A Novel Accurate Genetic Algorithm for Multivariable Systems

Transcription

1 World Applied Sciences Journal 5 (): , 008 ISSN IDOSI Publications, 008 A Novel Accurate Genetic Algorithm or Multivariable Systems Abdorreza Alavi Gharahbagh and Vahid Abolghasemi Department o Electronic and Computer, Islamic Azad University, Shahrood Branch, Shahrood, Iran Abstract: In this study a modiied and accurate Genetic Algorithm (GA) is proposed. Uncertain nature o probability o mutation (P mute ), in general GA methods, causes diiculty in deining this parameter. On the other hand there is always a critical unknown value or P mute. Choosing P mute above the critical value leads to high and also constant accuracy and is required high number o iterations, while smaller P mute values result in low accuracy and less iteration. There are ew values or this parameter that guarantee answer accuracy with low iteration. In order to soten this problem we propose the modiied mutation method. The proposed method has two major advantages in compare with simple mutation; irst P mute in this method is replaced with two new parameters (P start and De rate ) which alleviate the problem o P mute tuning. Second the answer accuracy in this algorithm is high but not constant similar to simple mutation methods. The maximum value o six multivariable sample unctions is computed using both general and the proposed methods. Two important parameters, convergence rate and answer accuracy, are considered as actors to compare these methods. The results conirm eectiveness and robustness o the proposed method against general methods. Key words : Genetic algorithm Mutation Crossover Convergance rate Accuracy INTRODUCTION John Holland's pioneering book Adaptation in Natural and Artiicial Systems [1] showed how the evolutionary process can be applied to solve a wide variety o problems using a highly parallel technique that is now called the Genetic Algorithm (GA). The genetic algorithm transorms a population o individual objects, each with an associated itness value, into a new generation o the population. It is based on Darwinian principle o reproduction and survival o naturally occurring genetic operations such as crossover and mutation. The genetic algorithm attempts to ind an optimum (or best) solution to the problem by genetically breeding the population o individuals over a series o generations. The two important parameters in GA methods are accuracy and rate o convergence. A good GA method needs low iteration (ast convergence) or getting accurate answer. The relative merits o crossover, mutation and other genetic operators have long been debated in the literature o genetic algorithms. Without mutation in complex systems, answer accuracy is poor. Also crossover and itness aect on number o iterations. Several researchers have been presented theoretical arguments that show mutation can be more useul than was previously thought. The empirical results conirm the strength o mutation [-4]. Others have produced new arguments in avor o crossover [5]. A major problem in using mutation is how to choose the Probability o mutation (P mute ). The uncertain nature o this parameter is a disadvantage. There is a critical unknown value or P mute. Choosing P mute above the critical value lead to high and also constant accuracy and the number o iterations increase, while smaller P mute values result low accuracy and ewer iterations. There are ew values or this parameter that guarantee answer accuracy with low iteration. In order to soten this problem we propose the modiied mutation method. The proposed method has two major advantages in compare with simple mutation; irst P mute in this method is replaced with two new parameters (P mute and De rate ) which alleviate the problem o P mute tuning. Second the answer accuracy in this algorithm is high but not constant similar to simple mutation methods. GA GENERAL METHODS We can divide methods by three main operators: selection, crossover and mutation. Corresponding Author: Dr. Abdorreza Alavi Gharahbagh, Deparment o Electronic Islamic, Azad University o Shahrood, Iran 137

