Scope of machine learning techniques for 5G wireless communication

 5G Introduction: --

Radio access network technology for 5G (5^th generation of wireless networks) is coined as 5G NR (new radio). It was introduced within the release 15 version of the 3GPP specifications.

5G 3GPP specifications are based on following three key use cases.

enhanced Mobile Broadband (eMBB)

Main goals:

1.       10-20 Gbps peak throughput (100 Mbps guaranteed)

2.       Support for high mobility (500 km/h speeds)

3.       Network energy saving by a factor of 100.

Ultra-Reliable and Low Latency Communications (URLLC)

Main goals:

1.       Ultra-responsive (5 msec end-to-end latency)

2.       Ultra-reliable and always available (99.9999%)

3.       Low to medium data rates (50 kbps - 10 Mbps)

4.       High speed mobility

Massive Machine Type Communications (mMTC)

Main goals:

1.       High density of devices (2x10^5 – 10^6 /square km)

2.       Long range  

3.       Low data rates (1 - 100 kbps)

4.       Ultra-low cost

5.       Long batter life (typically 10 years)

 

Key findings from Qualcomm study on economic impact of 5G

1.       5G will enable $13.1 trillion in global sales enablement in 2035.

2.       5G could enable 6.4% of public service (government) and 5.9% of agricultural output in 2035, driven by smart cities and smart agriculture deployments, respectively.

In addition, 5G technology has been at the heart of ongoing digital transformation across industries, with little impact of the Corona pandemic that has grappled the whole world.

The enormous challenge created by next generation 5G type wireless networks is that of assisting the radio in intelligent adaptive learning and decision making, so that the above specified diverse use cases can be met.

Machine learning is one of the most promising artificial intelligence tools, conceived to support smart radio terminals and self-driving networks in the making.

ML has a crucial role to play in such 5G and post 5G wireless networks, because it is capable of modelling systems that cannot be represented precisely by a mathematical equation but could be learned by model through massive amounts of data. It would be also used for management, optimization, and security of these networks as it would be impossible for network operators to manage these complex networks manually, which would consist of many antennas at the network and at the mobile end.

Machine learning conceptual overview.

ML models are used to learn the characteristics of a system that cannot be presented by an explicit mathematical model. These models are used in tasks such as classification, regression, and the interactions of an intelligent agent with an environment. Once a model learns the characteristics of a system (this is known as a trained model), it can efficiently perform the task using some basic arithmetical calculations.

ML comprises of four paradigms, known as:

a) Supervised learning: where the model is learned by presenting input samples and their known associated outputs.

b) Unsupervised learning, in which there are no output labels, and the model learns to classify samples of the input.

c) Reinforcement learning, where an agent interacts with an environment and learns to map any input to an action.

d) Deep learning, Deep learning methods based on artificial neural networks (ANNs) have recently been able to solve many learning problems, including famous AlphaGo computer program that plays the board game Go, developed by Google DeepMind, which beat the world champion, Lee Sedol a few years back. The rise of the deep learning paradigm has mainly been fueled by the availability of sufficient computational power and access to large datasets. It further comprises of the following.

1.       Multilayer perceptron (MLPs) are the basic models that are generally used in many learning tasks.

2.       Convolutional neural networks (CNN), which use convolution operation to reduce the input size, are often used in image recognition tasks.

3.       For learning tasks which require sequential models, recurrent neural networks (RNN) are most suitable.

4.       Autoencoder-based deep learning models are used for dimension reduction, and

5.       Generative adversarial networks (GANs) are used to generate samples like the available dataset.

 

Overview of the main machine paradigm application areas for 5G wireless systems is envisioned as follows: --

Supervised machine learning algorithms

Supervised machine learning	Learning techniques	Key characteristics	Application areas in 5G
	Regression models	Estimate the variables’ relationships. Linear and logistics regression	Energy learning
	Support vector machines	Non-linear mapping to high dimension Separate hyperplane classification	MIMO channel learning
	Bayesian learning	A posteriori distribution calculation	Massive MIMO learning Cognitive spectrum learning

  

Unsupervised machine learning algorithms

Un-Supervised machine learning	Learning techniques	Key characteristics	Application areas in 5G
	K-means clustering	K partition clustering Iterative updating algorithm	Heterogeneous networks
	Principal Component Analysis	Orthogonal transformation	Smart grid

Reinforcement machine learning algorithms

Reinforcement machine learning	Learning techniques	Key characteristics	Application areas in 5G
	MDP (Markov decision processes)/POMDP (Partially observable MDP)	Bellman equation maximization Value iteration algorithm	Energy harvesting
	Q-learning	Unknown system transition model Q-function maximization	Femto and small cells