Optimization of e-learning Model Using Fuzzy Genetic Algorithm

Transcription

1 Optmzaton of e-learnng Model Usng Fuzzy Genetc Algorthm Receved: 00?????? 2012, Accepted: 00????? 2013 M. A. Afshar Kazem Assocate professor, Informaton Technology Management Department, Electrnc Branch, Islamc Azad Unversty, Tehran, Iran Abbas Toloe Eshlagy Full Professor, Industral Management Department, Scence and Research Branch, Islamc Azad Unversty, Tehran, Iran Fateme Nazer Master of Scence, Department of Informaton Technology Management, Electrnc Branch, Islamc Azad Unversty, Tehran, Iran ABSTRACT E-learnng model s examned of three major dmensons. And each dmenson has a range of ndcators that s effectve n optmzaton and modelng, n many optmzaton problems n the modelng, target functon or constrants may change over tme that as a result optmzaton of these problems can also be changed. If any of these undetermned events be consdered n the optmzaton process, ths s called optmzaton problem. Many problems n the real-world are dynamc and uncertan and solve them as statc are not approprate. In ths paper, for the frst tme a fuzzy genetc algorthm for optmzaton and modelng e-learnng s presented. Method s that we create a fuzzy model at frst, and then we perform optmzaton by genetc. The results of proposed algorthm on moble peaks benchmark that are already best-known benchmark for evaluatng n the modelng are evaluated and the results of several vald algorthms have been compared. The results ndcate the hgh effcency of the proposed algorthm n comparson wth other algorthms. Keywords: E-Learnng Model, Fuzzy Logc, Optmzaton, Genetc Algorthm, Moble Peaks Benchmark 1- Introducton Fuzzy Genetc Algorthm (FGA*) s one of the algorthms are derved from nature and collectve ntellgence that s provded by doctor L Xao Le n 2002 [1]. The algorthm s a collectve behavor-based technque that s nspred by socal behavor n nature. Ths algorthm has a hgh convergence speed characterstcs, nsenstve to ntal values, hgh flexblty and fault tolerance. Ths algorthm s used n the applcatons of optmzaton such as learnng of forward neural network, [2], data clusterng [3], [4], data mnng [5], optmzaton of nonlnear functons [6], [7], Combnatonal optmzaton [8] [9]... Uncertanty n many real-world problems s qute clear and evdent. One way of dealng wth uncertanty s usng collectve ntellgence and evolutonary methods. Uncertan problems that have already been studed by evolutonary methods n general can be dvded nto four categores; Nose n the evaluaton functon and beng dynamc optmal solutons, perturbatons n the desgn varables, beng approxmate the evaluaton functon and beng dynamc optmal solutons. In ths paper, the uncertanty of the type of beng dynamc optmal solutons that s among the most common types of uncertanty s taken nto consderaton. Several methods to solve modelng problems usng evolutonary computng [10] and collectve ntellgence [11] methods are presented. Two major problems of evolutonary computng methods that make lack of the ablty drect use of these technques to optmze n the modelng envronment are nvald memory and dversty loss. When the model s changed, the resulted solutons n memory are not longer vald or they have completely forgotten or they can be assessed agan. Also, snce most of the evolutonary computng and collectve ntellgence methods due to ther nature are convergng to a pont, so cluster varety n the envronment s lost and f changes n the envronment converge to new optmal pont, f possble, would be very tme consumng. Thus the basc requrements of optmzaton n the modelng envronment can be used and correctly synchronze memory and also create a varety n the clusters after changes the model are consdered. Ths paper presents an FGA to optmze the e-learnng model n whch all requrements of modelng have been satsfed. In the proposed algorthm, fuzzy logc mechansm defnton of law, establsh a database [1] [7] [9] has changed. Also some parameters of FGA ncludng crowdng factor and step length were removed. The proposed algorthm used on the Moble Peak Benchmarks (MPB) *[12] that s the best known benchmarks n modelng and ts performance wth sx other algorthms called mqso [13], mcpso [13], AmQSO [14 ], CellularPSO [15], SPSO [16], and rspso [17] are compared. Comparson benchmark s an Offlne Error that s one of the man crtera Internatonal Journal of Informaton, Securty and System Management, 2014, Vol.3, No.1, pp