2 Selection Initial samples selection Randomly create an initial samples in search space. Create a sequential initial samples in search space. Selection scheme in algorithm The probability to choose a certain sample is proportional to its itness. Algorithm at last is permit to select N/ samples rom N initial samples (Roulette wheel). Algorithm chooses N/ samples with better itness and discards other samples (Ideal selection). The probability to choose a certain sample is proportional to its itness but i the sample with best itness discards, algorithm replaces this sample with one o selected samples and discards it (Elitism). Cross over One-point crossover: two strings cut at a randomly chosen position and swapping the two tails. Onepoint crossover is a simple method or GAs. N-point crossover: Instead o only one, N breaking points are chosen randomly. Every second section is swapped. Segmented crossover: Similar to N-point crossover with the dierence that the number o breaking points can vary. Uniorm crossover: For each position, it is decided randomly i the positions are swapped. Shule crossover: First a randomly chosen permutation is applied to the two parents and then N-point crossover is applied to the shuled parents. Mutation Inversion o single bits: With probability P mute, one randomly chosen bit is negated. Bitwise inversion: The whole string is inverted bit by bit with probability P mute Random selection: With probability P mute, the string is replaced by a randomly chosen one. Any combination o these operator types makes a GA method. In practice, a desired GA method rapidly and eectively optimizes complex, highly nonlinear, multidimensional systems. A desired GA method should be aster than other methods and more precise. World Appl. Sci. J., 5 (): , Among these operators, deining mutation is more crucial than others because o its uncertain nature. Setting this probability higher than critical value, lead to high answer accuracy. The drawback is increasing the numbers o iterations. I this value is assumed smaller than critical value, answer accuracy will be poor and number o iterations will be low. There is a narrow band or this parameter that guarantee answer accuracy with low iteration. Because o certain nature o other parameters in compare with mutation, they are not as important as mutation. THE PROPOSED METHOD In this method, mutation operator has been changed or improving GA parameters. Mutation only occurs in positions where bit value o all samples at that position is the same. For example at the ollowing samples the bit value in deined mutation point in all populations is the same while other positions have got two values 0 and 1. For this reason mutation only occurs in this position and other bits do not change by mutation. It is obvious that the mutation in proposed method occurs only in deined bits (in our example our bits) while general methods apply mutation in all bits (in our example 3 bits). This modiied mutation point selection lead to better system perormance. Assume one point crossover occur in group 1 and group at deined positions. Ater mutation or samples and 4 the bit value in mutation point is negated (Fig. 1). At the next step i child with mutation (number and 4) remain in selection process, the mutation is said to be good and algorithm continues without any change. On the other hand i these children do not remain in selection process, mutation is not appropriate. This indicates increasing the probability o appearing zeroed bit at this location in the inal answer. This means that probability that deined bit (at mutation point) be zero at inal answer is big. Facing this conditions lead us to inding a way that decrease the probability o mutation in deined bit. In the above example the irst assumed P mute is 0.5 because the mutation occurred in hal o child. As a solution we can assume at next step this probability is 0.5 as a result deined bit in one o our child is negated. In this method we deine dierent probability o mutation or each position (In Fig. 1 eight separate P mute ). This probability P mute is similar or all bits comprising the mutation point. In order to evaluate the probability changes in a better way we deine P start instead o P mute.p start is the probability o changing bits in mutation point at irst mutation which is 0.5 in the above example. P start is

3 World Appl. Sci. J., 5 (): , 008 Mutation point Crossover point group group Crossov Mutation P start = 0.5 First Step Crossover point Fig. 1: Parents and child in an example o proposed algorithm Pmute 1 = Pstart Derate = 0.5 Second Step [ Pmute = Pmute1 Derate = 0.15] Third Step Fig. : Probability changes in proposed method and De rate eect o De rate assumed to be equal in all bit positions. Another major parameter in the proposed method is decrease rate which is deined De rate (Decrease Rate). De rate describes the rate o decreasing P start between two successive mutation steps. In above example P start is assumed 0.5 and 0.5 or irst and second steps, respectively. De rate is deined by division o these two probabilities which is 0.5. Following this procedure, the third P mute computed or above example is 0.15 (Fig. ). It is obvious that De rate value would be bigger than 1 and should be positive. Setting this parameter close to 1 may result in low convergence rate (very high number o iterations). Setting this parameter to a big value leads to a general mutation system with low P mute. The experimental results given in section 5 demonstrate the optimum selected values or these parameters. MATERIAL AND METHODS As a comparison between the proposed method and general method the maximum value o six sample unctions is computed using both general and the 139 proposed methods. Two important parameters, convergence rate and answer accuracy, are considered as actors to compare these methods. Convergence rate obtained by consideration the number o iterations and initial population. For example or a system with the number o initial populations o 3 which needs 4 iterations or inding optimum point, the unction has to be computed 3+3*16=400 times. In irst iteration all samples (N sample) has to be computed, while at other iterations only new populations (i.e. hal o samples) are computed. I the iterations number is noted by Iter, then the number o computations is obtained through Eq 1. As can be seen rom Eq. 1, when the number o iterations rises, the number o computations increases. This means that the convergence rate has been descent. N N No. o Computations...N+ Iter = (+Iter) (1) The second important parameter which is answer accuracy is shown as a rational actor (%) or all unctions (Eq ). In Eq the Precise answer is computed through an exact (ideal) method and the computed answer is the output o dierent GA methods. Preciseanswer-Computedanswer AnswerAccuracy = 100 () Preciseanswer Hal o these six sample unctions have our variables which is noted by irst group (Eq 3 to 5) and others have seven variables noted by second group (Eq 6 to 8). Because o dierence between the number o variables in these sample unctions, the number o initial populations is dierent. Hence, the number o initial populations or irst and second groups is assumed to be 3 and 18, respectively. Another parameter which has to be set is the number o cross points. The number o cross points or

