Deep Reinforcement Learning Based Medium Access Control Protocol for LoRa Networks. Masters thesis, King Fahd University of Petroleum and Minerals.

PDF KFUPM_Thesis_Norah.pdf Restricted to Repository staff only until 20 April 2027. Download (5MB)

Arabic Abstract

ﺗﻌﺎﻧﻲ ﺷﺒﻜﺎﺕ LoRaWAN ﻣﻦ ﻣﺤﺪﻭﺩﻳﺔ ﻋﺎﻟﻴﺔ ﻓﻲ ﺍﻟﻘﺪﺭﺓ ﻋﻠﻰ ﺍﻟﺘﻮﺳّﻊ ﻭﺍﺭﺗﻔﺎﻉ ﻛﺒﻴﺮ ﻓﻲ ﻣﻌﺪﻝ ﺍﻟﺘﺼﺎﺩﻣﺎﺕ ﻧﺘﻴﺠﺔ ﺍﻋﺘﻤﺎﺩﻫﺎ ﻋﻠﻰ ﺑﺮﻭﺗﻮﻛﻮﻝ ﺍﻟﻮﺻﻮﻝ ﺍﻟﻌﺸﻮﺍﺋﻲ ALOHA، ﻣﻤﺎ ﻳﺆﺩﻱ ﺇﻟﻰ ﺍﻧﺨﻔﺎﺽ ﻣﻮﺛﻮﻗﻴﺔ ﺍﻹﺭﺳﺎﻝ ﺑﺸﻜﻞ ﻭﺍﺿﺢ ﻋﻨﺪ ﺍﺯﺩﻳﺎﺩ ﻋﺪﺩ ﺍﻷﺟﻬﺰﺓ ﻓﻲ ﺍﻟﺸﺒﻜﺔ. ﺗﻬﺪﻑ ﻫﺬﻩ ﺍﻟﺮﺳﺎﻟﺔ ﺇﻟﻰ ﺗﻄﻮﻳﺮ ﺑﺮﻭﺗﻮﻛﻮﻝ ﻭﺻﻮﻝ ﻳﻌﺘﻤﺪ ﻋﻠﻰ ﺍﻟﺘﻌﻠّﻢ ﺍﻟﻌﻤﻴﻖ ﺍﻟﻤﻌﺰﺯ ﺑﺎﺳﺘﺨﺪﺍﻡ ﺧﻮﺍﺭﺯﻣﻴﺔ ﺍﻟﺸﺒﻜﺎﺕ ﺍﻟﻌﺼﺒﻴﺔ ﺍﻟﻌﻤﻴﻘﺔ (DQN)، ﻭﺫﻟﻚ ﺑﻬﺪﻑ ﺗﺤﺴﻴﻦ ﺍﻷﺩﺍء ﺍﻟﺸﺒﻜﻲ ﻭﺗﺤﻘﻴﻖ ﻭﺻﻮﻝ ﺃﻛﺜﺮ ﻛﻔﺎءﺓ ﻭﻣﻮﺛﻮﻗﻴﺔ ﻓﻲ ﺍﻟﺒﻴﺌﺎﺕ ﻛﺜﻴﻔﺔ ﺍلأﺟﻬﺰﺓ. ﺗﻢّ ﻧﻤﺬﺟﺔ ﻣﺸﻜﻠﺔ ﺍﺧﺘﻴﺎﺭ ﻣﻌﻠﻤﺎﺕ ﺍﻹﺭﺳﺎﻝ ﻣﺜﻞ ﺍﻟﻘﻨﺎﺓ، ﻋﺎﻣﻞ ﺍﻻﻧﺘﺸﺎﺭ، ﻭﻗﻮﺓ ﺍﻹﺭﺳﺎﻝ، ﻋﻠﻰ ﺷﻜﻞ ﻋﻤﻠﻴﺔ ﻗﺮﺍﺭ ﺫﻛﻴﺔ، ﻣﻊ ﺗﺼﻤﻴﻢ ﻋﺎﻣﻞ  ﺗﻌﻠّﻤﻲ ﻣﻮﺯّﻉ ﻳﻘﻮﻡ ﻣﻦ ﺧﻼﻟﻪ ﻛﻞ ﺟﻬﺎﺯٍ ﺑﺎﻟﺘﻌﻠﻢ ﺫﺍﺗﻴًﺎ ﺍﻋﺘﻤﺎﺩًﺍ ﻋﻠﻰ ﺗﻐﺬﻳﺔ ﺭﺍﺟﻌﺔ ﺛﻨﺎﺋﻴﺔ (ﻧﺠﺎﺡ او ﻓﺸﻞ) ﺇﺿﺎﻓﺔً ﺇﻟﻰ ﻣﻌﺪﻝ ﺍﻟﻨﺠﺎﺡ ﺍﻟﺬﻱ ﻳﻌﻜﺲ ﺣﺎﻟﺔ ﺍﻻﺯﺩﺣﺎﻡ ﻓﻲ ﺍﻟﺸﺒﻜﺔ. ﺗﻢّ ﺗﻄﻮﻳﺮ ﺑﻴﺌﺔ ﻣﺤﺎﻛﺎﺓ ﻣﺨﺼﺼﺔ ﻣﺒﻨﻴﺔ ﻋﻠﻰ ﻣﺤﺎﻛﻲ LoRaSim ﺗﺘﻀﻤﻦ ﻧﻤﺎﺫﺝ ﻓﻴﺰﻳﺎﺋﻴﺔ ﺩﻗﻴﻘﺔ ﺗﺸﻤﻞ ﻓﻘﺪ ﺍﻟﻤﺴﺎﺭ، ﺣﺴﺎﺳﻴﺔ ﺍﻟﻤﺴﺘﻘﺒﻞ، ﺯﻣﻦ ﺍﻹﺭﺳﺎﻝ، ﻭﻧﻤﻮﺫﺟًﺎ ﺷﺎﻣﻼً ﻟﻠﺘﺼﺎﺩﻡ ﻭﺍﻻﺳﺘﻼﻡ. ﺃﻅﻬﺮﺕ ﺍﻟﻨﺘﺎﺋﺞ ﺃﻥ ﺍﻟﺒﺮﻭﺗﻮﻛﻮﻝ ﺍﻟﻤﻘﺘﺮﺡ ﺣﻘﻖ ﺃﻋﻠﻰ ﻧﺴﺒﺔ ﺗﺴﻠﻴﻢ ﻟﻠﺤﺰﻡ (PDR) ﻓﻲ ﺟﻤﻴﻊ ﺍﻟﺴﻴﻨﺎﺭﻳﻮﻫﺎﺕ ﺍﻟﺘﻲ ﺗﻢ ﺍﺧﺘﺒﺎﺭﻫﺎ، ﺣﻴﺚ ﻭﺻﻞ ﺇﻟﻰ % 99.07 ﻋﻨﺪ 100 ﺟﻬﺎﺯ ﻭ% 95.84 ﻋﻨﺪ 500 ﺟﻬﺎﺯ، ﻣﺘﻔﻮﻗًﺎ ﺑﻮﺿﻮﺡ ﻋﻠﻰ ﺑﺮﻭﺗﻮﻛﻮﻝ ALOHA ﻭﺍﻟﺘﻌﻠّﻢ ﺍﻟﻤﻌﺘﻤﺪ ﻋﻠﻰ ﺍﻟﺠﺪﺍﻭﻝ (Q-learning). ﻛﻤﺎ ﺃﻅﻬﺮ ﺍﻟﺒﺮﻭﺗﻮﻛﻮﻝ ﺍﻧﺨﻔﺎﺿًﺎ ﻣﻠﺤﻮﻅًﺎ ﻓﻲ ﻋﺪﺩ ﺍﻟﺘﺼﺎﺩﻣﺎﺕ ﻭﻋﺪﺩ ﻣﺤﺎﻭﻻﺕ ﺍﻹﺭﺳﺎﻝ ﺍﻟﻼﺯﻣﺔ ﻟﺘﺴﻠﻴﻢ ﺍﻟﺤﺰﻣﺔ (ETX)، ﺇﺿﺎﻓﺔ ﺇﻟﻰ ﺳﺮﻋﺔ ﺃﻛﺒﺮ ﻓﻲ ﺍﺳﺘﻘﺮﺍﺭ ﻋﻤﻠﻴﺔ ﺍﻟﺘﻌﻠّﻢ ﻣﻘﺎﺭﻧﺔ ﺑﺎﻟﻄﺮﻕ ﺍﻷﺧﺮﻯ. ﺗﺆﻛﺪ ﻫﺬﻩ ﺍﻟﻨﺘﺎﺋﺞ ﻓﻌﺎﻟﻴﺔ ﺍﺳﺘﺨﺪﺍﻡ ﺧﻮﺍﺭﺯﻣﻴﺎﺕ ﺍﻟﺘﻌﻠّﻢ ﺍﻟﻌﻤﻴﻖ ﺑﺎﻟﺘﻌﺰﻳﺰ ﻓﻲ ﺗﺤﺴﻴﻦ ﺃﺩﺍء ﺑﺮﻭﺗﻮﻛﻮﻻﺕ ﺍﻟﻮﺻﻮﻝ ﻓﻲ ﺷﺒﻜﺎﺕ LoRa، ﻣﻤﺎ ﻳﺠﻌﻠﻬﺎ ﻧﻬﺠًﺎ ﻭﺍﻋﺪًﺍ ﻟﺘﻄﻮﻳﺮ ﺗﻄﺒﻴﻘﺎﺕ ﺇﻧﺘﺮﻧﺖ ﺍﻷﺷﻴﺎء ﻭﺍﺳﻌﺔ ﺍﻟﻨﻄﺎﻕ.

