Implementasi Deep Reinforcement Learning dengan Arsitektur Multi-Layer Perceptron dan Graph Neural Network untuk Penjadwalan Proyek Pemeliharaan Pabrik

Avin, Moch. (2026) Implementasi Deep Reinforcement Learning dengan Arsitektur Multi-Layer Perceptron dan Graph Neural Network untuk Penjadwalan Proyek Pemeliharaan Pabrik. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221061-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (7MB) | Request a copy

Abstract

Resource-Constrained Project Scheduling Problem (RCPSP) pada proyek pemeliharaan pabrik merupakan permasalahan penjadwalan dimana setiap task memiliki hubungan ketergantungan dan dibatasi oleh ketersediaan sumber daya. Penjadwalan secara konvensional sering menghasilkan durasi proyek (makespan) yang kurang optimal. Penelitian ini menggunakan Deep Reinforcement Learning (DRL) karena proses penjadwalan dapat dimodelkan sebagai pengambilan keputusan sekuensial, di mana agen belajar memilih task berdasarkan kondisi proyek. Penelitian ini membandingkan empat kombinasi model, yaitu DQN-MLP, DQN-GNN, PPO-MLP, dan PPO-GNN. Deep Q-Network (DQN) digunakan sebagai algoritma berbasis nilai, sedangkan Proximal Policy Optimization (PPO) digunakan sebagai algoritma berbasis kebijakan. Multi-Layer Perceptron (MLP) merepresentasikan kondisi proyek sebagai vektor fitur, sementara Graph Neural Network (GNN) merepresentasikan proyek sebagai graf ketergantungan antar task. Keempat model dilatih menggunakan dataset proyek pemeliharaan pabrik tahun 2024, 2025, dan 2026. Evaluasi dilakukan dengan membandingkan hasil jadwal model terhadap jadwal manual dan batas bawah teoritis Critical Path Method (CPM) berdasarkan nilai makespan. Hasil pengujian menunjukkan bahwa DRL mampu menurunkan makespan hingga 2,16% pada dataset 2024 dan 1,64% pada dataset 2026 dibandingkan jadwal manual. Hasil penjadwalan divisualisasikan melalui antarmuka berbasis web untuk membantu analisis jadwal dan beban sumber daya.
=====================================================================================================================================
Resource-Constrained Project Scheduling Problem (RCPSP) in plant maintenance projects is a scheduling problem in which each activity has dependency relationships and is constrained by resource availability. Conventional scheduling often results in a less optimal project duration, or makespan. This study uses Deep Reinforcement Learning (DRL) because the scheduling process can be modeled as sequential decision-making, where an agent learns to select activities based on the project condition. This study compares four model combinations, namely DQN-MLP, DQN-GNN, PPO-MLP, and PPO-GNN. Deep Q-Network (DQN) is used as a value-based algorithm, while Proximal Policy Optimization (PPO) is used as a policy-based algorithm. Multi-Layer Perceptron (MLP) represents the project condition as a feature vector, while Graph Neural Network (GNN) represents the project as a graph of activity dependencies. The four models are trained using plant maintenance project datasets from 2024, 2025, and 2026. The evaluation is conducted by comparing the schedules generated by the models with the manual schedule and the theoretical lower bound obtained from the Critical Path Method (CPM), based on the makespan value. The experimental results show that DRL can reduce the makespan by up to 2.16% on the 2024 dataset and 1.64% on the 2026 dataset compared with the manual schedule. The resulting schedules are visualized through a web-based interface to support schedule and resource load analysis.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Deep Reinforcement Learning, RCPSP, DQN, PPO, Graph Neural Network
Subjects:	T Technology > T Technology (General) > T57.6 Operations research--Mathematics. Goal programming T Technology > T Technology (General) > T58.8 Productivity. Efficiency
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Moch. Avin
Date Deposited:	24 Jul 2026 03:32
Last Modified:	24 Jul 2026 03:32
URI:	http://repository.its.ac.id/id/eprint/136889

Actions (login required)

View Item