Indonesian J our nal of Electrical Engineering and Computer Science V ol. 24, No. 2, No v ember 2021, pp. 1161 1172 ISSN: 2502-4752, DOI: 10.11591/ijeecs.v24.i2.pp1161-1172 r 1161 Artificial neural netw ork based meta-heuristic f or perf ormance impr o v ement in ph ysical inter net supply chain netw ork Chouar Abdelsamad 1 , T etouani Samir 2 , Soulhi Aziz 3 , Elalami J amila 4 1,2 Laboratoire d’Analyse des Syst ` emes, T raitement de l’Information et Management Int ´ egr ´ e (LASTIMI), Uni v ersit ´ e Mohammed V -Agdal Ecole Mohammadia d’Ing ´ enieurs, Rabat, Morocco 1,2 Centre d’Excellence en Logistique (CELOG), ´ Ecole Sup ´ erieure de l’industrie du T e xtile et d’Habillement (ESITH), Casablanca, Morocco 3 Superior National School of Mines, Rabat, Morocco 4 National Center for Scientific and T echnical Research (CNRST), Rabat, Morocco Article Inf o Article history: Recei v ed Jun 3, 2021 Re vised Sep 11, 2021 Accepted Sep 15, 2021 K eyw ords: Artificial neural netw orks Slime mould algorithm Supply chain management Ph ysical internet ABSTRA CT No w adays, reducing total costs while enhancing customer satisf action is a major task for man y supply chain systems. T o deal with this issue, the ph ysical internet (PI) paradigm can be represented as a potential replacement for the current logistics sys- tem. This paper de v oted the cost reduction and lead time impro v ement in a PI-SCN using a h ybrid frame w ork based on an artificial neural netw ork (ANN) and an im- pro v ed slime mould algorithm (ISMA). T o address the performance of the proposed frame w ork, a real-case study in Morocco is considered. The ne w trainer ISMA s per - formance has been in v estig ated in three approximation datasets from the Uni v ersity of California at Irvine (UCI) machine-learning repository re g arding nine recent meta- heuristics. The e xperimental results highlight the ef fecti v eness of ISMA according to other meta heuristics for training feed-forw ard neural netw orks (FNNs) to con v er ge speed and to a v oid local minima. This is an open access article under the CC BY -SA license . Corresponding A uthor: Chouar Abdelsamad Centre d’Excellence en Logistique (CELOG) ´ Ecole Sup ´ erieure de l’industrie du T e xtile et d’Habillement (ESITH) Casablanca, Morocco Email: chouar@esith.ac.ma 1. INTR ODUCTION No w adays, the major strength for the global logistics operations is to ensure a sustainable syst ems through inte grating the de v eloped technologies and methodologies to the real w orld practices. F or a lar ge logistics scale, t he logistics web aims to connect the supply chain’ s netw ork including the dif ferent actors, ph ysical items and digital technologies in order to assist the global requirements. From a broadly perspecti v e, the ph ysical internet (PI) aims to optimize the supply cha in processes according to a defined frame w ork to enhance the logistics web ef ficienc y , ef fecti v eness and sustainability which de v elop the required reliability , resilience and adaptability . The purpose of the inno v ati v e Ph ysical Internet (PI or ) initiati v e is to re v erse the situation of e xisting unsustainable in current logistics systems. Indeed, due to the dynamic nature of real-w orld problems, logistics web design models must tak e into consideration the risks of disruption and unforeseen e v ents to ensure the resilience and ef ficienc y of the entire logistics web chain. F or e xample, taking into account J ournal homepage: http://ijeecs.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
