GA in optimization problems, one should pay attention to two points; first, whether the parameters to be optimized are correlated with each other or not. Second, whether there is some restriction on the smoothness of the resulting solution curve or not. In case of uncorrelated parameters or nonsmooth solution curves, the conventional GA will perform well. On the other hand, if the parameters are correlated with each other or smoothness of the solution curve is a must, then the CGA is preferable in this case [14–22]. The steps of CGA used in this work are as follows.
(1) Initialization: The initialization function used in the algorithm should be smooth from one side and should satisfy constraint boundary conditions from the other side. Two smooth functions that satisfy the boundary conditions are chosen in this work, which include the modified normal gaussian (MNG) function and the modified tangent hyperbolic (MTH) function for each and , where is the th variable value for the th parent, is the ramp function of the th variable value and defined as , is the population size, and are random numbers within the range and , respectively.
The two initialization functions differ from each other by two main criteria: the convex/concave nature and the possibility of any overshoot/undershoot of the concerned function. The MNG function is either convex or concave within the given range of the independent variable while the MTH function is convex in a subinterval of the independent variable and concave in the remaining interval. The MNG function and MTH function, on the other hand, might result in an overshoot or an undershoot, which might exceed the values of the given boundary conditions at some interior mesh points but not at the boundary point as will be shown later. The two initialization functions are multiplied by the corrector function, , which guarantees that the two functions always satisfy the given boundary conditions.
The choice of depends on the boundary conditions and as follows: is any random numbers within the range if differ from zero, within the range if vanished, and within the range if and are both vanished. It is to be noted that for both initialization functions, specifies the amplitude of the corrector function and specifies the degree of dispersion. For small the parameter specifies the center of the MNG function, while specifies the intersection point between the ramp function and the MTH function, which determines the convexity point. The two initialization functions together with the ramp function are shown in Figure 1.
(a)
(b)
The previously mentioned parameters , , and are generated randomly due to the fact that the required solutions are not known for us, and in order to make the initial population as much diverse as we can, randomness should be there to remove any bias toward any solution. The mentioned diversity is the key parameter in having an informationrich initial population. In other cases where one of the boundaries of the solution curves is unknown, the reader is kindly requested to go through [18] for comparison and more details.
(2) Evaluation: the fitness, a nonnegative measure of quality used to reflect the degree of goodness of the individual, is calculated for each individual in the population.
(3) Selection: in the selection process, individuals are chosen from the current population to enter a mating pool devoted to the creation of new individuals for the next generation such that the chance of selection of a given individual for mating is proportional to its relative fitness. This means that the best individuals receive more copies in subsequent generations so that their desirable traits may be passed onto their offspring. This step ensures that the overall quality of the population increases from one generation to the next.
Six selection schemes are incorporated in the algorithm, which include rankbased [26], tournament with replacement [26], tournament without replacement [27], roulette wheel [24], stochastic universal [27], and halfbiased selection [28]. Rankbased selection chooses a prescribed number of parent individuals with the highest fitness according to the rankbased ratio, , and performs the mating process by choosing parents at random from this subpopulation of the size .
In the tournament selection scheme, two individuals are randomly selected from the parent population, and a copy of the individual with the large fitness value, better individual, of the two is replaced in the mating pool. Tournament selection has two forms depending on whether the selection individuals will be placed back into the parent population or not. In a tournament without replacement, the two individuals are set aside for the next selection operation, and they are not replaced into the population until all other individuals have also been removed. Since two individuals are removed from the population for every individual selected, the original population is restored after the mating pool is half filled. The process is repeated for a second round in order to fill the mating pool. In a tournament with replacement, upon the selection of the better individual of the two, both individuals are placed back into the original population for the next selection operation. This step is performed until the mating pool is full. When tournament selection schemes are applied, the number of copies of each individual in the original population cannot be predicted except that it is guaranteed that there will be no copies of the worst individual in the original population.
Roulette wheel selection is a fitness proportionate selection scheme in which the slots of a roulette wheel are sized according to the fitness of each individual in the population. In stochastic universal selection, equidistant markers are placed around the roulette wheel. The number of copies of each individual selected in a single spin of the roulette wheel is equal to the number of markers inside the corresponding slot (the size of slot is still fitness proportional).
