DRL For WSN Book

SPRINGER BRIEFS IN
ELEC TRIC AL AND COMPUTER ENGINEERING
F. Richard Yu
Ying He
Deep
Reinforcement
Learning
for Wireless
Networks
123
SpringerBriefs in Electrical and Computer
Engineering
Series editors
Woon-Seng Gan, School of Electrical and Electronic Engineering, Nanyang
Technological University, Singapore, Singapore
C.-C. Jay Kuo, University of Southern California, Los Angeles, CA, USA
Thomas Fang Zheng, Research Institute of Information Technology, Tsinghua
University, Beijing, China
Mauro Barni, Department of Information Engineering and Mathematics, University
of Siena, Siena, Italy
SpringerBriefs present concise summaries of cutting-edge research and practical
applications across a wide spectrum of fields. Featuring compact volumes of 50
to 125 pages, the series covers a range of content from professional to academic.
Typical topics might include: timely report of state-of-the art analytical techniques,
a bridge between new research results, as published in journal articles, and a
contextual literature review, a snapshot of a hot or emerging topic, an in-depth case
study or clinical example and a presentation of core concepts that students must
understand in order to make independent contributions.
More information about this series at http://www.springer.com/series/10059

F. Richard Yu • Ying He
Deep Reinforcement
Learning for Wireless
Networks
123
F. Richard Yu Ying He
Carleton University Carleton University
Ottawa, ON, Canada Ottawa, ON, Canada
ISSN 2191-8112 ISSN 2191-8120 (electronic)

SpringerBriefs in Electrical and Computer Engineering
ISBN 978-3-030-10545-7 ISBN 978-3-030-10546-4 (eBook)
https://doi.org/10.1007/978-3-030-10546-4
Library of Congress Control Number: 2018967964
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2019

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Switzerland AG.
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
A Brief Journey Through “Deep Reinforcement Learning

for Wireless Networks”
There is a phenomenal burst of research activities in machine learning and wireless

systems. Machine learning evolved from a collection of powerful techniques in AI
areas and has been extensively used in data mining, which allows the system to learn
the useful structural patterns and models from training data. Reinforcement learning
is an important branch of machine learning, where an agent learns to take actions
that would yield the most reward by interacting with the environment. The main
advantage of reinforcement learning is that it works well without prior knowledge
of an exact mathematical model of the environment. However, the traditional
reinforcement learning approach has some shortcomings, such as low convergence
rate to the optimal behavior policy and its inability to solve problems with high-
dimensional state space and action space. These shortcomings can be addressed
by deep reinforcement learning. The key idea of deep reinforcement learning is to
approximate the value function by leveraging the powerful function approximation
property of deep neural networks. After training the deep neural networks, given a
state-action pair as input, deep reinforcement learning is able to estimate the long-
term reward. The estimation result can guide the agent to choose the best action.
Deep reinforcement learning has been successfully used to solve many practical
problems. For example, Google DeepMind adopts this method on several artificial
intelligent projects with big data (e.g., AlphaGo) and gets quite good results. In
this book, we utilize deep reinforcement learning approach to wireless networks to
improve the system performance.
As the first chapter of this book, Chap. 1, we introduce machine learning
algorithms, which are classified into four categories: supervised, unsupervised,
semi-supervised, and reinforcement learning. In this chapter, widely used machine
learning algorithms are introduced. Each algorithm is briefly explained with some
examples.
v
vi Preface
Chapter 2 focuses on reinforcement learning and deep reinforcement learning.