1162 r ISSN: 2502-4752 uncertainties (demand and road traf fic) and assessing the risks of disruptions caused by major crises, such as the crisis of the corona virus disease (CO VID-19) epidemic. The capacity to measure the strate gical, tactical and operational performance is considered as a main frame w ork to assert ine vitably in order to strengthen the enterprises competiti v eness within ph ysical internet supply chain netw orks. This allo ws the long-term outputs ef fects ass essment to further support the competi- ti v eness and decision-making po wer [1]. Therefore, a well-defined set of performance indicators is mandatory to consolidate the required objecti v es for the o v erall performance measurement. The underpinning to unco v er these indicators i n order to increase the chances of success is an o v erall analysis of the compan y’ s en vironment processes [2]. Thus, correction’ s adv antages are performed through e v aluation perspecti v es which must g ather financial and non-financial measures. Since the presence of man y criteria, the performance measurement in ph ysical internet supply chain netw ork (PISCN) is defined as a problem which belongs to multiple criteria de- cision making. In f act, multiple methodologies ha v e been de v eloped to e v aluate a multiple criteria scheme such as data en v elopment analysis (DEA) [3]. Accordingly , the DEA contrib utes to assess the ef ficienc y surf aces through a mathematical programming model. Wherea s, the resolution process can be af fected statistical noises which e xtent to wrap the deri v ed frontier [4]. Due to the panoply of choice for decision-making processes, the non-parametric tool for non-linear relations between inputs and outputs modeling approach has been widely de v eloped with the artificial neural netw ork (ANN) [5]. In f act, the ANN presents a wide v ariety in the literature such as spiking neural net- w orks [6] and recurrent neural netw orks [7], though, the most popular type is the feed-forw ard neural netw orks (FNNs) [8]. Besides, the learning process (i.e., training process) has a huge impact on the process performance of ANNs. As a whole, training algorithms can be arranged into tw o cate gories: gradient-based algorithms v ersus stochastic search algorithms. It can be noted from [9] that the widely adopted gradient-based training algorithm is back-propag ation (BP). T o some e xtent, the cons of this method are e vinced, for instance, through the tardy con v er gence beha vior and hanging on local minima. Besides, for optimization problems, considering some nature-inspired metaheuristics algorithms as alternati v e trainers ha v e pro v ed a higher ef ficienc y to di v er ge from local minima. The lar ge potential of meta- heuristic methods to train the feed-forw ard neural netw orks (FNNs) has been widely asserted in the literature. In t his respect, the krill herd algorithm (KHA) has been established for data classification to train t he FNNs [10]. Not l ong ago in 2016, a nature-inspi red algorithm kno wn as multi v erse optimizer (MV O) has been used for training the FNNs [11]. It is w orth mentioning that through the reported numerical results, the MV O re v eals a high competiti v eness and it performs better than another training algorithms in most of datasets. In spite of the high quality of the pre vious presented w orks, the local optima entrapment’ s issue con- tinues to be f aced. Besides, as mentioned by [12], a theorem kno wn as No Free Lunch within the heuristics area highlights t he lack of a generic problem solving optimization algorithm. Gi v en that, the performance g ap between algorithms occurs after FNNs training for multiple data sets. Thus, ne w algorithms ef ficiencies for learning FNNs are considered as a w orth y field to be addressed by researchers. In this respect, this paper aims to embed the ne wly slime mould algorithm (SMA) algorithm [13] into FNNs. This paper appraises the ph ysical internet supply chain netw ork performance (PI-SCN). The proposed structre is tw o steps based. Firstly , we depict three fe atures’ cate gories (i.e., economic, social, and en viron- mental) in addition to the tar get v ariables (reducing costs and lead time impro v ement) af fecting the ph ysical internet supply chain netw ork (PI-SCN). Secondly and in order to train the FNNs, a ne w method formulation has been applied using the impro v ed slime mould algorithm (ISMA) to reach the ef ficienc y v alues. The remainder of this article is or g anized as follo ws. The performance measurement system is dis- cussed in the sec tion 2. Section 3 details the methodologies applied in this study . Sect ion 4 presents numerical computations and discussion. At the end, section 5 summarizes the conclusions and points out future re- searches. 