Stowage Planning Assistant Berbasis CBR Dan GRAPHRAG Dengan LLM Berdasarkan Pengalaman Historis Dan Aturan Kontekstual

Wieansah, Muhammad Razan (2026) Stowage Planning Assistant Berbasis CBR Dan GRAPHRAG Dengan LLM Berdasarkan Pengalaman Historis Dan Aturan Kontekstual. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5002211168-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Perencanaan penempatan kontainer pada kapal (stowage planning) merupakan proses penting untuk menjamin keselamatan dan efisiensi operasi bongkar-muat, namun menuntut pemenuhan berbagai kendala teknis dan operasional secara bersamaan. Untuk menjawab tantangan tersebut, penelitian ini mengembangkan asisten perencanaan stowage (Stowage Planning Assistant) yang mengintegrasikan Case-Based Reasoning (CBR), GraphRAG, dan Large Language Model (LLM) yang diorkestrasikan melalui LangGraph untuk menghasilkan rencana penempatan kontainer (stowage plan) berbasis pengalaman historis dan aturan kontekstual. Sistem melakukan temu-balik kasus (case) pada penyimpanan vektor (Milvus) dan temu-balik kendala pada penyimpanan graf (Neo4j), kemudian menggabungkannya menjadi paket kendala (constraint packet) sebagai acuan generasi, validasi deterministik, dan siklus perbaikan. Evaluasi sistem dilakukan melalui pengujian pada 30 kueri Container Loading List (CLL) menggunakan metrik top-k berbasis normalized Discounted Cumulative Gain (nDCG); skor agregat didefinisikan sebagai rata-rata nDCG jalur CBR dan GraphRAG, menghasilkan Score@5 = 0.907 dan Score@10 = 0.958. Pengujian stowage plan pada skenario baseline 678 TEU dan near-stress 1189 TEU menghasilkan rencana lengkap pada kedua skenario, dengan seluruh kontainer teralokasi ke slot yang tersedia. Pada skenario near-stress, rencana baru menunjukkan kepatuhan kendala yang lebih baik dibanding kasus rujukan melalui penurunan pelanggaran dan penalti hingga mencapai status Feasible, sedangkan pada skenario baseline rencana berada pada status Caution. Hasil evaluasi sistem menunjukkan grounding sebesar 100% dan rata-rata faithfulness sebesar 66,5%, mengindikasikan mayoritas klaim didukung konteks, namun validasi deterministik dan siklus perbaikan tetap diperlukan untuk meminimalkan inkonsistensi residual pada generasi berbasis LLM.
====================================================================================================================================
Container stowage planning is a critical process for ensuring the safety and operational efficiency of loading and unloading activities, yet it requires satisfying multiple technical and operational constraints simultaneously. This study develops a Stowage Planning Assistant that integrates Case-Based Reasoning (CBR), GraphRAG, and a Large Language Model (LLM) orchestrated via LangGraph to generate container stowage plans grounded in historical experience and contextual rules. The system retrieves similar cases from a vector store (Milvus) and retrieves relevant constraints from a graph store (Neo4j), then combines them into a constraint packet to guide plan generation, deterministic validation, and an iterative repair loop. The system is evaluated on 30 Container Loading List queries using top-k ranking metrics based on normalized Discounted Cumulative Gain (nDCG); the aggregate score is defined as the mean nDCG of the Case-Based Reasoning and GraphRAG retrieval paths, achieving Score@5 = 0.907 and Score@10 = 0.958. Testing on a baseline scenario of 678 TEU and a near-stress scenario of 1189 TEU produces complete plans in both scenarios, with all containers successfully allocated to available ship slots. In the near-stress scenario, the generated plan demonstrates better constraint compliance than the reference case, as indicated by reduced violations and penalties until it reaches a Feasible status, whereas in the baseline scenario, the plan remains at a Caution status. The system evaluation results show 100% grounding and 66.5% faithfulness average, indicating that most claims are supported by the retrieved context. Nevertheless, deterministic validation and the repair loop remain necessary to minimize residual inconsistencies in LLM-based generation.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Perencanaan Penempatan Kontainer, Case-Based Reasoning, Large Language Model, Retrieval-Augmented Generation, GraphRAG, LangGraph, Stowage Planning, Case-Based Reasoning, Large Language Model, Retrieval-Augmented Generation, GraphRAG, LangGraph
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence
Divisions:	Faculty of Mathematics and Science > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Muhammad Razan Wieansah
Date Deposited:	28 Jan 2026 05:42
Last Modified:	28 Jan 2026 05:42
URI:	http://repository.its.ac.id/id/eprint/130771

Actions (login required)

View Item