LT
L.M.N. Tabaksblat
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
2 records found
1
**Abstract**
Cell-Free Networks (CFNs) have emerged as a promising technology to improve the spatial fairness of mobile networks by combining with a functional split to create a network architecture capable of serving any User Equipment (UE) with multiple spatially distributed Multiple-Input Multiple-Output (MIMO) antennas. While the distributed architecture of CFNs offers significant performance benefits, it also introduces challenges related to energy efficiency, as networks are typically designed to satisfy peak-hour traffic demands and therefore consume more power than necessary during periods of low traffic. Sleep Mode Management (SMM), which selectively deactivates unnecessary Access Points (APs), is a promising approach to reducing this energy consumption. However, existing Machine Learning (ML)-based solutions often rely on simplified observations, use only one sleep mode, and do not explicitly account for coverage constraints and realistic sleep mode transition times. This thesis investigates whether ML, and in particular Deep Reinforcement Learning (DRL), can be used to address these limitations.
For this reason, a scalable Multi-Agent Proximal Policy Optimisation (MAPPO)-based SMM algorithm is proposed. The MAPPO algorithm operates on a short timescale and supports three AP sleep modes: active mode, light sleep mode, and deep sleep mode. It takes into account transition delays between sleep modes and uses only realistically obtainable observations regarding the network based on what each AP can observe in its neighbourhood. During training, the algorithm learns to minimize the average network power consumption subject to coverage and Quality of Service (QoS) constraints.
The proposed algorithm is evaluated using a variety of realistic scenarios located in the city centre of Amsterdam. The scenarios are of various sizes and use a realistic network deployment, as well as a realistic basis for the spatial traffic distribution and daily traffic fluctuations. When comparing the proposed algorithm with two state-of-the-art heuristic algorithms, it was found that the algorithm could not outperform the heuristic baselines in terms of energy savings. Additionally, it was found that the 10th UE throughput percentile is not a suitable metric to use as direct feedback for a DRL algorithm because it is an unreliable metric on short timescales. More research therefore needs to be performed to investigate how the DRL algorithm can be improved.
...
Cell-Free Networks (CFNs) have emerged as a promising technology to improve the spatial fairness of mobile networks by combining with a functional split to create a network architecture capable of serving any User Equipment (UE) with multiple spatially distributed Multiple-Input Multiple-Output (MIMO) antennas. While the distributed architecture of CFNs offers significant performance benefits, it also introduces challenges related to energy efficiency, as networks are typically designed to satisfy peak-hour traffic demands and therefore consume more power than necessary during periods of low traffic. Sleep Mode Management (SMM), which selectively deactivates unnecessary Access Points (APs), is a promising approach to reducing this energy consumption. However, existing Machine Learning (ML)-based solutions often rely on simplified observations, use only one sleep mode, and do not explicitly account for coverage constraints and realistic sleep mode transition times. This thesis investigates whether ML, and in particular Deep Reinforcement Learning (DRL), can be used to address these limitations.
For this reason, a scalable Multi-Agent Proximal Policy Optimisation (MAPPO)-based SMM algorithm is proposed. The MAPPO algorithm operates on a short timescale and supports three AP sleep modes: active mode, light sleep mode, and deep sleep mode. It takes into account transition delays between sleep modes and uses only realistically obtainable observations regarding the network based on what each AP can observe in its neighbourhood. During training, the algorithm learns to minimize the average network power consumption subject to coverage and Quality of Service (QoS) constraints.
The proposed algorithm is evaluated using a variety of realistic scenarios located in the city centre of Amsterdam. The scenarios are of various sizes and use a realistic network deployment, as well as a realistic basis for the spatial traffic distribution and daily traffic fluctuations. When comparing the proposed algorithm with two state-of-the-art heuristic algorithms, it was found that the algorithm could not outperform the heuristic baselines in terms of energy savings. Additionally, it was found that the 10th UE throughput percentile is not a suitable metric to use as direct feedback for a DRL algorithm because it is an unreliable metric on short timescales. More research therefore needs to be performed to investigate how the DRL algorithm can be improved.
...
**Abstract**
Cell-Free Networks (CFNs) have emerged as a promising technology to improve the spatial fairness of mobile networks by combining with a functional split to create a network architecture capable of serving any User Equipment (UE) with multiple spatially distributed Multiple-Input Multiple-Output (MIMO) antennas. While the distributed architecture of CFNs offers significant performance benefits, it also introduces challenges related to energy efficiency, as networks are typically designed to satisfy peak-hour traffic demands and therefore consume more power than necessary during periods of low traffic. Sleep Mode Management (SMM), which selectively deactivates unnecessary Access Points (APs), is a promising approach to reducing this energy consumption. However, existing Machine Learning (ML)-based solutions often rely on simplified observations, use only one sleep mode, and do not explicitly account for coverage constraints and realistic sleep mode transition times. This thesis investigates whether ML, and in particular Deep Reinforcement Learning (DRL), can be used to address these limitations.