2. PHYSICAL INTERNET SUPPL Y CHAIN NETW ORK (PI-SCN) PERFORMANCE SYSTEM The assessment of operations’ performance in a ph ysical internet supply chain netw ork stands among the main manageri al b usiness af f air . But, it is v ery tough to e v aluate an or g anization’ s performance when se v eral measures belongs to a defined system or operation [14]. Besides, the e xpanding competiti v eness in the supply c hain domain requires more adv anced performance le v el. According to the global objecti v es of the compan y , the related system of performance measurement’ s indicator will be dra wn [15]. The importance Indonesian J Elec Eng & Comp Sci, V ol. 24, No. 2, No v ember 2021 : 1161 1172 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 r 1163 of financial measures to assess the or g anizations’ profits point out their e xistence within man y performance measurement frame w orks proposed for ph ysical supply chain netw ork. This study highlights the objecti v es to rely on as well as the in v olv ed performance indicators to reach the tar geted PI-SCN performances. The whole steps are presented in Figure 1. Figure 1. Ph ysical internet supply chain netw ork system 3. RESEARCH METHOD In this part, we illustrate the recommended frame w ork to e v aluate the performance of (PI-SCN). 3.1. F eed-f orward neural netw ork One of the most popular type of ANNs is the FNNs. In this netw ork, the information has a unique progression’ s direction which start from the inputs to outputs by mo ving across set “neurons” [16] in hidden layer . Ne v ertheless, the netw ork does not intend an y c ycles or loops. The Figure 2 illustrates an elementary FNN with a single hidden layer . As highlighted, the sum of the inputs’ weight are computed by each neuron considering a bias. Subsequently , the sum is pass ed across sigmoid function and then reach the output of NN. The procedure is represented by (1)-(3): H j = R X i =1 ! i;j I j + b j (1) Where R is the number of nodes of input layers, ! i;j denotes the connection weight between the i th neuron of the input layer and j th neuron of the hidden layer , b j is the threshold (bias) in hidden layers and I i is the i th input data. f ( x ) = 1 1 + e x (2) Here f ( x ) is the sigmoid function. The output of the netw ork is calculated as follo ws: y k = f k ( N X j =1 i;j H j + b k ) (3) Artificial neur al network based meta-heuristic for performance impr o vement... (Chouar Abdelsamad) Evaluation Warning : The document was created with Spire.PDF for Python.
1164 r ISSN: 2502-4752 Where i;j denotes the connection weight between the j th neuron of the hidden layer and k th neuron of the output layer , b k is the threshold (bias) in output layers. As long as some error criterion is not reached, the training procedure is performed to re gulate t he weights and bias. As a matter of f act, pick up the proper training algorithm is the principal challenge. Moreo v er , the design comple xity of the neural netw ork increases gi v en that man y elements af fect the training performance, for instance, the total nodes in hidden layers, in addition to the error and acti v ation functions. Figure 2 sho ws a simple FNN structure. Figure 2. FNN architecture 3.2. Brief description of SMA A ne wly optimization technique has been presented not long ago by [13] called the SMA, the general concept is simulated from slime moulds, ph ysarum polyce p ha lum intelligent beha viour . Henceforw ard, this algorithm which is based on a population stochastic search procedure has been emplo yed into man y comple x engineering problems in the field of optimizati on. The principal concepts of SMA are outlined in the follo wing subsection. 