Due to high advances in production of technological devices and a sharp increase in the Internet usage in contemporary decades, the Internet and other small network connections have become a basic need in life as nearly almost all the daily consumed information has to be gathered online from news, movies, books etc. However, as the number of the Internet users escalate, the momentum of the Internet decelerates as a result of data traffic congestion within the networks. To regulate this issue, a couple of predestined algorithms such as TCP-Reno and TCP-Vegas are employed to counterbalance the sending and receiving rates of data packets within the network to eliminate delays and data loss but come with limitations such rigidity in dynamically changing congestion environment which has led to low efficiency. This Tez introduces the comparison in efficiency of deep reinforcement learning (DRL) algorithms developed to diminish congestion in computer networks. There are many DRL algorithms and simulators that can be deployed in such scenarios but this work focuses on deep deterministic policy gradient (DDPG), twin delayed deep deterministic (TD3), and proximal policy optimization (PPO) and later weighed with the existing non-DRL ones. The results showed that DDPG outperformed the other two by a considerable margin, succeeded by TD3 and finally by PPO. Even though one of the non-DRL algorithms scored higher than two of the DRL algorithms, DDPG still showed remarkable results.
Eser Adı (dc.title) | Comparison Of Deep Reinforcement Learning Algorithms In Computer Network Traffic Congestion Control |
Yayın Türü (dc.type) | Tez |
Yazar/lar (dc.contributor.author) | MAYANJA, Abubakar |
Atıf Dizini (dc.source.database) | Diğer |
Konu Başlıkları (dc.subject) | Deep Reinforcement Learning |
Konu Başlıkları (dc.subject) | Deep Deterministic Policy Gradient |
Konu Başlıkları (dc.subject) | Twin Delayed Deep Deterministic |
Konu Başlıkları (dc.subject) | Proximal Policy Optimization |
Yayıncı (dc.publisher) | KTO Karatay Üniversitesi |
Tez Danışmanı (dc.contributor.advisor) | ALTUN, Hüseyin Oktay |
Yayın Tarihi (dc.date.issued) | 2021 |
Kayıt Giriş Tarihi (dc.date.accessioned) | 2022-01-04T07:39:01Z |
Açık Erişim tarihi (dc.date.available) | 2022-01-04T07:39:01Z |
Tez Türü (dc.type.thesis) | Yüksek Lisans |
Özet (dc.description.abstract) | Due to high advances in production of technological devices and a sharp increase in the Internet usage in contemporary decades, the Internet and other small network connections have become a basic need in life as nearly almost all the daily consumed information has to be gathered online from news, movies, books etc. However, as the number of the Internet users escalate, the momentum of the Internet decelerates as a result of data traffic congestion within the networks. To regulate this issue, a couple of predestined algorithms such as TCP-Reno and TCP-Vegas are employed to counterbalance the sending and receiving rates of data packets within the network to eliminate delays and data loss but come with limitations such rigidity in dynamically changing congestion environment which has led to low efficiency. This Tez introduces the comparison in efficiency of deep reinforcement learning (DRL) algorithms developed to diminish congestion in computer networks. There are many DRL algorithms and simulators that can be deployed in such scenarios but this work focuses on deep deterministic policy gradient (DDPG), twin delayed deep deterministic (TD3), and proximal policy optimization (PPO) and later weighed with the existing non-DRL ones. The results showed that DDPG outperformed the other two by a considerable margin, succeeded by TD3 and finally by PPO. Even though one of the non-DRL algorithms scored higher than two of the DRL algorithms, DDPG still showed remarkable results. |
Yayın Dili (dc.language.iso) | en |
Tek Biçim Adres (dc.identifier.uri) | http://hdl.handle.net/20.500.12498/5195 |