4 both general and the proposed methods is assumed to be 7. The next parameter is number o bits in each variable. This parameter or each variable is selected as 0. Hence, we have 80 bits in irst group and 140 bits in second group or each sample. The number o cross points and bits aects the answer accuracy and number o iterations. But based on our simulation results the eect o choosing these parameters or two (general and the proposed) algorithm are the same, so the comparative results are independent o these parameters. Other requirements are selected as ollows; Initial samples are selected randomly (deined as type a in section -1-1), the selection method choose a certain sample proportional to its itness (deined as type b in section -1-), cross over scheme is type b that was described in section -1, inally mutation scheme is chosen o type b as shown in section -3. The last unknown parameter which we aim to analyze and evaluate its eect on total perormance is probability o mutation. This probability P mute in general method is equivalent to P start and De rate in proposed method. Based on our simulation results De rate variations is more eective than P start variations in system perormance. Hence, we assume a constant value P start = and then analyze De rate eect on system. In the proposed method, other parameters such as number o cross points, initial population, etc are similar to general method. Also because o random nature o GA methods, simulation result o each solving process is not the same. In our experiments all unctions is solved 100 times. In order to soten this random nature eect and get an appropriate answer close to method answer, we discard 10% o maximum and minimum obtained results and then compute average o the remained results. The advantage is that the accuracy is increased and errors have been reduced. The our variable unctions (irst group) described beore are: =(1-x+y^-y)*cosh(x/k)*... 1 (1/(cos(z-)+))*...sinh(z/k)^*(x-k/4) cos(k-z) (k+ π/4) =(1-x y ) cos(xy)...(1/cosh(z-)) 3=abs(cos(k+sinh(xy-y +3))......tanh(cos(yz -3z+x)) log(k -3k+4) (y x -3y xk cos(kz ))) For all above unctions x, y [- ], z [-3 3], k [ 5] which are real variables. The maximum value World Appl. Sci. J., 5 (): , 008 (3) (4) (5) 140 (precise answer) or unction Eq 3 is 6.5, or unction Eq 4 is 5. and or unction Eq 5 is 139. The seven variable unctions (second group) are: 4(x, 1,x)=-(x+)sin(30x 7 1 1)x 1(x+1)cos(xx 1 1 -x^) (x ^3-xcos(10x 3 )-cos(x 3x4^3)) cos(xx 3 5^-x1x) 6 (6) ^(x -x ^ abs(x )-cosh(x x )) ( x 1,, x 7) = (x1+ )sin(0x1+ 1)cos(10x + ) (x 1) sin(3x3 1).cos(7x 3 + 3) x4 5 (x3 3) sin(10x4 1) cos(0x5 ) (x 4)(x 3)(x + 1) x cos(x ) (7) (x, 1,x)= 7 -(x+)sin(30xx 1 1 )x (xx )) sin(πx 5x 7)(x 5-)(x 5x 4-4)(x 5-3) cos(x x ^-x x ) For all unctions the variables spaces is equal as ollows: (8) [ ] [ ] [ ] [ ] [ ] [ ] [ ] x1 1,x 1,x3 3 3, x 1 5,x 4,x 1 3,x The maximum value or unction Eq. 6 is 0.111, or unction Eq. 7 is and or unction Eq. 8 is 48. RESULTS First group, Number o Initial Population: 3, No. O Cross Point: 7, Type o Selection: Ideal, (proposed method P start = 0.5). The simulation results or irst group unctions set is shown in Table 1-3. We deine a critical P mute point in general method, which is described as ollows. A critical P mute point is an optimum point in which answer accuracy and number o computations are acceptable. Choosing P mute s lower or higher than this value cause to degradation one o answer accuracy or number o computations parameters. In Eq 3, critical value or P mute is 0.006, i P mute less than this value answer accuracy is poor, while or P mute higher than 0.006, answer accuracy is acceptable. The disadvantage is that convergence rate decreases (number o computations increase). In Eq. 4 and Eq. 5 the critical P mute values are 0.00 and respectively. For higher values o P mute number o computations increase without an eicient eect on accuracy. We have not any prior knowledge about P mute so we have to assume an initial value or it. This lack o deining P mute may decrease the system perormance.