3.2.1. A ppr oach f ood T o model the approaching beha vior of slime mould as a mathematical equation, the follo wing rule is proposed to imitate the contraction mode by using (4): X t +1 = X b ( t ) + v b : ( W :X A ( t ) X B ( t )) r < p v c :X t r p (4) where v b is a parameter with a range of [ a; a ] , v c decreases linearly from one to zero. The t represents the current iteration, X b represents the indi vidual location with the highest odor concentration currently found, X represents the location of slime mould, X A and X B represent tw o indi viduals randomly selected from the sw arm, W represents the weight of slime mould. The formula of p is as follo ws using (5): p = tanh j S ( i ) D F j (5) where i 2 1 ; 2 ; : : : ; n , S ( i ) represents the fitness of X , D F represents the best fitness obtained in all iterations. The formula of v b is as follo ws by (6): v b = [ a; a ] (6) The formula of a is as follo ws by (7): a = arctanh t max t + 1 (7) The formula of W is listed as follo ws by (8): W ( S mel l I ndex ( i )) = 1 + r log (( b F S ( i )) = ( b F w F ) + 1) condition 1 r log (( b F S ( i )) = ( b F w F ) + 1) other s (8) Indonesian J Elec Eng & Comp Sci, V ol. 24, No. 2, No v ember 2021 : 1161 1172 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 r 1165 S mel l I ndex = sor t ( S ) (9) where condition indicates that S ( i ) ranks first half of the population, r denotes the random v alue in the interv al of [0,1], b F and w F illustrates, respecti v ely , the obtained optimal and w orst fitnesses in the current iterati v e process, while S mel l I nde x denotes the sorted sequence of fitness v alues (ascends in the minimum v alue problem). 3.2.2. Wrap f ood The mathematical formula for updating the location of slime mould is described by (10): X = 8 < : r and ( U B LB ) + LB r and < z X b ( t ) + v b ( W X A ( t ) X B ( t )) r < p v c X ( t ) r p (10) where LB and U B outline the lo wer and upper boundaries of the search range, rand and r define the random v alue within [0 : 1] . 3.2.3. Oscillation The v alue of v b oscillates randomly between [ a; a ] and it gradually con v er ges to w ard zero as well as the iterations increase. The v alue of v c oscillates between [ 1 ; 1] and tends to zero e v entually . The main steps of the SMA are illustrated in the figure and the algorithm. 3.3. L ´ evy flights Le vy flights is a non-Gaussian stochasti c process, the related step sizes are distrib uted based on a Le vy stable distrib ution to generate ne w solutions. Once a no v el solution is defined, the follo wing Le vy flight is carried out in (11): X t +1 = X t Lev y ( ) (11) where indicates t he step size related to the problem’ s scales. The product means entry-wise multiplications. The pre v ailing idea is that Le vy flights furnish a random w alk gi v en that for lar ge steps. In this study , the algorithm proposed by [16] will be used on account of its prominent ef ficient hi ghlighted with Le vy flights implementation. 3.4. Impr o v ed slime mould algorithm The pre v ailing criteria for an ef ficient optimization algorithm are based on the strong e xploration ability in addition to a f ast e xploitation rate. W ith the aim for SMA performance’ s impro v ement and to e xpand the algorithm e xploration, an update position-based accelerated particle sw arm optimization (APSO) [17] and L ´ evy flight technique are included into the SMA. The principal intention of the suggested algorithm is in that w ay . The basic idea of the proposed algorithm is as follo ws. First, a fraction of the population is chosen according to the w orst fitness v alue. Then a L ´ evy flight is performed according section 2.2, while the standard SMA is applied to the