Stochastic universal selection guarantees that the number of copies of an individual selected is almost proportional to its fitness, which is not necessarily the case for roulette wheel selection. In halfbiased selection, one mate is selected as in roulette wheel selection, while the other mate is selected randomly from the original population.
(4) Crossover: crossover provides the means by which valuable information is shared among the individuals in the population. It combines the features of two parent individuals, say and , to form two children individuals, say and , that may have new patterns compared to those of their parents and plays a central role in algorithm. The crossover process is expressed as for each , where and represent the two parents chosen from the mating pool, and are the two children obtained through crossover process, and represents the crossover weighting function within the range . The parameters and are as given in the initialization process. Figure 2 shows the crossover process in a solution curve for the two random parents. It is clear that new information is incorporated in the children while maintaining the smoothness of the resulting solution curves.
(a)
(b)
(5) Mutation: the mutation function may be any continuous function within the range such that the mutated child solution curve will start with the solution curve of the child produced through the crossover process and gradually changes its value till it reaches the solution curve of the same child at the other end. Mutation is often introduced to guard against premature convergence. Generally, over a period of several generations, the gene pool tends to become more and more homogeneous. The purpose of mutation is to introduce occasional perturbations to the parameters to maintain genetic diversity within the population. The mutation process is governed by the following formulas: for each and , where represents the th variable value for the th child produced through the crossover process, is the mutated th child for the th variable value, and is the gaussian mutation function. The parameter is as given in the initialization process.
Regarding the mutation center, , and the dispersion factor, , used in the mutation process, three methods are used for generating the mutation center where each method is applied to onethird of the population and two methods are used for generating the dispersion factor where each method is applied to onehalf of the population. The reader is asked to refer to [17] in order to know more details and descriptions about these methods. The mutation process for a random child is shown in Figure 3. As in the crossover process, some new information is incorporated in the mutated child while maintaining the smoothness of the resulting solution curves.
(a)
(b)
(6) Replacement: after generating the offspring's population through the application of the genetic operators to the parents' population, the parents' population is totally or partially replaced by the offspring's population depending on the replacement scheme used. This is known as nonoverlapping, generational, replacement. This completes the “life cycle” of the population.
(7) Termination: the algorithm is terminated when some convergence criterion is met. Possible convergence criteria are as follows: the fitness of the best individual so far found exceeds a threshold value, the maximum nodal residual of the best individual of the population is less than or equals some predefined threshold value, the maximum number of generations is reached, or the improvement in the fitness value of the best member of the population over a specified number of generations is less than some predefined threshold, is reached. After terminating the algorithm, the optimal solution of the problem is the best individual so far found. If the termination conditions are not met, then the algorithm will go back to Step 2.
It is to be noted that the two functions used in the initialization phase of the algorithm will smoothly oscillate between the two ends with a maximum number of single oscillation. If the final solution curves will have more smooth oscillations than one oscillation, then this will be done during the crossover and mutation mechanisms throughout the evolution process. This is actually done by those two operators during the run of the algorithm while solving a problem. However, the evaluation step in the algorithm will automatically decide whether they are rejected or accepted modifications due to their fitness function value.
Two additional operators were introduced to enhance the performance of the CGA, the “elitism” operator, and the “extinction and immigration” operator. These operators are summarized in the form of the following [14–22].
(1) Elitism: elitism is utilized to ensure that the fitness of the best candidate solution in the current population must be larger than or equal to that of the previous population.
(2) Extinction and immigration: this operator is applied when all individuals in the population are identical or when the improvement in the fitness value of the best individual over a certain number of generations is less than some threshold value. This operator consists of two stages; the first stage is the extinction process where all of the individuals in the current generation are removed except the bestofgeneration individual. The second stage is the massimmigration process where the extinct population is filled out again by generating individuals to keep the population size fixed. The generated population is divided into two equal segments each of size; the first segment, with , is generated as in the initialization phase, while the other segment is generated by performing continuous mutation to the bestofgeneration individual as given by the formula for each and , where is the th variable value for the th parent generated using immigration operator, represents the bestofgeneration individual, is the gaussian mutation function, and represents a random number as given in the initialization process.
To summarize the evolution process in CGA an individual is a candidate solution that consists of 1 curve of nodal values. The population of individuals undergoes the selection process, which results in a mating pool among which pairs of individuals are crossed over with probability . This process results in an offspring generation where every child undergoes mutation with probability . After that, the next generation is produced according to the replacement strategy applied. The complete process is repeated till the convergence criterion is met where the parameters of the best individual are the required nodal values. The final goal of discovering the required nodal values is translated into finding the fittest individual in genetic terms.