2 to compare the algorthms s desgned for dynamc envronments [11]. Test results show that the proposed algorthm has acceptable performance. Remanng of ths paper s organzed n the followng way. The second secton descrbes the proposed algorthm. Test results are dscussed n the thrd secton and the fnal secton presents the conclusons. 2 - The Proposed Model In ths secton, a new model based on fuzzy genetc algorthm for optmzaton n dynamc envronments s presented. In ths algorthm, behavors and parameters and the process of performng FGA has been changed tll the proposed algorthm can fnd the optmal model and followng them after the changng envronment. In the proposed algorthm three steps of fuzzcaton, modelng and optmzaton for e-learnng model are done that has major dfferences wth prevous standard model [1] [7] [9]. Then, after descrbng a New Fuzzy Genetc Algorthm (NFGA), a new optmzed model s ntroduced. Fgure 1 Model of e-learnng Fuzzy Rules After the model s desgned the rule base to determne assocatons between varables should be formed n form of fuzzy that the model of rule base s as follows. 2-1 E-Learnng Model Optmzaton Algorthm Frst, we wll descrbe the FGA step by step. E-learnng model has three man dmensons of the followng: 1 - Hard preparaton 2 - Software preparaton 3 - Coordnaton montorng support preparaton Each of these dmensons, have smaller dmensons. Software preparaton, ncludng nne other smaller dmensons named 1-securty preparaton, 2-polcy preparaton, 3- culture preparaton, 4- human recourses preparaton 5- fnancal preparaton, 6- laws and regulatons preparaton, 7- content preparaton, 8- standard preparaton 9- management preparaton. Hardware preparaton, ncludng network preparaton, equpment preparaton and thrd dmensons ncludes: 1- Support Preparaton 2- coordnaton preparaton, 3- montorng preparaton. Each of these smaller dmensons s also ncluded benchmarks n the attached table. These benchmarks are 116, whch belong to a collecton that s clear n the table. The table becomes fuzzy n the next secton of modelng algorthm E-Learnng Fuzzy Modelng To fuzzcaton the proposed model Mamdan Inference System s used. In ths system, for each dmensons and ther characterstcs fuzzy varable are defned. Then on each of these dmensons wll form the base of fuzzy rules. It s should be noted the model n the software dmenson has feedback then ths dmensons wll be consdered as optmzaton crteron for genetc algorthm. Below you can see the fuzzy modelng. Fgure 2 Model of Fuzzy Rules Base for e-learnng Now results of the next secton added to the genetc algorthm Genetc Algorthm As descrbed n Secton for each ndcators fuzzy model was defned n ths secton we explan how to apply genetc algorthms, general trend s that when the chromosomes were not able to shft nto better postons, t does not move. In contrast to the standard model n the FGA f not fnd better postons n search behavor of model, randomly and freely move a step [1] and lose ts prevous poston, n the FGA f fal, they retaned ther prevous poston. Ths makes the best of the cluster placed n the best poston found so far by the members of the cluster because the behavor of the model n the FGA, only one chromosome s moved as to move nto a better poston. In the behavor of followng, each of the models a step towards the best cluster usng equaton 3: (3) X t 1 X t X X Ds, t Vsual Rand 0,1 Where X s of -th that performs the behavor of followng and X s equal to poston vector of best cluster. Thus, the ndex of -th maxmum sze of the vsual (feld of vew) n each dmensons move toward the best cluster. 282

3 In fact, after an ndex to fnd more space other members of the cluster followng that to acheve more space. Run the behavor of followng the best cluster ncreases the speed of cluster convergence and helps to mantan consstency n a cluster. Ths behavor s group behavor and nteracton between members of a cluster s done among all members of the cluster Collectve Moton Behavor of Dmensons Ths behavor s smlar to behavor of followng of a group behavor, and n the whole of the level of cluster members s done. In the group moton behavor, frst of all the central poston of cluster whch s same center of gravty of cluster based on the arthmetc mean of all members of cluster n each dmenson s calculated. To -th ndex condton of move toward central poston, whch s f (X) f (X) s checked, and f ths was true, the next poston of -th ndex, usng equaton (4) s obtaned: (4) X X X t 1 X t Ds, t Vsual Rand 0,1 Equaton 4 to move toward the central poston whch all the ndexes of cluster that ther stuaton s worse than the central poston s used but for the best ndex of cluster s placed n the X poston, If the ftness value X s better than X, next poston of best ndcator of the cluster s obtaned by equaton 5: (5) X X Equaton 5 s used for best cluster because movng toward the central poston usng equaton 4 may be placed n the worse poston than ts current poston. There may be worse postons at the end of way toward the central poston. As a result, ths could result to loss of the best poston ever found by all members of the cluster that usng the equaton 5 to best cluster ths problem resolved. Reasons for not usng equaton 5 for all crtera of cluster s that transfer crtera of cluster nto a dentcal poston dramatcally reduce the varaton n the group and come down dramatcally speed of convergence The Implementaton of the Proposed Model In the FGA for each of the ndcators, n each repeats, each of optmal search behavors, followng and collectve moton s performed. Unlke FGA standard that mplementaton