Pengembangan Pembelajaran Hibrida yang Mengintegrasikan Extreme Learning Machine dan Feedforward Neural Network

Al-Khanza, Gissella Nasywa (2026) Pengembangan Pembelajaran Hibrida yang Mengintegrasikan Extreme Learning Machine dan Feedforward Neural Network. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5002221109-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Extreme Learning Machine (ELM) merupakan metode pembelajaran yang memiliki proses pelatihan cepat karena bobot hidden layer ditetapkan acak dan bobot output layer diperoleh melalui penyelesaian least squares secara analitik. Meskipun demikian, mekanisme tersebut menyebabkan kualitas representasi fitur yang dihasilkan sangat bergantung pada inisialisasi bobot acak. Di sisi lain, Feedforward Neural Network (FNN) mampu membentuk representasi fitur yang adaptif melalui optimisasi berbasis gradien, tetapi memerlukan proses pelatihan iteratif dengan biaya komputasi yang lebih tinggi. Tugas akhir ini mengusulkan metode pembelajaran hibrida Feedforward Neural Network-Extreme Learning Machine (FNN–ELM) yang mengintegrasikan kemampuan FNN dalam membentuk representasi fitur adaptif dengan mekanisme penentuan bobot output layer pada ELM. Pada metode yang diusulkan, parameter hidden layer dipelajari secara
bertahap melalui optimisasi berbasis gradien, kemudian dibekukan (freeze), sedangkan bobot output layer ditentukan melalui penyelesaian masalah least squares. Kinerja metode
dievaluasi pada dua kategori permasalahan, yaitu permasalahan berbasis simulasi yang meliputi aproksimasi fungsi nonlinier dan penyelesaian persamaan diferensial Bratu 1D menggunakan pendekatan Physics-Informed Neural Network (PINN), serta permasalahan berbasis data berupa regresi pada California Housing Dataset. Evaluasi dilakukan
menggunakan Mean Squared Error (MSE), Root Mean Squared Error (RMSE), koefisien determinasi (R2), dan relative L2 error. Hasil evaluasi menunjukkan bahwa FNN-ELM konsisten menghasilkan akurasi yang lebih tinggi dibandingkan FNN standar dan ELM standar pada kedua kategori permasalahan yang diuji. Selain itu, FNN–ELM juga membutuhkan waktu pelatihan yang sebanding dengan FNN standar, bahkan lebih
singkat pada beberapa kasus. Hasil tersebut menunjukkan bahwa integrasi FNN dan ELM mampu memanfaatkan keunggulan kedua metode, sehingga menjadi alternatif metode pembelajaran yang efektif untuk permasalahan berbasis simulasi maupun berbasis data.
=====================================================================================================================================
Extreme Learning Machine (ELM) is a learning method characterized by a fast training process, in which the hidden layer weights are randomly initialized, while the
output layer weights are determined analytically by solving a least squares problem. However, this mechanism makes the quality of the learned feature representations highly dependent on the random initialization of the hidden-layer weights. In contrast, Feedforward Neural Network (FNN) can learn adaptive feature representations through gradient-based optimization, but requires an iterative training process with higher computational cost. This undergraduate thesis proposes a hybrid learning method, namely Feedforward Neural Network–Extreme Learning Machine (FNN–ELM), which integrates the adaptive feature representation capability of FNN with the ELM mechanism for determining the output layer weights. In the proposed method, the hidden layer parameters are learned progressively through gradient-based optimization and subsequently frozen, whereas the output layer weights are determined by solving a least squares problem. The performance of the proposed method is evaluated on two categories of problems: simulation-based problems, including nonlinear function approximation and the solution of the one-dimensional Bratu differential equation using the Physics-Informed Neural Network (PINN) approach, and data-driven problems, namely regression on the California Housing Dataset. The evaluation is conducted using Mean Squared Error (MSE), Root Mean Squared Error (RMSE), the coefficient of determination (R2), and the relative L2 error. The experimental results demonstrate that FNN–ELM consistently
achieves higher prediction accuracy than both the standard FNN and the standard ELM across the two categories of problems. Furthermore, FNN–ELM requires a training time
comparable to that of the standard FNN, and even shorter in several cases. These results indicate that the integration of FNN and ELM effectively combines the strengths of both
methods, making FNN–ELM an effective alternative learning method for both simulation based and data-driven problems.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Extreme Learning Machine,Feedforward Neural Network, Hybrid FNN-ELM, Aproksimasi Fungsi, Persamaan Diferensial, Function Approximation, Differential Equations
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Gissella Nasywa Al-khanza
Date Deposited:	04 Aug 2026 01:56
Last Modified:	04 Aug 2026 01:56
URI:	http://repository.its.ac.id/id/eprint/142704

Actions (login required)

View Item