We mention here the following facts about the previously mentioned parameters , , and : firstly, the value of these parameters can gradually increase or decrease out of the mentioned intervals that are given in the initialization phase, crossover, and mutation mechanisms throughout the evolution process. Secondly, these values are changed from process to process, from generation to generation, and from curve to curve; this is due to the fact that they are generated randomly.
4. Numerical Results and Discussion
In order to evaluate the performance of the proposed CGA, some problems of singular twopoint BVPs are studied. The results obtained by the CGA are compared with the analytical solution of each problem. Results demonstrate that the present method is remarkably effective. The effects of various CGA operators and control parameters on the convergence speed of the proposed algorithm are also investigated in this section. The analysis includes the effect of various initialization methods on the convergence speed of the algorithm in addition to an analysis of the effect of the most commonly used selection schemes, the rankbased ratio, the crossover and mutation probabilities, the population size, the maximum nodal residual, and the step size effect.
The CGA was implemented using visual basic platform. The input data to the algorithm are summarized in Table 1.

Mixed methods for initialization schemes are used where half of the population is generated by the MNG function, while the other half generated using the MTH function. The rankbased selection strategy is used. Generational replacement scheme is applied where the number of elite parents that are passed to the next generation equals onetenth of the population size. Extinction and immigration operator is applied when the improvement in the fitness value of the best individual of the population over 400 generations is less than 0.001. The termination criterion used for each problem is problem dependent and varies from one case to another. However, the CGA is stopped when one of the following conditions is met.(1)The fitness of the best individual of the population reaches a value of 0.999999.(2)The maximum nodal residual of the best individual of the population is less than or equal to 0.00000001.(3)A maximum number of 3000 generations is reached.(4)The improvement in the fitness value of the best individual in the population over 500 generations is less than 0.001.
It is to be noted that the first two conditions indicate a successful termination process (optimal solution is found), while the last two conditions point to a partially successful end depending on the fitness of the best individual in the population (near optimal solution is reached) [14–22]. Due to the stochastic nature of CGA, twelve different runs were made for every result obtained in this work using a different random number generator seed; results are the average values of these runs. This means that each run of the CGA will result in a slight different result from the other runs.
Problem 1. Consider the following linear singular twopoint BVP with singularity at left endpoint: subject to the boundary conditions The exact solution is .
Problem 2. Consider the following nonlinear singular equation with singularities at both endpoints: subject to the boundary conditions The exact solution is .
Problem 3. Consider the following nonlinear singular equation with singularities at both endpoints: subject to the boundary conditions The exact solution is .
Problem 4. Consider the following nonlinear singular twopoint BVP with singularities at both endpoints: subject to the boundary conditions The exact solution is .
Throughout this paper, we will try to give the results of the four problems; however, in some cases we will switch between the results obtained for the problems in order not to increase the length of the paper without the loss of generality for the remaining problems and results. The convergence speed of the algorithm, whenever used, means the average number of generations required for convergence. The step size for the four problems is fixed at 0.1, and thus, the number of interior nodes equals 9 for all problems.
The convergence data of the four problems is given in Table 2. It is clear from the table that the problems take about 1086 generations, on average, within about 229.13 seconds to converge to a fitness value of 0.99999653 with an average absolute nodal residual of the value and an average absolute difference between the exact values and the values obtained using CGA of the value .

The detailed data of the four problems that includes the exact nodal values, the CGA nodal values, the absolute error, and the absolute nodal residuals is given in Tables 3, 4, 5, and 6, respectively. It is clear that the accuracy obtained using CGA is moderate since it has a truncation error of the order .




The evolutionary progress plots of the bestfitness individual of the four problems are shown in Figures 4 and 5. It is clear from the figures that in the first of generations the best fitness approaches to one very fast; after that, it approaches to one slower. That means the approximate of CGA converges to the actual solution very fast in the first of the generations.