rest of better solutions. Secondly , an update position-based is implemented based on APSO. Generally , APSO has the ability to prospect rapidly the search space and find out ef ficiently the optimal solution. Hence, the position is updated by the follo wing (12): X t +1 = (1 ) X t + g + r (12) The v elocity is not included in the equation 12, thereof, the APSO does not require v elocities’ initiali zation, in this re g ard, it a v oids the dra wbacks related to re gular PSO v elocities. At this stage the third term r compels the system to be more mobile and to a v oid entrapment within an y local optima if its selection has been performed properly , the corresponding definition of r can be dra wn from a statistical distrib ution. According to the other parameters such as and are choosing according to [17] as follo w in (13): = t (13) where = 0 : 2 0 : 7 and = 0 : 1 0 : 99 . Here t 2 [0 ; t max ] . The proposed method is outlined in Algorithm 1: Artificial neur al network based meta-heuristic for performance impr o vement... (Chouar Abdelsamad) Evaluation Warning : The document was created with Spire.PDF for Python.
1166 r ISSN: 2502-4752 Algorithm 1 Proposed ISMA trainer Inputs : N : Population size and max t : maximum number of iterations Outputs : The best solution Slime mould positions X i ( i = 1 ; 2 ; : : : ; n ) at t = 0 while (stop criterion) do Calculate the fitness of all slime mould f or (each portion N pop F r action of w orst of solutions) do Perform Le vy flight for X i to generate a ne w slime X 0 i using Eq. (11 ) X i   X 0 i f i   f 0 i end f or Calculate the W by (8 ) Update bestFitness and X b f or (each portion 3 N pop F r action of rest of solutions) do Update p , v b , v c Update positions by : 8 < : (1 ) X t + X b + r r and < z X b ( t ) + v b ( W X A ( t ) X B ( t )) r < p v c X ( t ) r p end f or end while Retur n bestFitness and X b 4. ISMA FOR TRAIN FNNS 4.1. Ar chitectur e of FNNs During NNs emplo yment, the structure should be essentially defined according to the layers’ number in addition to layers’ neurons number . The NN comple xity’ s is correlated to the number of neurons in the hidden layers. As long as the number is important, the comple xity increases. In this study , the characteristic of a problem-dependent has been associated to the input and output neurons’ number in MLP netw ork so that the K olmogoro v theorem [18] has been adopted to compute the number of hidden nodes through the (14): H = 2 I + 1 (14) The netw ork’ s weights and bias ha v e been optimized through SMA. Moreo v er , D reflects each or g anism’ s dimension: D = ( I H ) + ( H O ) + H bias + O bias (15) Notice that I , H and O describe respecti v ely the input, hidden and output FNN’ s neurons. While H bias and O bias presents the biases’ number and output layers. Figure 3 sho ws the FNN architecture for PI-SCN. Figure 3. PI-SCN system ANN based representation Indonesian J Elec Eng & Comp Sci, V ol. 24, No. 2, No v ember 2021 : 1161 1172 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 r 1167 4.2. Method e v aluation The e v aluation according to ISMA of each slime mould is performed gi ving its fitness which reflect s to its status. The corresponding process is the follo wing; the v ector that include the weights and biases is passed to FNNs, afterw ard, the neural netw ork prediction emplo ying the training dataset is used to figure out the mean squared error (MSE) criterion. The optimal solution is reached after consecuti v e iterations which describe the neural netw ork weights and biases. The (16) presents the MSE criterion, M describes the samples’ number in training dataset and ( b Y ; Y ) are respecti v ele y to the estimated and the original v alues according to the suggested model. M S E = 1 M M X i =1 ( y r b y r ) (16) 4.3. Encoding strategy Dif ferent encoding strate gies ha v e been introduced by [19]. F or instance, in the field of e v olut ionary algorithm, the