For this reason, a scalable Multi-Agent Proximal Policy Optimisation (MAPPO)-based SMM algorithm is proposed. The MAPPO algorithm operates on a short timescale and supports three AP sleep modes: active mode, light sleep mode, and deep sleep mode. It takes into account transition delays between sleep modes and uses only realistically obtainable observations regarding the network based on what each AP can observe in its neighbourhood. During training, the algorithm learns to minimize the average network power consumption subject to coverage and Quality of Service (QoS) constraints.
The proposed algorithm is evaluated using a variety of realistic scenarios located in the city centre of Amsterdam. The scenarios are of various sizes and use a realistic network deployment, as well as a realistic basis for the spatial traffic distribution and daily traffic fluctuations. When comparing the proposed algorithm with two state-of-the-art heuristic algorithms, it was found that the algorithm could not outperform the heuristic baselines in terms of energy savings. Additionally, it was found that the 10th UE throughput percentile is not a suitable metric to use as direct feedback for a DRL algorithm because it is an unreliable metric on short timescales. More research therefore needs to be performed to investigate how the DRL algorithm can be improved.
Cell-Free Networks (CFNs) have emerged as a promising technology to improve the spatial fairness of mobile networks by combining with a functional split to create a network architecture capable of serving any User Equipment (UE) with multiple spatially distributed Multiple-Input Multiple-Output (MIMO) antennas. While the distributed architecture of CFNs offers significant performance benefits, it also introduces challenges related to energy efficiency, as networks are typically designed to satisfy peak-hour traffic demands and therefore consume more power than necessary during periods of low traffic. Sleep Mode Management (SMM), which selectively deactivates unnecessary Access Points (APs), is a promising approach to reducing this energy consumption. However, existing Machine Learning (ML)-based solutions often rely on simplified observations, use only one sleep mode, and do not explicitly account for coverage constraints and realistic sleep mode transition times. This thesis investigates whether ML, and in particular Deep Reinforcement Learning (DRL), can be used to address these limitations.
For this reason, a scalable Multi-Agent Proximal Policy Optimisation (MAPPO)-based SMM algorithm is proposed. The MAPPO algorithm operates on a short timescale and supports three AP sleep modes: active mode, light sleep mode, and deep sleep mode. It takes into account transition delays between sleep modes and uses only realistically obtainable observations regarding the network based on what each AP can observe in its neighbourhood. During training, the algorithm learns to minimize the average network power consumption subject to coverage and Quality of Service (QoS) constraints.
The proposed algorithm is evaluated using a variety of realistic scenarios located in the city centre of Amsterdam. The scenarios are of various sizes and use a realistic network deployment, as well as a realistic basis for the spatial traffic distribution and daily traffic fluctuations. When comparing the proposed algorithm with two state-of-the-art heuristic algorithms, it was found that the algorithm could not outperform the heuristic baselines in terms of energy savings. Additionally, it was found that the 10th UE throughput percentile is not a suitable metric to use as direct feedback for a DRL algorithm because it is an unreliable metric on short timescales. More research therefore needs to be performed to investigate how the DRL algorithm can be improved.
Bachelor thesis
(2023)
-
L. Alonso Antona, S. Aurori, M. Beenders, E.G. Chen, C.F.M. Kendall, D.J.D. Norbart, T. Odijk, M.H. Rusch, L.M.N. Tabaksblat, S. Yorucu, I. Akay, A.O. Başkaya, P. Piron
The goal of project Altus is to do an in-situ investigation of Polar Mesospheric Clouds (PMCs). These clouds form around an altitude of 84 km, and only for 60 to 80 days per year, during the summer. Normally, these clouds only form in the polar regions, from around 50◦ latitude north and south. Recently, however, PMCs have been observed as low as 40◦ north. There are theories linking this change in location, and other unexpected behaviours of PMCs, to climate change. However, further research is still required to confirm these theories. As these changes are happening at a slow rate, a database of PMC measurements would be extremely beneficial to track indicator values over time. Project Altus sets out to bridge this knowledge gap by taking regular measurements of PMCs over an extended period of time.
...
The goal of project Altus is to do an in-situ investigation of Polar Mesospheric Clouds (PMCs). These clouds form around an altitude of 84 km, and only for 60 to 80 days per year, during the summer. Normally, these clouds only form in the polar regions, from around 50◦ latitude north and south. Recently, however, PMCs have been observed as low as 40◦ north. There are theories linking this change in location, and other unexpected behaviours of PMCs, to climate change. However, further research is still required to confirm these theories. As these changes are happening at a slow rate, a database of PMC measurements would be extremely beneficial to track indicator values over time. Project Altus sets out to bridge this knowledge gap by taking regular measurements of PMCs over an extended period of time.