(a)
(b)
(a)
(b)
The way in which the nodal values evolve for Problems 1 and 4 is studied next. Figure 6 shows the evolution of the first, , middle, , and ninth, , nodal values for Problem 1 while Figure 7 shows the evolution of the second, , middle, , and eighth, , nodal values for Problem 4. It is observed that from the evolutionary plots that the convergence process is divided into two stages: the coarsetuning stage and the finetuning stage, where the coarsetuning stage is the initial stage in which oscillations in the evolutionary plots occur, while the finetuning stage is the final stage in which the evolutionary plots reach steadystate values and do not have oscillations by usual inspection. In other words, evolution has initial oscillatory nature for all nodes, in the same problem. As a result, all nodes, in the same problem, reach the near optimal solution together. The average percentage of the finetuning stage till convergence from the total number of generations across all nodes of the four problems is given in Table 7. It is clear from the table that the problems spent about of generations, on average, in the coarsetuning stage, while the remaining is spent in the finetuning stage.
(a)
(b)
(c)
(a)
(b)
(c)
The effect of the different types of initialization methods on the convergence speed of the algorithm is discussed next. Three initialization methods are investigated in this work; the first method uses the MNG function, the second uses the MTH function, while the third is the mixedtype initialization method that initializes the first half of the population using the MNG function and the second half of the population using the MTH function. Table 8 shows that the used initialization method has a minor effect on the convergence speed because usually the effect of the initial population dies after few tens of generations and the convergence speed after that is governed by the selection mechanism, crossover, and mutation operators. For Problems 1, 2, and 3, the MNG function results in the fastest convergence speed while for Problem 4, the mixedtype initialization method results in the fastest convergence speed. For a specific problem, the initialization method with the highest convergence speed is the one that provides initial solution curves which are close to the optimal solution of that problem; that is, the optimal solution of the Problems 1, 2, and 3 is close to the MNG function and so on. However, since the optimal solution of any given problem is not assumed to be known, it is better to have a diverse initial population by the use of the mixedtype initialization method. As a result, the mixedtype initialization method is used as the algorithm default method [14–22].
The effect of the most commonly used selection schemes by GA community of the performance on the CGA is explored next. Table 9 represents the convergence speed using the six selection schemes previously described. It is clear from the table that the rankbased selection scheme has the faster convergence speed for all problems. The tournament selection (with and without replacement) approaches come in the second place with almost similar convergence speeds. It is obvious that the fitness proportionate methods (i.e., roulette wheel, stochastic universal, and halfbiased selection schemes) have slower convergence speed of the rest of the methods. The halfbiased selection scheme has the slowest convergence speed.
The effect of the rankbased ratio, , on the convergence speed is studied next. Table 10 gives the convergence speed of the algorithm for different value within the range . It is clear that results in the best convergence speed for all problems. Furthermore, it is observed that the average number of generations required for convergence increases as the ratio increases.
The effect of the vector norm used in the fitness evaluation is studied here. Two vector norms are used: norm and norm. norm is governed by the following equation: while norm is governed by (2.7). Figure 8 shows the evolutionary progress plots for the bestofgeneration individual for Problems 2 and 3 using and norms while Table 11 gives the convergence speed for the four problems. Two observations are made in this regard; first, the evolutionary progress plots of both norms show that norm has higher fitness values than those of norm throughout the evolution process. Second, norm converges a little bit faster than norm. The key factor behind these observations is the square power appearing in norm. Regarding the first observation, it is known that for a given set of nodal residuals with values less than 1, norm results in a higher value than norm, and correspondingly, the fitness value using norm will be higher than that using norm. Regarding the second observation, norm tries to select individual solutions, vectors, with distributed nodal residuals among the nodes rather than lumped nodal residuals where one nodal residual is high and the remaining nodal residuals are relatively small. This distributed selection scheme results in closer solutions to the optimal one than the lumped selection scheme. In addition to that, the crossover operator will be more effective in the former case than the latter one. These two points result in the faster convergence speed in norm as compared with norm. Furthermore, it is observed that norm is less sensitive to variations in the genetic related parameters and problem related parameters. As a result, norm is preferred over norm, and it is used as the algorithm’s default norm [14–22].
(a)
(b)
The particular settings of several CGA tuning parameters including the probabilities of applying crossover operator and mutation operator are investigated here. These tuning parameters are typically problem dependent and have to be determined experimentally. They play a nonnegligible role in the improvement of the efficiency of the algorithm. Table 12 shows the effect of the crossover probability, , and the mutation probability, , on the convergence speed of the algorithm for Problem 1. The probability value is increased in steps of starting with and ending with for both and . It is clear from the tables that when the probabilities values