FNNs’ weights and biases for e v ery agent can be structured in multiple forms such as v ector , matrix, or binary . The Figure 4 highlights an encoding strate gy which belongs to v ector structure, this method has been adopted for the present study . In this respect, the FNNs’ weights and biases stand for each mould which con v erted afterw ards into a real number single v ector . Figure 4. Solution representation 4.4. Pr oposed model In this part, the suggested model is e xplored according to a t hree part. Firstly , three cate gories of fea- tures are ide n t ified (i.e., economic, s ocial, and en vironmental) and the tar get v ariables (reducing costs and lead time impro v ement) that af fect our system. At the end, the ef ficienc y scores are determined while implementing ISMA as no v el method to train FNNs. The related al gorithm for the proposed h ybrid frame w ork is reported in the Figure 5. Figure 5. Proposed h ybrid frame w ork Artificial neur al network based meta-heuristic for performance impr o vement... (Chouar Abdelsamad) Evaluation Warning : The document was created with Spire.PDF for Python.
1168 r ISSN: 2502-4752 5. RESUL T AND DISCUSSION In this part, the suggested h ybrid frame w ork ef ficienc y i s e xplored to e v aluate the PI-SCN perfor - mance. The proposed method is compared ag ainst recent nine algorithms such as GA [20], GW O [21], SCA [22], W O A [23], HHO [24], SMA [13], MV O [25], MFO [26]. The whole algorithms were programmed in MA TLAB R2014a. The e xperiments computations ha v e been performed through 20 distinct runs; the other algorithms’ parameters are tak en the same as the published paper’ s v alues. The dataset sho wn in T able 1 has been g athered e xploiting a brainstorming. Because of condientiality concerns, the first data has been changed and partitioned into 66% for training and 34% for testing. T able 1. T e xtile datasets Features T ar gets Economic Social En vironment SG PG R OC G L TTF R OS R OCS Spo E WWS SSD Spi Costs L T 10 2 6 6 8 3 6 1 6 2 4 5 4 69 68 9 6 6 9 1 2 6 3 9 2 5 7 6 76 61 10 9 6 3 5 2 3 9 1 8 3 5 3 70 53 6 2 5 10 1 5 2 9 5 1 10 1 6 79 73 3 2 9 7 9 6 9 4 6 8 6 1 10 66 53 1 10 10 8 3 7 3 6 3 1 6 6 9 74 73 7 8 6 5 1 3 8 3 9 8 1 8 5 71 55 3 10 8 4 9 1 2 1 6 1 3 5 4 51 51 In order to compare all algorithms and during all benchmarks, the a v erage (A VE) and the standar g de viation (STD) ha v e been emplo yed. These tw o mesures ha v e been implemented to highlight the algorithm’ s ef fecti v eness to escape from local minima entrapment. A case study has been applied based on 100 companies operating in te xtile industry in order to v alidate our model. The results of dataset are reported in T able 2. By analysi n g the results of the T able 2, the pre v ailing element to share is the best performance outlined by the proposed method as well as MFO and HHO. This beha vior is depicted by the highest abi lity to escape local optima which is fundamentally finer comparing to another algorithms. Besides, a con v er gence comparati v e e xperimentation w as implem ented to appro v e that ISMA has greater con v er gence performance than the other algorithms. Figure 6 sho ws the con v er gence curv es. T able 2. Experimental results for te xtile compan y dataset GA GW O SCA W O A HHO SMA MV O MFO ISMA Min 9,86E-04 9,80E-04 1,02E-03 1,13E-03 5,02E-04 9,71E-04 9,64E-04 5,00E-04 5,00E-04 Max 9,86E-04 9,80E-04 1,02E-03 1,13E-03 5,02E-04 9,71E-04 9,64E-04 5,00E-04 5,00E-04 A V G 9,86E-04 9,80E-04 1,02E-03 1,13E-03 5,02E-04 9,71E-04 9,64E-04 5,00E-04 5,00E-04 Std 1,6E-02 4,0E-03 5,0E-02 4,7E-02 8,9E-04 3,9E-02 2,7E-02 8,5E-04 8,1E-04 Error 8,45 8,32 13,19 11,56 5,45 9,36 8,10 5,40 5,25 p-v alue 8,60E-08 6,09E-03 6,80E-08 1,06E-07 5,31E-01 7,41E-09 4,60E-04 6,61E-0 N/A Figure 6. Con v er gence curv es-case study Indonesian J Elec Eng & Comp Sci, V ol. 24, No. 2, No v ember 2021 : 1161 1172 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 