A brief review of reinforcement learning and Q-learning is first described. Then,
recent advances of deep Q-network are presented, and double deep Q-network and
dueling deep Q-network that go beyond deep Q-network are also given.
Chapter 3 presents a deep reinforcement learning approach in cache-enabled
opportunistic interference alignment wireless networks. Most existing works on
cache-enabled interference alignment wireless networks assume that the channel
is invariant, which is unrealistic considering the time-varying nature of practical
wireless environments. In this chapter, we consider realistic time-varying channels.
The complexity of the system is very high when we consider realistic time-varying
channels. We use Google TensorFlow to implement deep reinforcement learning in
this chapter to obtain the optimal interference alignment user selection policy in
cache-enabled opportunistic interference alignment wireless networks. Simulation
results are presented to show that the performance of cache-enabled opportunistic
interference alignment networks in terms of the network’s sum rate and energy
efficiency can be significantly improved by using the proposed approach.
Chapter 4 considers trust-based mobile social networks with mobile edge
computing, in-network caching, and device-to-device communications. An opti-
mization problem is formulated to maximize the network operator’s utility with
comprehensive considerations of trust values, computation capabilities, wireless
channel qualities, and the cache status of all the available nodes. We apply a deep
reinforcement learning approach to automatically make a decision for optimally
allocating the network resources. The decision is made purely through observing
the network’s states, rather than any handcrafted or explicit control rules, which
makes it adaptive to variable network conditions. Simulation results with different
network parameters are presented to show the effectiveness of the proposed scheme.
The book will be useful to both researchers and practitioners in this area. The
readers will find the rich set of references in each chapter particularly valuable.
Ottawa, ON, Canada F. Richard Yu

October 2018 Ying He
Contents
1 Introduction to Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Supervised Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 k-Nearest Neighbor (k-NN) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.2 Decision Tree (DT) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.3 Random Forest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.4 Neural Network (NN) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.1.5 Support Vector Machine (SVM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1.6 Bayes’ Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1.7 Hidden Markov Models (HMM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2 Unsupervised Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.1 k-Means . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.2 Self-Organizing Map (SOM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3 Semi-supervised Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2 Reinforcement Learning and Deep Reinforcement Learning . . . . . . . . . . . . 15
2.1 Reinforcement Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Deep Q-Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3 Beyond Deep Q-Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3.1 Double DQN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3.2 Dueling DQN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3 Deep Reinforcement Learning for Interference Alignment
Wireless Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2.1 Interference Alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2.2 Cache-Equipped Transmitters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.3 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.3.1 Time-Varying IA-Based Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.3.2 Formulation of the Network’s Optimization Problem . . . . . . . . . 26
vii
viii Contents
3.4 Simulation Results and Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3.4.1 TensorFlow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.4.2 Simulation Settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.4.3 Simulation Results and Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.5 Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4 Deep Reinforcement Learning for Mobile Social Networks . . . . . . . . . . . . . . 45
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.1.1 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.1.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.2.1 System Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.2.2 Network Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.2.3 Communication Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.2.4 Cache Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.2.5 Computing Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.3 Social Trust Scheme with Uncertain Reasoning . . . . . . . . . . . . . . . . . . . . . . . . 54
4.3.1 Trust Evaluation from Direct Observations. . . . . . . . . . . . . . . . . . . . . 55
4.3.2 Trust Evaluation from Indirect Observations . . . . . . . . . . . . . . . . . . . 56
4.4 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.4.1 System State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.4.2 System Action . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.4.3 Reward Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.5 Simulation Results and Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.5.1 Simulation Settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.5.2 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.6 Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
Chapter 1
Introduction to Machine Learning
Abstract Machine learning is evolved from a collection of powerful techniques in

AI areas and has been extensively used in data mining, which allows the system to
learn the useful structural patterns and models from training data. Machine learning
algorithms can be basically classified into four categories: supervised, unsupervised,
semi-supervised and reinforcement learning. In this chapter, widely-used machine
learning algorithms are introduced. Each algorithm is briefly explained with some
examples.
A machine learning approach usually consists of two main phases: training phase
and decision making phase as illustrated in the Fig. 1.1. At the training phase,
machine learning methods are applied to learn the system model using the training
dataset. At the decision making phase, the system can obtain the estimated output
for each new input by using the trained model. Figure 1.2 shows supervised,
unsupervised, semi-supervised and reinforcement learning, which are described in
the following.
1.1 Supervised Learning
Supervised learning is a kind of labelling learning technique. Supervised learning

algorithms are given a labeled training dataset (i.e., inputs and known outputs) to
build the system model representing the learned relation between the input and
output. After training, when a new input is fed into the system, the trained model
can be used to get the expected output [1, 2]. In the following, we will give a detailed
representation of widely-used supervised learning algorithms, such as k-nearest
neighbor, decision tree, random forest, neural network, support vector machine,
Bayes’ theory, and hidden markov models.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2019 1

