Page 199 - Kaleidoscope Academic Conference Proceedings 2021

Connecting physical and virtual worlds

Table 1 – AI-based network slicing functions algorithms

Function Algorithm Key Features Reference
F.1. Design Support vector machine, Gradient Analysis of user requirements as well [16]
boosting decision tree, Spectra clustering as environment factors
Q-learning
Hybrid learning algorithm, NN glowworm swarm-based DHOA [21]
F.2. Deployment (Resource Supervised learning: DEEPCOG, a cost prediction [19]
provisioning & allocation) 3D-CNN algorithm for reallocation decisions
Lasso regression (provisioning) Allocating resources through a [22]
RNN, DNN, RL, Sparse regression process consisting 4 functional blocks
Reinforcement learning: RL, DQL Embrace the deep relation between [23]
user requests and resource allocation
Reinforcement learning: RL [24]
Outperforms the three deterministic
heuristics
Reinforcement learning: RL Model free and scalable [25]
Reinforcement learning: RL, QL Model free and robust [26]

Reinforcement learning: Ape-X, a DRL A one-slice-by-one-agent basis [27]
method resource management
discrete normalized advantage functions Exploiting a deterministic policy [28]
(DNAF) + DQL gradient descent (DPGD)
Reinforcement learning: QL Maximizing the profit of tenants [29]
while considering the QoS
Reinforcement learning: DRL Improves latency performance [30]
Reinforcement learning: RL A two-stage network slice resource [31]
allocation algorithm
Reinforcement learning: DRL Dynamic resource allocation [32]
Reinforcement Learning: DRL Generative adversarial network based [33]
deep distributional Q network
Reinforcement learning: RL Collaborative relationship between [34]
node mapping and link mapping
Reinforcement Learning: DRL A dynamic two-tier slice allocation [35]
scheme
Reinforcement learning: exponential RL An auction-based resource allocation [36]
Reinforcement learning: QL Maximizing network utility [37]

F.3.1. Operation & management: Reinforcement learning: QL Inter-slice control with enablers [38]
Performance management
Reinforcement Learning: DRL adaptive to system changes, useful for [17]
large networks
Reinforcement learning: RL Aims for network lasting profitability [25]
Reinforcement learning: Value-based RL Maximizing mobile network [26]
operator’s revenue
Reinforcement learning: RL Maximizing network efficiency [24]
F.3.2. Operation & management: Logistic regression, Bayesian networks, Analysis of system behavior and [16]
Fault management Principal component analysis, capable of locating the faults
independent component analysis
F.3.3. Operation & management: Supervised Learning: DNN Finding the threats and isolating the [16]
Security harmed areas
Machine learning real-time and offline attack detection [39]
Supervised Learning: DL a self-adaptive system for anomaly [40]
detection
Supervised Learning: DL Designing a quarantine slice towards [4]
5g secure slicing

– 137 –