r 1169 Additionally , the proposed method has led to higher ef ficac y threshold. This result has been demon- strated through a benchmark with three selected standard classification data sets from the Uni v ersity of Cal- ifornia at Irvine (UCI) machine learning repository [27]: sigmoid, cosine and sine. The dataset i s sho wn in T able 3. T able 3. Function approximation dataset Function approximation dataset T raining samples T est samples Sigmoid: y = 1 = (1 + e x ) 61: x 2 [3 : 0 : 1 : 3] 121: x 2 [3 : 0 : 05 : 3] Cosine: y = ( cos ( x = 2)) 7 31: x 2 [1 : 25 : 0 : 05 : 2 : 75] 38: x 2 [1 : 25 : 0 : 04 : 2 : 75] Sine: y = sin (2 x ) 126: x 2 [ 2 : 0 : 1 : 2 ] 252: x 2 [ 2 : 0 : 05 : 2 ] 5.1. Sigmoid function The sigmoid dataset belongs to the interv al [ 3 ; 3] with increases of 0.1 which sum up the number of training data to 61. The number of test samples is 121, lying in the same range. The test errors in T able 4, con v er gence curv e in Figure 7 highlights that the ISMA algorithm admit the higher approximate e xactitude in addition to the f astest con v er gence rate. T able 4. Sigmoid dataset results GA GW O SCA W O A HHO SMA MV O MFO ISMA Min 8,9E-09 8,6E-09 1,3E-02 1,7E-03 3,9E-04 2,9E-04 3,7E-04 4,0E-04 5,0E-09 Max 5,8E-02 4,5E-02 1,4E-01 1,5E-01 7,8E-03 6,4E-02 6,5E-02 5,6E-03 2,9E-03 A V G 1,6E-02 3,6E-03 6,3E-02 4,6E-02 5,1E-03 1,2E-02 1,0E-02 2,5E-03 8,0E-04 Std 1,6E-02 1,0E-02 3,0E-02 4,7E-02 2,9E-03 1,9E-02 1,7E-02 1,4E-03 7,3E-04 Error 0,78 0,20 9,63 3,50 1,41 1,56 1,77 1,79 0,14 p-v alue 8,60E-08 5,61E-01 6,80E-08 1,06E-07 2,60E-06 7,41E-09 4,60E-04 9,28E-09 N/A Figure 7. Con v er gence curv es-sigmoid dataset 5.2. Cosine function The cosine dataset belongs to the interv al [1 : 25 ; 2 : 75] with increases of 0 : 05 , therefore the amount of training data is 31. The number of test samples is 38, lying in the same range. The test errors in T able 5, con v er gence and curv e in Figure 8 highlights that the ISMA algorithm admit the higher approximate e xactitude in addition to the f astest con v er gence rate. T able 5. Cosine dataset results GA GW O SCA W O A HHO SMA MV O MFO ISMA Min 3,8E-04 2,1E-04 2,9E-03 3,1E-03 2,1E-03 1,7E-04 6,7E-04 1,3E-02 1,2E-04 Max 4,3E-03 2,0E-03 1,4E-02 6,8E-02 1,3E-01 2,5E-03 2,4E-03 1,3E-01 1,2E-03 A V G 1,4E-03 6,8E-04 6,6E-03 2,0E-02 4,4E-02 8,2E-04 1,6E-03 8,0E-02 6,0E-04 Std 8,8E-04 4,6E-04 3,1E-03 2,0E-02 4,9E-02 6,7E-04 4,9E-04 3,9E-02 3,3E-04 Error 0,79 0,94 1,43 1,81 1,42 0,65 0,92 3,56 0,44 p-v alue 1,20E-01 1,58E-08 6,80E-06 6,80E-06 1,23E-07 2,22E-04 5,17E-08 6,80E-06 N/A Artificial neur al network based meta-heuristic for performance impr o vement... (Chouar Abdelsamad) Evaluation Warning : The document was created with Spire.PDF for Python.
1170 r ISSN: 2502-4752 Figure 8. Con v er gence curv es-cosine dataset 5.3. Sine function The sine dataset belongs to the interv al [ 2 ; 2 ] with increases of 0 : 1 , therefore the amount of training data is 126. The number of test samples is 256, lying in the same range. The test errors in T able 6, con v er gence and curv e in Figure 9 highlights that the ISMA algorithm admit the higher approximate e xactitude in addition to the f astest con v er gence rate. T able 6. Sine dataset results GA GW O SCA W O A HHO SMA MV O MFO ISMA Min 8,0E-02 6,1E-02 1,4E-01 1,1E-01 1,1E-01 4,9E-02 1,6E-02 1,1E-01 4,4E-02 Max 2,9E-01 1,2E-01 2,8E-01 2,9E-01 1,2E-01 2,8E-01 2,2E-01 1,2E-01 2,2E-01 A V G 1,5E-01 1,2E-01 2,0E-01 2,1E-01 1,2E-01 1,3E-01 1,2E-01 1,2E-01 1,2E-01 Std 5,9E-02 5,0E-02 3,6E-02 5,4E-02 2,3E-03 5,5E-02 6,3E-02 2,5E-03 2,4E-03 Error 112,50 71,50 175,80 168,80 177,80 147,30 136,54 178,12 28,90 p-v alue 4,8E-02 1,6E-02 6,08E-06 1,20E-08 8,35E-03 5,61E-02 6,17E-03 4,16E-04 N/A Figure 9. Con v er gence curv es-sine dataset Indonesian J Elec Eng & Comp Sci, V ol. 24, No. 2, No v ember 2021 : 1161 1172 Evaluation Warning : The document was created with Spire.PDF for Python.