F. R. Yu, Y. He, Deep Reinforcement Learning for Wireless Networks, SpringerBriefs
in Electrical and Computer Engineering, https://doi.org/10.1007/978-3-030-10546-4_1
2 1 Introduction to Machine Learning
Training dataset
Training input set
Training output set
Training phase
Machine learning methods
Decision making phase

Estimated
output
Trained model
New input
Fig. 1.1 The general processing procedure of a machine learning approach
1.1.1 k-Nearest Neighbor (k-NN)
The k-NN is a supervised learning technique, where the classification of a data

sample is determined based on the k nearest neighbors of that unclassified sample.
The process of the k-NN algorithm is very simple: if the most of the k nearest
neighbors belong to a certain class, the unclassified sample will be classified into
that class. Figure 1.3 shows a simple example of how k-NN algorithm works.
Specially, when k = 1, it becomes the nearest neighbor algorithm. The higher
the value of k is, the less effect the noise will have on the classification. Since
the distance is the main metric of the k-NN algorithm, several functions can be
applied to define the distance between the unlabeled sample and its neighbors, such
as Chebyshev, City-block, Euclidean and Euclidean squared. For a more insightful
discussion on k-NN, please refer to [3].
1.1.2 Decision Tree (DT)
The DT is one of the classification techniques which performs classification through

a learning tree. In the tree, each node represents a feature (attribute) of a data, all
branches represent the conjunctions of features that lead to classifications, and each
leaf node is a class label. The unlabeled sample can be classified by comparing
its feature values with the nodes of the decision tree [4, 5]. The DT has many
advantages, such as intuitive knowledge expression, simple implementation and
high classification accuracy. ID3 [6], C4.5 [7] and CART [8] are three widely-
used decision tree algorithms to perform the classification of training dataset
1.1 Supervised Learning 3
Machine learning techniques
Supervised learning
k-nearest neighbour
Decision tree
Random forest
Neural network
Support vector machine
Bayes' theory
Hidden markov model
Unsupervised learning
k-means
Self-organizing map
Semi-supervised learning
Reinforcement learning
Reinforcement learning
Deep reinforcement learning
RL-based game theory
Fig. 1.2 Common machine learning algorithms applied to SDN
automatically. The biggest difference among them is the splitting criteria which are
used to build decision trees. The splitting criteria used by ID3, C4.5 and CART are
Information gain, Gain ratio and Gini impurity respectively. Reference [9] gives a
detailed comparison of the three DT algorithms.
1.1.3 Random Forest
The random forest method [10], also known as random decision forest, can be used
for classification and regression tasks. A random forest consists of many decision
trees. To mitigate over-fitting of decision tree method and improve accuracy, the
Fig. 1.3 Example of k-NN

algorithm, for k = 5. Among A
the five closest neighbors, one
neighbor belongs to class A A
and four neighbors belong to
class B. In this case, the
unlabeled example will be
classified into class B A A
B
B
B ?
B
A
B
W ij(t) Wij(2)
(2)
X1 a1 Yt
(2)
X2 a2 Y2
. . .
. . .
. . .
(2)
Xm ak Yn
Input Layer Hidden Layer Output Layer
Fig. 1.4 A basic neural network with three layers: an input layer, a hidden layer and an output
layer. An input has m features (i.e., X1 , X2 , . . . , Xm ) and the input can be assigned to n possible
classes (i.e., Y1 , Y2 , . . . , Yn ). Also, Wijl denotes the variable link weight between the ith neuron of
layer l and the j th neuron of layer l + 1, and akl denotes the activation function of the kth neuron
in layer l
random forest method randomly chooses only a subset of the feature space to
construct each decision tree. The steps to classify a new data sample by using
random forest method are: (a) put the data sample to each tree in the forest. (b)
Each tree gives a classification result, which is the tree’s “vote”. (c) The data sample
will be classified into the class which has the most votes.
1.1.4 Neural Network (NN)
A neural network is a computing system made up of a large number of simple

