Page 197 - Kaleidoscope Academic Conference Proceedings 2021
P. 197
Connecting physical and virtual worlds
layer, and service layer. In this architecture, we have three [14]. So, with the help of these intelligent techniques such as
similar layers and two new layers, management layer and machine learning algorithms, we can have an adaptive and
business layer [2]. flexible network that can adjust to the changes in users’
needs. Machine learning algorithms are mainly divided to
2.4 Key technologies in network slicing three groups of supervised, unsupervised and reinforcement
learning [15].
There are a number of technologies which are used in • Supervised learning: This group of techniques are trained
network slicing. Here we elaborate on two of the most based on a set of data with desired input and output.
effective technologies in network slicing performance.
• Unsupervised learning: These techniques train
1. SDN (Software Defined Network) themselves with unlabeled data where the learning agent
A Software Defined Network (SDN) is a structure made for finds the pattern between inputs and outputs with no
adding flexibility to the architecture of networks by feedback.
abstracting the control phase of network devices [10]. The • Reinforcement learning: In these algorithms, the learning
SDN architecture consist of three layers: application plane, agent receives a reward from its current action from the
data plane and control plane. The controller has an ability to environment and it aims to maximize the reward based on
unify network management and other functions. SDN also the previous actions and related rewards.
has two interfaces between these layers. The Southbound
API is a connector that distributes information between the Machine learning algorithms are widely used in network
controller and data plane. The Northbound API is a functions and they are used for designing, deploying,
connector that transfers information between the controller operating and managing. They have the ability to detect
and application plane [11]. patterns that humans may not notice [16].
2. NFV (Network Function Virtualization) 4. AI BASED NETWORK SLICING FUNCTIONS
Network Function Virtualization (NFV) is separating the
network functions from hardware parts of the system. It also In this section different functions of network slicing are
implements these functions as independent software in presented from the standpoint of implementing AI and
Virtual Machines (VMs). In this way, we can efficiently use machine learning techniques. The overall view of the relation
data centers, being on a virtual infrastructure. The NFV between network slicing functions in the proposed model can
framework consists of three main components: [11] be seen in Figure 2. The mentioned functions are design,
deployment (resource provisioning and allocation),
• Virtual Network Functions (VNFs): software operation and management in terms of performance
implementations of network functions. These virtual management (admission control), fault management and
functions can be exploited in NFVI [12]. security. Table 1 presents a comparative view of different
• Network Function Virtualization Infrastructure (NFVI) is exploited machine learning techniques for network slicing
both the hardware and software parts of the NFV functions along with the algorithms and key features.
environment. NFV infrastructure is also responsible for
connecting the covered locations. 4.1 Design
• Network Functions Virtualization Management and
Architectural Framework (NFV-MANO Architectural The first step to have an efficient network is to plan and
Framework) is a set of all functional blocks used for data design. Network slices should be designed somehow to
transfer [10]. answer distinct service needs defined in 5G network. To
achieve this, a vast amount of data should be processed and
3. ARTIFICIAL INTELLIGENCE IN 5G analyzed. This data includes user needs and requirements,
NETWORK working environment and service goals [16].
For designing, deploying, and managing a network slice, 4.2 Deployment (resource provisioning and allocation)
huge amounts of data must be analyzed. Processing this huge
load of data is hard for any human being, that’s where In the next step after designing an efficient network, the
artificial intelligence comes to the picture. Intelligent amount of computing and network resources should be set
algorithms have the ability to automate network slicing by VNO (Virtual Network Operator) to support the service
operations and they can perform effectively for design, needs. Network resource provisioning and allocation is to
deployment, operation and management and configuring allocate the available resources of the network to the valid
system parameters [13]. slices efficiently. It should also adapt the designed resource
allocation to the dynamic behavior of network slices,
Machine learning has the ability to analyze huge amounts of keeping an eye on the capacity to avoid capacity outage.
data and learn the system behavior and predict future events After the admission control and accepting number of
in a short period of time. Intelligent techniques address the requests for slices, sufficient resources should be allocated
challenges faced in network automation. These algorithms to each slice. One of the major tasks in network slice
extract useful information from a huge amount of complex management is adaptive resource allocation to the admitted
data and choose the best decisions based on this information slices this step. In this regard, network operators should
determine the resources that should be dedicated to each
slice in order to use the
– 135 –