one of both collectve moton and followng behavor had no effect on the optmal moton and a lot of calculatons would be waste, n the FGA every three optmal search behavors, followng and collectve moton n the dsplacement of the ndcators and moton of cluster toward better postons are effectve. In the FGA all aspects of optmal search behavor s performed at frst and ther status s updated based on the mplementaton of the. Wth the mplementaton of ths behavor, each dmenson can have shft up try_number. Then all of them based on ther new poston and other clusters that performed the optmal search behavor would perform followng behavor and all aspects except the aspect that has feedback (after soft preparaton), wll move nto a new poston n the way movement toward best poston that found by the cluster. At fnal, each aspect and ndcators of the collectve moton behavor s done. Wth the mplementaton of collectve behavor, ndcators that have remaned away from the cluster are placed on worse poston than other members of the cluster, more quckly returnng to cluster. Ths behavor ncreases the speed of convergence and unlke the followng behavor can even be mproved best poston of cluster Confguraton of the Model as Optmzaton In MPB, when there s more than one peak n the problem space, each peak can after change the settng to convert a global optmum, so all the peaks have to cover under aspects of e-learnng model f any of them were to become a global optmzaton, algorthm able to fnd t quckly. So at each peak shall be placed a cluster and covers t. In ths paper, n the begnnng there s only one cluster on the problem space, f ths cluster s convergng to a peak, another cluster comes up n the problem envronment. When the cluster s converged the poston of best ndcator of cluster n the problem space after a few teratons s approxmately fxed. When a cluster converged parameter value of the feld of vew s reduced to a cluster to do well local search n the vcnty of the target [9]. Each peak s only covered just under a cluster because ncreasng the number of clusters n a peak just consume an unreasonable of ftness assessment and as a result ncrease the speed of changng the envronment than teratons of algorthm performance. So when both of them close to each other more than certan lmt, that s, they have converged to a peak and therefore the cluster ftness of best dmenson of that s less, then loss. In the proposed algorthm, when all the dmensons n the problem envronment were converged, a new cluster n the problem space s ntalzed. In each repeat, there should be only one free cluster. Free cluster s cluster that stll not converged. When more than one cluster were free, only the best of them reman n the problem space and the rest are lost. After some mplementaton tme passed all the peaks are covered and there wll be a cluster free on the problem space. To detect changes n the envronment a random pont n the frst mplementaton algorthm are consdered and amount of ts ftness s recorded. In each repeat of the algorthm amount of ftness of ths pont s assessed and f t s dfferent from the prevous value t means the envronment has changed. After detectng changes n the envronment ftness of all artfcal fsh n all clusters s assessed and ftness value recorded for each artfcal fsh s vald and algorthm can then properly execute ts behavor. Then n each cluster, only the best ndcator of cluster stays n ts place and other members of the cluster wthn a certan radus around that are randomly dstrbuted. Values of the radus equal to length of peaks dsplacement are consdered. Thus, most lkely after change the peak (target) s located wthn members' feld of vew. The act also ncreases the dversty n the cluster. 283

4 3- Test Results To assess the accuracy and effcency of the proposed algorthm, ths algorthm wth sx known algorthms called mqso [13], mcpso [13], AmQSO [14], CellularPSO [15], SPSO [16] and rspso [17] on the MPB have been compared. Two algorthms m CPSO and mqso wth ant-convergence are gven n the comparson as well. Results have been obtaned wth respect to scenaro 2 n the MPB [12]. The only dfferent parameter s the number of peaks for better assessment methods changed from 1 peak to 200 peaks at a change frequency of 5000 has been consdered. In experments, n the proposed algorthm ndcators number of dmensons n each cluster equal to 2 and amount of Try_number equal to 1 are consdered. Vsual amount for those that have not converged equal to 20 and for those that have converged equal to 0.5 are consdered. The value of ths parameters based on experments s set too hgh. Experments were performed 30 tmes and the results obtaned from the experments (Offlne Error and Standard Error) on MPB wth 100 tmes envronment change at each experment n the dfferent peak number wth envronment changes of frequency 5000 n Table (2) shown. Standard Errors are shown n parentheses next to the Offlne Error. Two best results n each row are hghlghted and best result s n talc. P s equal