processing units, which operate in parallel to learn experiential knowledge from
historical data [11]. The concept of neural networks is inspired by the human
brain, which uses basic components, known as neurons to perform highly complex,
nonlinear and parallel computations. In a NN, its nodes are the equivalent compo-
nents of the neurons in the human brain. These nodes use activation functions to
perform nonlinear computations. The most frequently used activation functions are
the sigmoid and the hyperbolic tangent functions [12]. Simulating the way neurons
are connected in the human brain, the nodes in a NN are connected to each other by
variable link weights.
A NN has many layers. The first layer is the input layer and the last layer is the
output layer. Layers between the input layer and the output layer are hidden layers.
The output of each layer is the input of the next layer and the output of the last layer
is the result. By changing the number of hidden layers and the number of nodes in
each layer, complex models can be trained to improve the performance of NNs. NNs
are widely used in many applications, such as pattern recognition. The most basic
NN has three layers, including an input layer, a hidden layer and an output layer,
which is shown in Fig. 1.4.
There are many types of neural networks, which are often divided into two
training types, supervised or unsupervised [13]. In the following, we will give
a detailed representation of supervised neural networks which have been applied
in the field of SDN. In Sect. 1.2.2, self-organizing map, a representative type of
unsupervised neural networks, will be described.
Random NN
The random NN can be represented as an interconnected network of neurons which

exchange spiking signals. The main difference between random NN and other neural
networks is that neurons in random NN exchange excitatory and inhibitory spiking
signals probabilistically. In random NN, the internal excitatory state of each neuron
is represented by an integer, which is called “potential”. The potential value of
each neuron rises when it receives an excitatory spiking signal and drops when it
receives an inhibitory spiking signal. Neurons whose potential values are strictly
positive are allowed to send out excitatory or inhibitory spiking signals to other
neurons according to specific neuron-dependent spiking rates. When a neuron sends
out a spiking signal, its potential value drops one. The random NN has been used
in classification and pattern recognition [14]. For a more insightful discussion on
random NN, please refer to [14–16].
Deep NN
Neural networks with a single hidden layer are generally referred to as shallow
NNs. In contrast, neural networks with multiple hidden layers between the input
layer and the output layer are called deep NNs [17–19]. For a long time, shallow
NNs are often used. To process high-dimensional data and to learn increasingly
complex models, deep NNs with more hidden layers and neurons are needed.
However, deep NNs increase the training difficulties and require more computing
resources. In recent years, the development of hardware data processing capabilities
(e.g., GPU and TPU) and the evolved activation functions (e.g., ReLU) make it
possible to train deep NNs [20]. In deep NNs, each layer’s neurons train a feature
representation based on the previous layer’s output, which is known as feature
hierarchy. The feature hierarchy makes deep NNs capable of handling large high-
dimensional datasets. Due to the multiple-level feature representation learning,
compared to other machine learning techniques, deep NNs generally provide much
better performance [20].
Convolutional NN
Convolutional NN and recurrent NN are two major types of deep NNs. Convo-
lutional NN [21, 22] is a feed-forward neural network. Local sparse connections
among successive layers, weight sharing and pooling are three basic ideas of
convolutional NN. Weight sharing means that weight parameters of all neurons in
the same convolution kernel are same. Local sparse connections and weight sharing
can reduce the number of training parameters. Pooling can be used to reduce the
feature size while maintaining the invariance of features. The three basic ideas
reduce the training difficulties of convolutional NNs greatly.
Recurrent NN
In feed-forward neural networks, the information is transmitted directionally from

