Development of deep reinforcement learning based resource allocation techniques in cloud radio access network
Nextgeneration networks are envisioned to support management to maximize the user s ’ dynamic and agile network quality of service (QoS). Cloud radio access network (CRAN) emerges as a promising candidate since the limited network resources can be virtualized and shared among distributed remote radi...
Saved in:
| Main Author: | |
|---|---|
| Format: | Final Year Project / Dissertation / Thesis |
| Published: |
2022
|
| Subjects: | |
| Online Access: | http://eprints.utar.edu.my/4606/1/Amjad_Iqbal.pdf http://eprints.utar.edu.my/4606/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Nextgeneration networks are envisioned to support management to maximize the user s ’ dynamic and agile network quality of service (QoS). Cloud radio access network (CRAN) emerges as a promising candidate since the limited network resources can be virtualized and shared among distributed remote radio heads (RRHs). Conventional approaches formulate resource allocation as an optimization problem and solve it with instantaneous environment knowledge without considering the consequences of actions. A step towards long network performance optimization is theterm use of deep reinforcement learning (DRL), which can learn the best policy via interaction with the environment. This thesis proposes three DRLbased resource allocation algorithms that optimize the CRAN performance in terms of energy efficiency (EE), spec tral efficiency (SE), and total power consumption. The first proposed algorithm aims to optimize the EE by controlling the on/off status of RRH via a deep Q network (DQN) and subsequently solving a power optimization problem. To capture the spatiotemporal channel state information (CSI), the second proposed algorithm adopts machine learning techniques with anchor graph hashing to extract generalized features before feeding them into the DQN. The goal here is to optimize the longterm tradeoff between EE and SE. In the last proposed scheme, additional EE savings are facilitated by designing and integrating a convolution al neural network (CNN), which can better learn the feature of environment states. Simulation results show that all proposed DRL algorithms outperform 2025% compared to existing techniques while achieving faster convergence. All performance benchmarking was carried out based on 100 testing episodes after properly training the DRL agent with 1000 episodes. |
|---|