Research Overview of Intelligent Wireless
Communication Technology
Reported by: Tiankui
Zhang
Date ：
2023.11.23
Page 1
Contents
Background
01 Introduction 02 Knowledge
Wireless Spectrum
03 Transmission
04 Management
Resource Network
05 Configuration
06 Access
Network and Summary and Future
07 System Optimization 08 Research Prospects
Page 2
01 Introduction
Page 3
Introduction
Deep
Reinforcement
Deep Learning
Learning
(DL)
(DRL)
Difficulties solved by
in the development of
Machine Learning (ML)
wireless communication
technology
developing······
Federated Learning
(FL)
Page 4
Background
02 Knowledge
Deep Learning Overview
Deep Reinforcement Learning Overview
Federated Learning Overview
Page 5
Deep Learning Overview
asic idea:
use DNN(deep neural network) to fit the relationship between source data and target data
t3
neuron
h
ig
ht 4
we
source target wei
g weig
ht 1 wei
data ght
data ht 2 sum
5
g activation
wei
we
function
i
gh
t6
a neuron of
input hidden hidden output input hidden layer 1 hidden
input
layer layer 1 layer n layer layer layer 2
layer
The Structure of DNN Internal Structure of a Neuron
in Hidden Layer 1
in DNN
Page 6
Deep Reinforcement Learning Overview
Page 7
Federated Learning Overview
FL is proposed for multi-agent collaborative training of DNN without sharing training data.
Page 8
Wireless
03 Transmission
Channel Estimation
Channel Prediction and Feedback
Signal Detection
Page 9
Channel Estimation
Step 1:Send the signal
Step 2:Estimate the CSI
Examples:
1 、 MMSE+CNN
2 、 LS+DNN
3 、 CFO+TO+DNN Page 10
Channel Prediction and Feedback
Q: Why channel prediction is necessary?
A: In a massive MIMO system using frequency division multiplexing, it is difficult to obtain
CSI to precode the signal for parallel transmission of multiple information streams.
Solution 1:
Use deep learning to mine the mapping relationship between uplink and downlink CSI.
By reducing the number of neurons in the middle layer, deep learning is forced to extract the
key common information common to uplink and downlink CSI, thereby obtaining better
prediction accuracy.
Solution 2:
Reduce feedback overhead.
Utilize an autoencoder to compress CSI, reducing backhaul link load by transmitting only the
CSI compressed value.
Page 11
Signal Detection
Challenges in Traditional Signal Deep Learning as a Solution
Detection
Limited by incomplete prior knowledge Efficiently solves decision problems in
and signal complexity signal detection
Advancements with DNN Practical Application
Improvements in OAMP, CGD, and ML DNN for direct signal-to-data mapping,
algorithms for lower computational optimal design of pilots, and channel
overhead estimation
Deep Neural Networks (DNN) revolutionize signal detection in wireless transmission by simulating encoding-decoding
processes and abstracting channel responses, leading to significantly improved performance even under constraints like
fewer pilots or missing cyclic prefixes. Page 12
Spectrum
04 Management
Spectrum Sensing
Channel Access
Modulation Identification
Page 13
Spectrum Sensing
Cognitive radio:
alleviate surge traffic and limited spectrum,
and improve spectrum utilization.
Secondary users (SU) are allowed to use the
spectrum when the primary user (PU) who
owns the spectrum is idle.
