Hasil Pencarian  ::  Simpan CSV :: Kembali

"One of the problems faced in applying neural network to some realworld application is related to difticulties in finding an optimum set of weightsand thresholds during the training phase. A general most method in tindingthese solutions for these problems is backpropagation.A different method to tind the solutions of the same problems isGenetic Algorithms. Genetic algorithm is relatively new search algorithm thathas not been fully explored in this area. ln this thesis, genetic algorithms areapplied to train neural networks and to evolve an optimum set of weights andthresholds. Process begin with encode neural networks parameters to binarychromosomes, and evaluate. The Spinning wheel selections are using toproduce offspring with high titness_ then recombinate with crossover andmutation as genetic operator.The proiect carried out investigates whether genetic atgonthms can beapplied to neural networks to solve pattem classitication and functionapproximation problems. This thesis describes tl1e simulation works thathave been perfomwed. It describes the design ofa genetic algorithm and theresults obtained. ln pattem classilication problem that use feedforwardnetwork show, that genetic algorithm is superior to backpropagation trainingrule in error and speed calculation. ln function approximation, the resultshows that genetic algorithm approach is very much slower than thebackpropagation method. Results' show that even for relatively simplenetwork, genetic algorithm requires a much longer time to Uain neuralnetworks-"

"Salah satu permasalahan yang dihadapi dalam mengaplikasikan neural network adalah menentukan parameter-parameter weight dan threshold yang optimum selama fasa pelatihan. Metode yang umum digunakan untuk mendapatkan solusi permasalahan ini adalah metode backpropagation. Suatu pendekatan berbeda yang digunakan untuk mendapatkan solusi dari permasatahan diatas adalah algoritma genetik. Dalam tesis ini algoritma genetik diaplikasikan untuk melatih neural network guna mendapatkan suatu parameter weight dan threshold yang optimum. Proses diawali dengan mengkodekan parameter-parameter neural network menjadi kromosom biner, yang kemudian dilanjutkan dengan suatu proses evaluasi kromosom. Proses seleksi dengan metode 'Spinning Wheel' digunakan untuk menyeleksi turunan dengan kelayakan tinggi. Proses pencarian solusi optimal dikerjakan dengan melakukan operator-operator genetik persilangan dan mutasi dari kromosom yang terseleksi. Hasil pengujian menunjukkan bahwa pelatihan dengan algoritma genetik untuk permasalahan klasifikasi pola terbukti lebih unggul kinerjanya daripada dengan metode backpropagation untuk mencapai error minimum yang diinginkan. Pada pengujian pendekatan fungsi, algoritma genetik ter1ihat lebih tambat dari segi waktu untuk mencapai error minimum yang sama dibandingkan dengan metode backpropagation.

---

One of the problems faced in applying neural network to some real wond application is related to difficulties in finding an optimum set of weights and thresholds during the training phase. A general most method in finding these solutions for these problems is backpropagation. A different method to find the solutions of the same problems is Genetic Algorithms. Genetic algorithm is relatively new search algorithm that has not been fully explored in this area. In this thesis, genetic algorithms are applied to train neural networks and to evolve an optimum set of weights and thresholds. Process begin with encode neural networks parameters to binary chromosomes, and evaluate. The Spinning wheel selections are using to produce offspring with high fitness, then recombinate with crossover and mutation as genetic operator. The project carried out investigates whether genetic algorithms can be applied to neural networks to solve pattern classification and function approximation problems. This thesis describes the simulation works that have been performed. It describes the design of a genetic algorithm and the results obtained. In pattern classification problem that use feedforward network show, that genetic algorithm is superior to backpropagation training rule in error and speed calculation. In function approximation, the result shows that genetic algorithm approach is very much slower than the backpropagation method. Results show that even for relatively simple network, genetic algorithm requires a much longer time to train neural networks."