Analisis Implementasi LLLM Sebagai Penyeleksi Algoritma Metaheuristik untuk Kasus Prioritas Pesanan dan Penjadwalan Pengiriman dalam Rantai Pasok FMCG

Sembiring, Yoda Nata Orlando Sembiring (2026) Analisis Implementasi LLLM Sebagai Penyeleksi Algoritma Metaheuristik untuk Kasus Prioritas Pesanan dan Penjadwalan Pengiriman dalam Rantai Pasok FMCG. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5010221062-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Industri FMCG menghadapi kompleksitas distribusi yang tinggi akibat besarnya volume permintaan, keterbatasan armada, dan luasnya wilayah layanan, sehingga berdampak pada rendahnya pencapaian indikator On-Time In-Full (OTIF), terutama untuk pengiriman ke luar Pulau Jawa. Permasalahan ini meliputi order prioritization dan delivery scheduling yang termasuk ke dalam Capacitated Vehicle Routing Problem with Time Windows (CVRPTW) yang bersifat NP-hard, sehingga metode eksak sulit diterapkan secara efisien pada skala operasional nyata. Di sisi lain, No Free Lunch Theorem menyatakan bahwa tidak ada satu algoritma metaheuristik yang selalu unggul untuk seluruh karakteristik permasalahan, sehingga diperlukan mekanisme pemilihan algoritma yang adaptif sesuai dengan karakteristik setiap instance distribusi. Sehingga penelitian ini mengusulkan pendekatan LLM-Assisted Algorithm Selection dengan memanfaatkan model Phi-3-mini yang di-fine-tune menggunakan teknik QLoRA pada 500 instance eksperimen berbasis data historis distribusi pada kasus CVRPTW untuk memilih algoritma terbaik antara Hybrid GA-ALNS, dan Green Anaconda Optimization (GAO) pada kasus pemilihan dua algoritma. Dan algoritma GA, ALNS, serta GAO untuk kasus pemilihan tiga algoritma. Algoritma-algoritma tersebut diuji pada kasus penjadwalan pengiriman yang memiliki karakteristik Capacitated Vehicle Routing Problem with Time Windows (CVRPTW). Hasil pengujian dua algoritama menunjukkan model mencapai akurasi prediksi algoritma sebesar 98,41% dengan Macro F1-score sebesar 0,83. Dibandingkan metode Nearest Neighbor Heuristic, pendekatan yang diusulkan mampu menurunkan Total Weighted Tardiness (TWT) rata-rata sebesar 35,3%, meningkatkan OTIF keseluruhan sebesar 23,5%, serta meningkatkan OTIF segmen MT Outer Island sebesar 25,1% dengan waktu komputasi yang tetap layak untuk operasional harian. Pengembangan model awal dengan dua algoritma menunjukkan bahwa ketidakseimbangan kelas menyebabkan collapsed class pada Phi-3-mini. Sehingga dilakukan penyederhanaan menjadi tiga algoritma, penerapan SMOTE, dan penggunaan Mistral-7B yang berhasil meningkatkan akurasi menjadi 78% (Macro F1 0,77), meskipun Random Forest masih lebih unggul pada data tabular. Penelitian ini membuktikan bahwa integrasi LLM sebagai decision layer eksternal untuk pemilihan algoritma metaheuristik dapat meningkatkan konsistensi dan kualitas solusi penjadwalan distribusi pada industri FMCG, terutama untuk wilayah luar pulau.
===================================================================================================================================
The Fast-Moving Consumer Goods (FMCG) industry faces significant distribution challenges due to high demand volumes, limited fleet capacity, and extensive service coverage, resulting in difficulties in achieving the On-Time In-Full (OTIF) target, particularly for deliveries to regions outside Java Island. These challenges involve order prioritization and delivery scheduling, which can be formulated as a Capacitated Vehicle Routing Problem with Time Windows (CVRPTW), an NP-hard optimization problem that makes exact solution methods impractical for real-world operations. Furthermore, the No Free Lunch Theorem suggests that no single metaheuristic algorithm consistently outperforms others across all problem instances, highlighting the need for an adaptive algorithm selection mechanism. This study proposes an LLM-Assisted Algorithm Selection framework by fine-tuning the Phi-3-mini model using the QLoRA technique on 500 experimental instances derived from historical distribution data represented as CVRPTW instances. Two experimental settings were investigated: a two-algorithm selection problem involving Hybrid GA-ALNS and Green Anaconda Optimization (GAO), and a three-algorithm selection problem involving GA, ALNS, and GAO. The selected algorithms were subsequently applied to distribution scheduling problems with CVRPTW characteristics. For the two-algorithm setting, the fine-tuned model achieved a Top-1 algorithm selection accuracy of 98.41% with a Macro F1-score of 0.83. Compared with the Nearest Neighbor Heuristic, the proposed approach reduced the average Total Weighted Tardiness (TWT) by 35.3%, improved the overall OTIF by 23.5%, and increased the MT Outer Island OTIF by 25.1%, while maintaining computational times suitable for daily operational use. However, the initial model suffered from severe class imbalance, resulting in a collapsed-class problem. To address this issue, the candidate algorithms were restructured into a three-algorithm setting, SMOTE was applied for data balancing, and the model was upgraded to Mistral-7B. This improved the algorithm selection accuracy to 78% with a Macro F1-score of 0.77, although Random Forest still outperformed the LLM on purely tabular data. Overall, the results demonstrate that employing an LLM as an external decision layer for metaheuristic algorithm selection can improve the consistency and quality of distribution scheduling decisions in FMCG logistics, particularly for deliveries to outer-island regions.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Algorithm Selection, Delivery Scheduling, FMCG, Metaheuristik, Large Language Model
Subjects:	T Technology > TS Manufactures > TS155 Production control. Production planning. Production management T Technology > TS Manufactures > TS157.5 Production scheduling
Divisions:	Faculty of Industrial Technology > Industrial Engineering > 26201-(S1) Undergraduate Thesis
Depositing User:	Yoda Nata Orlando Sembiring
Date Deposited:	30 Jul 2026 01:26
Last Modified:	30 Jul 2026 01:26
URI:	http://repository.its.ac.id/id/eprint/140051

Actions (login required)

View Item