English Abstract

This thesis presents a Deep Q-Network (DQN)-based Medium Access Control (MAC) protocol designed to improve scalability and reliability in LoRa networks. LoRaWAN relies on an ALOHA-like multi-channel MAC mechanism, which suffers from increasing collisions and declining packet delivery ratio (PDR) as network density grows. To address these limitations, the transmission parameter selection problem is modeled as a Markov Decision Process (MDP), and a decentralized learning framework is developed in which each node autonomously selects its channel (CH), spreading factor (SF), and transmission power (TP). A custom LoRa simulator based on LoRaSim is implemented, incorporating realistic physical-layer effects such as packet airtime, multi-condition collision, and capture effect model. Extensive experiments were conducted for networks ranging from 100 to 3000 nodes. The proposed DQN-based MAC has been shown to achieve the highest PDR in all scenarios, reaching 99.07% for 100 nodes and 95.84% for 500 nodes. It also significantly reduced collisions, achieved lower Expected Transmission Count (ETX), and demonstrated faster convergence compared to both pure ALOHA and tabular Q-learning. These results confirm that Deep Reinforcement Learning (DRL) provides a scalable solution for LoRa medium access. The findings demonstrate the potential of intelligent learning-based MAC protocols to enhance large-scale IoT deployments.

Item Type:	Thesis (Masters)
Subjects:	Computer Systems Research Research > Information Technology
Department:	College of Computing and Mathematics > Computer Engineering
Thesis Advisor:	Louai Al-awami,
Thesis Committee Members:	Muhammad Mahmoud, Tarek Sheltami,
Depositing User:	NORAH SUMAYLI (g202306530)
Date Deposited:	22 Apr 2026 07:09
Last Modified:	22 Apr 2026 07:09
URI:	http://eprints.kfupm.edu.sa/id/eprint/144134