5 World Appl. Sci. J., 5 (): , 008 Table 1: Function 1 (Eq. 3) P mute Answer accuracy as % Number o computations Answer accuracy as % Number o computations Table : Function (Eq. 4) P mute Answer accuracy as % Number o computations Answer accuracy as % Number o computations Table 3: Function 3 (Eq. 5) P mute Answer accuracy as % Number o computations Answer accuracy as % Number o computations In proposed system, De rate controls system parameters. Choosing low values or De rate, increase accuracy but decrease convergence rate (increase number o computations) simultaneously. However this parameter has two major advantages in compare with P mute. First, reducing De rate increases accuracy as much as possible, while increasing accuracy stop or even degrade with increasing P mute in general method. As can be seen rom Table (general method) increasing the P mute rom 0.00 with accuracy o % and number o computations to 0.008, leads to accuracy o 95. % with number o computations Also our experiments show that in general method increasing the system perormance may require changing the number initial populations or other parameters, while in the proposed method De rate itsel can improve system perormance. Second comparative parameter, convergence rate, in the proposed system is approximately lower than or equal to general system in irst group. 141 For veriying result correctness and present a more reliable result, simulations are repeated or second group. Second group, Seven Variables Functions, Number o Initial Population 18, No O Cross Point 7, Type o Selection: Ideal, (proposed method P start = 0.5). The simulation results or second group have been given in Table 4-6. As can be seen rom Table 4 regarding Eq 6, the result o general method is better than proposed method. This implies that the proposed method may achieve poorer results rather than general methods, in some systems. In Eq 7 and 8 the critical value or P mute approximately is It is obvious rom Table 5 and 6 that increasing P mute rom critical value does not change signiicantly accuracy parameter, similar to irst group. The results o second group (Eq 7, 8) analysis conirm the advantages stated or irst group. However, the number o computations in these unctions is higher than that or general method which is not an appropriate property.

6 World Appl. Sci. J., 5 (): , 008 Table 4: Function 4 (Eq. 6) P mute Answer accuracy as % Number o computations Answer accuracy as % Number o computations Table 5: Function 5 (Eq. 7) P mute Answer accuracy as % Number o computations Answer accuracy as % Number o computations Table 6: Function 6 (Eq. 8) P mute Answer accuracy as % Number o computations Answer accuracy as % Number o computations CONCLUSION We proposed a new method or solving P mute selection problem in conventional GA systems. In this method the eect o P mute has been removed and two new parameters is added to system. The accuracy o new method is better than conventional GA methods and is not limited similar to conventional method. But in some situations convergence rate o proposed method is lower than general method; especially in systems with high complexity. The new method has been tested or six dierent systems. The results were compared with conventional systems in similar conditions and conirmed the reliability o this novel algorithm. As a uture work, we recommend the ollowing items. The eect o dierent values or P start has to be evaluated. The system robustness via unction shape has to be studied (which unction type, works better in proposed method and vise versa).. Eshelman, L. and J. Schaer, Real-coded genetic algorithms and interval-schemata. In Foundations o Genetic Algorithms, L.D. Whitley, Ed. Morgan Kaumann, pp: Tate, D.M. and A.E. Smith, Expected Allele Coverage and the Role o Mutation in Genetic Algorithms. In Forrest, S. (Ed.). Proceedings o the Fith International Conerence on Genetic Algorithms, Morgan Kaumann, pp: Mi Hinterding, R., H. Gielewski and T.C. Peachey, The Nature o Mutation in Genetic Algorithms. In Eshelman, L.J. and Morgan Kaumann (Eds.), Proceedings o the Sixth International Conerence on Genetic Algorithms. pp: Spears, W.M., Crossover or Mutation? In Foundations o Genetic Algorithms. Whitley, L.D. and Morgan Kaumann (Eds.). REFERENCES 1. Ulrich Bodenhoer, Genetic Algorithms: Theory and Applications, Fuzzy logic laboratorium Linz- Hagenberg, 3rd Edn., 003/