to the number of peaks n these tables. Two algorthms wth mqso and mcpso wth antconvergence marked wth an astersk. As can be seen n the Table 2, the proposed algorthm n all cases has acheved better results than other algorthms. Snce the number of clusters n the proposed algorthm determned based on the number of peaks, so effcency of the algorthm n the low peaks s qute acceptable. Ths dstncton exsts n the AmQSO as well because n ths algorthm the number of clusters determned based on the number of peaks s found but n the other algorthms due to the constancy of the number of clusters, when the number of clusters s greater than the number of peaks, clusters compare wth each other over peaks and out together from the peaks usng exclusvty, whch thereby results were less than expected for the number of peaks. For the proposed algorthm ncreasng the number of peaks decrease the performance of algorthm. The problem arses when there are a large number of peaks; there wll be more clusters n the problem space. Increasng the number of clusters, the number rses to evaluate the ftness of each teraton, snce the frequency change of e-learnng model s based on amount of ftness assessment, t causes the dmensons has less teraton tll changes subsequent envronment. As a result, the accuracy of search can lower and the results encountered decrease n the qualty. AmQSO algorthm encountered ths problem more than proposed algorthm and the number of peaks 100 and 200 ts performance s also less than prevous model.e. mqso n whch the number of categores s fxed. In the proposed algorthm, due to the rapd ncrease n dversty after envronment change and quckly cover peak, results from the hgher number of peaks encountered less qualty loss. 4- Concluson Ths artcle for frst tme proposes a new method for e- learnng fuzzy modelng and optmzaton procedure was performed by genetc algorthm. And ts results on the moble peaks benchmarks functon were compared wth several other well-known methods. Results shown that the proposed algorthm has acceptable performance n all cases,.e. has effect on both low number of peaks and hgh number of peaks Table (2): Offlne Error and Standard Error of the algorthms n the 5000changes frequency wth dfferent peaks numbers References 1. Darab, b., (2011). An electc model for assessng e- learnng readness n the Iranan unverstes. 2. Asgharpour, M.J., (1386). Group Decson Makng and Games Theores wth Operaton Research Vew. 3. Bngham, R., (1388) Student Sklls, Scence and Technology Unversty. 4. Tayefeh, M. M., (1382). Tranng and Improvng Operatonal Sklls as an Electronc Learnng Servce n Learnng Envronment, a set of artcles of the frst Informaton Technology and Knowledge Internatonal Conference, Tehran, December. 5. Asgharpour, M. J., (1386). Group Decson Makng and Games Theores wth Operaton Research Vew. 6. Fallown, K.l, Brown, Sh., Bagher, F., Kouchak Mehd, H., (1387). Electronc Learnng Standards, Tehran, Madares Houshmand Publcaton. 7. Deylmaghan, M.,(1386). Challenges and Necesstes, a Member of Scentfc Board of Unversty Department of Industres Engneerng Unversty. 8. Tolou Ashlagh, A., (1387). Expert System, Unversty Courses, Scence and Research Unversty. 9. Tehran, S., Tadayon, Sh., (1380). Informaton Technology Management, Tehran, Frst publsh. 284

5 10. Hakmpour, A., Decson Makng n Management, second publsh. 11. Afshar Kazem, M. A., (1387). Expert Decson Makng, Scence & Research Unversty. 12. Whtman, M.E. and Mattord, H.J. (2009). "Management of nformaton securty", Thomson, second edton. 13. Bhaskar, S.M., and Ahson, S.I. (2008)."Informaton securty A practcal Approach," Alpha scence. 14. Turban, E., and Aronson, J.E, and Lang, T.P.,(2007). "Decson support systems and Intellgent systems," prentce- Hall, seven Edton, Alter, S., (2004)."Informaton systems", pearson Educaton, fourth Edton. 16. Whtman, M.E., and Mattord, H.d., (2006)."Readngs and cases n the management of Informaton securty,isen 17. Wthman, M.E. and Mattord, H.d., (2003)."Hands on Informaton securty lab Manual", second Edton. 18. Wllams, A.H.,; "In a trustng envronment, everyone s responsble for nformaton securty. 19. Krutz, R.l., and Vnes, R.D., (2007). "The CISM prep Gude: Masterng the fve Domans of Informaton securty management". John wley & sons, ISBN: Rahm, A. and Basr, otman; Hassanen, khaled (2005). "An Intellgent Decson support system For Layout system. 21. Decenzo, D. (2002) "Human Resource Management" Stephen. P. Robns. 22. Harrson, Tobas; wennstorm, Danel (2006)."A Decson support system for Executve Busness management" ABB. 23. Pouruakhshour, Z., and shattr, M., (2003)." Decson support system In ol spll cases" Dsaster preventon and management volume 2, Nubmber Serpell A., and Maturana, S., (2002)."A Decson supoort system for Human Resource Management In constructon Projects. 25. Swanepoel, K., (2007). "Decson support system: Real Tme Control of manufacturng processes" Journal of manufacturng Technology management, volume s, number 1 285