MARCOS, Hendra and Gernowo, Rahmat and Wibowo, Adi and Tahyudin, Imam (2026) PENGEMBANGAN MODEL PENGENDALIAN SINYAL LALU LINTAS ADAPTIF MENGGUNAKAN MULTI-OBJECTIVE DEEP REINFORCEMENT LEARNING DENGAN MODIFIKASI REWARD. Doctoral thesis, UNIVERSITAS DIPONEGORO.

	Text cover aja.pdf Download (187kB)
	Text Cover_Full.pdf Restricted to Repository staff only Download (5MB)
	Text BAB I.pdf Download (249kB)
	Text BAB II.pdf Download (1MB)
	Text BAB III.pdf Restricted to Repository staff only Download (893kB)
	Text BAB IV.pdf Restricted to Repository staff only Download (1MB)
	Text BAB V.pdf Restricted to Repository staff only Download (151kB)
	Text Daftar Pustaka.pdf Download (200kB)
	Text Lampiran.pdf Restricted to Repository staff only Download (1MB)

Abstract

Penelitian ini mengembangkan model pembelajaran penguatan deep reinforcement learning multi-objektif untuk optimasi pengendalian sinyal lalu lintas adaptif berbasis multi-agent. Tujuan utama penelitian adalah untuk mengatasi tantangan pengelolaan lalu lintas perkotaan yang kompleks, dengan memperkenalkan system yang dapat beradaptasi dengan kondisi lalu lintas yang dinamis dan meminimalkan kemacetan, waktu tunggu, serta konsumsi bahan bakar secara bersamaan. Model yang diajukan menggunakan Dueling Double Deep Q-Network (D3QN) untuk meningkatkan stabilitas dan kinerja dalam multi-agent reinforcement learning (MARL). Sistem ini dibangun dengan agen-agen yang independen, di mana setiap traffic light (lampu lalu lintas) bertindak sebagai agen yang belajar mengoptimalkan waktu hijau dengan memperhatikan kondisi lalu lintas lokal dan informasi dari agen-agen tetangga. Curriculum reward shaping diterapkan untuk mengatasi masalah kompleksitas desain fungsi reward multi-objektif yang melibatkan waktu tunggu, panjang antrian, throughput, dan efisiensi bahan bakar. Evaluasi dilakukan di lingkungan simulasi SUMO (Simulation of Urban MObility) untuk berbagai skenario lalu lintas, termasuk jam sibuk dan kondisi over-saturated. Hasil eksperimen menunjukkan bahwa model ini mampu mengurangi waktu tunggu kendaraan hingga 48% dibandingkan dengan pengendalian sinyal tetap, serta menurunkan konsumsi bahan bakar lebih dari 21% dibandingkan dengan model DRL berbasis objektif tunggal. Selain itu, strategi curriculum learning yang diterapkan mempercepat konvergensi model dan meningkatkan stabilitas pelatihan. Penelitian ini memberikan kontribusi penting terhadap pengelolaan lalu lintas cerdas yang lebih efisien dan berkelanjutan, serta menunjukkan potensi penerapan model multi-agent DRL dalam skenario dunia nyata dengan peningkatan koordinasi antar agen.
Kata kunci : Adaptif, Deep Reinforcement Learning, Multi-Agent, Optimasi Sinyal Lalu Lintas, Curriculum Reward Shaping

This study develops a multi-objective deep reinforcement learning model for optimizing adaptive traffic signal control based on a multi-agent system. The main objective of this research is to address the challenges of complex urban traffic management by introducing a system that can adapt to dynamic traffic conditions while simultaneously minimizing congestion, vehicle waiting time, and fuel consumption. The proposed model employs a Dueling Double Deep Q-Network (D3QN) to improve stability and performance in a multi-agent reinforcement learning environment, where each traffic light operates as an independent agent that learns to optimize green signal timing by considering local traffic conditions and information from neighboring agents. Curriculum reward shaping is applied to overcome the complexity of designing a multi-objective reward function involving waiting time, queue length, throughput, and fuel efficiency. The evaluation was conducted using the Simulation of Urban Mobility (SUMO) environment under various traffic scenarios, including peak-hour and oversaturated conditions. Experimental results show that the proposed model reduces vehicle waiting time by up to 48% compared with fixed-time signal control and decreases fuel consumption by more than 21% compared with single-objective DRL models. In addition, the applied curriculum learning strategy accelerates model convergence and improves training stability. This research contributes to the development of more efficient and sustainable intelligent traffic management systems and demonstrates the potential implementation of multi-agent deep reinforcement learning in real-world traffic scenarios through improved interagent coordination.
Keywords: Adaptive Traffic Signal Control, Deep Reinforcement Learning, MultiAgent, Traffic Signal Optimization, Curriculum Reward Shaping.

Actions (login required)

View Item