received signal
Core of Spectrum Sensing
Id e a 1 Id e a 2
how to accurately determine the primary user
Identification
More ideas······ covariance matrix
Nature of Spectrum Sensing
a problem of determining whether the primary
user exists test statistic
Deep Learning Page 14
Judgement
Channel Access
Channel Access: a decision-making problem about channels and access times
Use DRL to directly analyze spectrum
Sharing of unlicensed spectrum between LTE information and intelligently adjust
cellular mobile communication system and Wi-Fi time slot allocation for LTE and Wi-Fi
by predicting Wi-Fi traffic information
Dynamic channel access Distributed channel access solution based on DRL
CA problem in multi-channel and
multi-user situations
Competition - Dual Deep Q-Learning Algorithm
Complex electromagnetic DQL-based anti-interference channel selection
environment with malicious jammers algorithm
Scenario Strategy Page 15
Modulation Identification
Purpose: Application examples:
Monitor the illegal occupation of spectrum, 1. CNN+ Signal spectrum waterfall chart
detect illegal signals and locate the signal
source 2. CNN+ Signal constellation diagram
3. Double-layer CNN architecture
4. CNN+ LSTM
5. CNN+ LSTM+ Attention mechanism
Algorithm Scheme
solving based on Based on artificial
old likelihood function selection of features
MI
Use deep learning to
using deep learning to
automatically extract
new directly make
features and modulate
decisions using DNN
the signal Page 16
Resource
05 Configuration
Power Distribution
Computing Resource Configuration
Page 17
Power Distribution
Power Distribution: An Effective Means for Interference Management
Single-user Scenario Multi-user Cellular
Networks
Difficulties:
1: Non-convex problem
The DQL algorithm can be
2: Exponentially increasing
selected to implement a DRL-
CSI data
based power distribution
solution.
Solution strategy:
Distributed DQL algorithm +
centralized training
framework + decomposing
the global optimization goal
Other strategy:
Combining power distribution and channel access to jointly optimize transmit power and
Page 18
channels, can be applied to fields such as the Internet of Vehicles.
Computing Resource Configuration
Why?
In fact, the computing resources of mobile edge computing are also limited and need to be flexibly
configured according to user needs.
Requirement Requirement
Analyze and predict Offload computing tasks
users’ preferences for to mobile edge
various services computing
Solution Strategy Solution Strategy
Design distributed deep Improve DQL into a
learning algorithms federated-deep Q
based on the idea of learning algorithm based
federated learning on federated learning
Page 19
06 Network Access
Page 20
Network Access
Why?
In order to continuously ensure access performance, users need to dynamically select and switch
connected base stations from a global and long-term perspective.
1 、 Add an offset value to the original handover threshold
Base station access issues in UDN 2 、 Use A3C algorithm in DRL to intelligently select
the offset value.
Multi-layer cellular systems with Using the distributed competition-double deep Q
NA
varying transmit powers learning algorithm
Using the DQL algorithm allows users to learn the
Access issues for LTE and Wi-Fi
optimal RAT switching control strategy
Scenario Strategy Page 21
Network and System
07 Optimization
Page 22
Network and System Optimization
The introduction of artificial intelligence is expected to
achieve self-optimization of wireless communication
systems and reduce costs while improving system
performance.
System fault repair
User Traffic Demand is
unevenly distributed in DQL algorithm
Time and Space
Detect abnormal devices in the IoT
The joint use of DL algorithm and FL
two methods
IRS DQL DQL+UAV
Page 23
Summary and Future
08 Research Prospects
Deep Learning
Deep Reinforcement Learning
Federated Learning
Page 24
Deep Learning
Future
Background Research point
Small amount of data DNN Structure and
Problem
and noise interference Data Preprocessing
Promote mature technologies Manually tuning the DNN structure based on data
Solution in fields such as computer characteristics and meta-learning are also possible ways to
vision solve this problem.
Compared with existing manual selection DNN structure
Extremely low
Meaning strategies, the above method can effectively improve system
computational overhead performance.
Page 25
Deep Reinforcement Learning
Future
Background Research point
Unstable performance during Build a distributed multi-agent
Problem convergence; Hardware limitations decision-making solution
Combine existing Game theory and stochastic game theory can provide a
Solution knowledge to speed up; theoretical framework for solving the multi-agent non-
Use MEC to share tasks stationarity problem of DRL.
DRL training speeds up Future wireless communication systems will be built from
Meaning
and performance numerous smart devices.
improves.
Page 26
Federated Learning
Future
Background
Problem Huge communication overhead; Possibly different data sources; Poor ability to resist attacks
Quantify and reduce communication overhead;
Solution Study federated learning for non-IID data sets;
Adopt strong security measures
Meaning Reduce communication overhead, improve DNN accuracy, and effectively ensure system
security.
Page 27
THANK YOU ！
Reported by: Tiankui
Zhang
Date ：
2023.11.23
Page 28
