Kerangka Kerja Multi-Agen untuk Deteksi Konflik Kebutuhan Fungsional Berbasis Sistem Aturan

Jati, Andrea Bemantoro (2026) Kerangka Kerja Multi-Agen untuk Deteksi Konflik Kebutuhan Fungsional Berbasis Sistem Aturan. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025241047_Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Konflik pada kebutuhan fungsional merupakan salah satu permasalahan kritis dalam rekayasa perangkat lunak karena dapat menyebabkan ketidakkonsistenan dan kegagalan sistem. Pendekatan berbasis aturan (Rule-Based System/RBS) yang dikombinasikan dengan clustering terbukti efektif dalam mendeteksi konflik secara deterministik, namun hanya mampu menangkap pola konflik yang bersifat eksplisit. Konflik semantik yang lebih halus, seperti asimetri negasi, perbedaan modal verb, atau pertentangan kondisi-respons, umumnya tidak dapat dideteksi melalui pencocokan aturan sederhana. Penelitian ini mengusulkan kerangka kerja multi-agent yang mengintegrasikan RBS sebagai penyaring deterministik awal dengan lima semantic specialist agent, yaitu Lexical Agent, Numeric Agent, Logic Agent, Semantic Verb Agent, dan NLI Agent, serta LLM Coordinator berbasis Gemini sebagai pengintegrasi keputusan akhir. Pasangan kebutuhan yang tidak dapat diselesaikan oleh RBS dieskalasikan ke seluruh specialist agent, kemudian LLM Coordinator mensintesis laporan dari setiap agent menjadi keputusan akhir beserta justifikasi yang eksplisit. Evaluasi pada enam dataset kebutuhan perangkat lunak, yaitu pure, uav, opencoss, warehouse, worldvista, dan ecommerce, menggunakan tujuh metrik klasifikasi biner menunjukkan bahwa sistem yang diusulkan meningkatkan F1-score pada lima dari enam dataset, dengan peningkatan terbesar pada ecommerce sebesar 23,81%, diikuti oleh worldvista sebesar 8,55% dan pure sebesar 7,04%. Rata-rata F1-score meningkat sebesar 7,37%, dari 0,6654 menjadi 0,7390, dan meningkat sebesar 8,66% apabila dataset opencoss yang sangat tidak seimbang dikecualikan. Dari 21 pasangan False Negative yang tidak terdeteksi oleh RBS, sebanyak 57,1% berhasil diidentifikasi, dengan NLI Agent sebagai kontributor utama sebesar 83%. Evaluasi menggunakan F2-score, Balanced Accuracy, dan Matthews Correlation Coefficient (MCC) menunjukkan bahwa sistem secara keseluruhan unggul pada skenario yang berorientasi pada recall tinggi, sedangkan konfigurasi LLM terstruktur memberikan keseimbangan kinerja yang lebih baik secara keseluruhan.
=======================================================================================================================================
Conflicts in functional requirements are a critical issue in software engineering because they can lead to inconsistencies and system failures during design and implementation. Rule-Based System (RBS) approaches combined with clustering have proven effective in detecting conflicts deterministically; however, they are limited to identifying explicit conflict patterns. More subtle semantic conflicts, such as negation asymmetry, differences in modal verbs, or condition-response contradictions, are often overlooked by simple rule-matching techniques. This study proposes a hybrid multi-agent framework that integrates an RBS as the initial deterministic filter with five semantic specialist agents, namely the Lexical Agent, Numeric Agent, Logic Agent, Semantic Verb Agent, and NLI Agent, along with a Gemini-based LLM Coordinator that serves as the final decision integrator. Requirement pairs that cannot be resolved by the RBS are escalated to all specialist agents, after which the LLM Coordinator synthesizes the reports from each agent into a final decision accompanied by explicit justification. The proposed framework was evaluated using six software requirement datasets, namely pure, uav, opencoss, warehouse, worldvista, and ecommerce, across seven binary classification metrics. The results show that the proposed system improved the F1-score on five of the six datasets, with the largest improvement observed for ecommerce (+23.81%), followed by worldvista (+8.55%) and pure (+7.04%). The average F1-score increased by 7.37%, from 0.6654 to 0.7390, and by 8.66% when the highly imbalanced opencoss dataset was excluded. Of the 21 False Negative requirement pairs missed by the RBS, 57.1% were successfully recovered, with the NLI Agent contributing 83% of the recovered cases. Additional evaluation using the F2-score, Balanced Accuracy, and Matthews Correlation Coefficient (MCC) indicates that the complete system performs best in high-recall scenarios, while the structured LLM configuration provides a more balanced overall performance.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Deteksi konflik kebutuhan fungsional, LLM Coordinator, Multiagent, NLI, Rule-Based System
Subjects:	T Technology > T Technology (General) > T57.5 Data Processing
Divisions:	Faculty of Information Technology > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Andrea Bemantoro Jati
Date Deposited:	31 Jul 2026 02:33
Last Modified:	31 Jul 2026 02:33
URI:	http://repository.its.ac.id/id/eprint/140708

Actions (login required)

View Item