the input layer to the output layer. However, recurrent NN [23, 24] is a stateful
network, which can use internal state (memory) to handle sequential data. A typical
recurrent NN and its unrolled form are shown in Fig. 1.5. Xt is the input at time step
t. ht is the hidden state at time step t. ht captures information about what happened
in all the previous time steps, so it is called “memory”. Yt is the output at time step
t. U , V and W are parameters in the recurrent NN. Unlike a traditional deep NN,
which uses different parameters at each layer, the recurrent NN shares the same
parameters (i.e., U , V and W ) across all time steps. This means that at each time
step, the recurrent NN performs the same task, just with different inputs. In this way,
the total number of parameters needed to be trained is reduced greatly. Long Short-
Term Memory (LSTM) [25, 26] is the most commonly-used type of recurrent NNs,
which has a good ability to capture long-term dependencies. LSTM uses three gates
(i.e., an input gate, an output gate and a forget gate) to compute the hidden state.
Fig. 1.5 A typical recurrent NN and its unrolled form. Xt is the input at time step t. ht is the
hidden state at time step t. Yt is the output at time step t. U , V and W are parameters in the
recurrent NN
1.1.5 Support Vector Machine (SVM)
SVM is another popular supervised learning method, invented by Vapnik and others
[27], which has been widely used in classification and pattern recognition. The
basic idea of SVM is to map the input vectors into a high-dimensional feature
space. This mapping is achieved by applying different kernel functions, such as
linear, polynomial and Radial Based Function (RBF). Kernel function selection is an
important task in SVM, which has effect on the classification accuracy. The selection
of kernel function depends on the training dataset. The linear kernel function works
well if the dataset is linearly separable. If the dataset is not linearly separable,
polynomial and RBF are two commonly-used kernel functions. In general, the RBF-
based SVM classifier has a relatively better performance than the other two kernel
functions [28, 29].
The objective of SVM is to find a separating hyperplane in the feature space to
maximize the margin between different classes. Note that, the margin is the distance
between the hyperplane and the closest data points of each class. The corresponding
closest data points are defined as support vectors. An example of SVM classifier is
shown in Fig. 1.6. From the figure, there are many possible separating hyperplanes
between two classes, but only one optimal separating hyperplane can maximize the
margin. For a more insightful discussion on SVM, please refer to [30–32].
1.1.6 Bayes’ Theory
Bayes’ theory uses the conditional probability to calculate the probability of an

event occurring given the prior knowledge of conditions that might be related to
the event. The Bayes’ theory is defined mathematically as the following equation:
Hl
X2
H 2 (Optimal hyperplane )
H3
support vectors
margin
Xl
Fig. 1.6 An example of SVM classifier with an optimal linear hyperplane. There are two classes
in the figure, and each class has one support vector. As it can be seen, there are many possible
separating hyperplanes between two classes, such as H1 , H2 and H3 , but only one optimal
separating hyperplane (i.e., H2 ) can maximize the margin
P (E|H )P (H )
P (H |E) = (1.1)
P (E)
where E is a new evidence, H is a hypothesis, P (H |E) is the posterior probability

that the hypothesis H holds given the new evidence E, P (E|H ) is the posterior
probability that of evidence E conditioned on the hypothesis H , P (H ) is the prior
probability of hypothesis H , independent of evidence E, and P (E) is the probability
of evidence E.
In a classification problem, the Bayes’ theory learns a probability model by using
the training dataset. The evidence E is a data sample, and the hypothesis H is the
class to assign for the data sample. The posterior probability P (H |E) represents
the probability of a data sample belonging to a class. In order to calculate the
posterior probability P (H |E), P (H ), P (E) and P (E|H ) need to be calculated first
based on the training dataset using the probability and statistics theories, which
is the learning process of the probability model. When classifying a new input
data sample, the probability model can be used to calculate multiple posterior
probabilities for different classes. The data sample will be classified into the class
with the highest posterior probability P (H |E). The advantage of the Bayes’ theory
is that it requires a relatively small number of training dataset to learn the probability
model [33]. However, there is an important independence assumption when using
the Bayes’ theory. To facilitate the calculation of P (E|H ), the features of data
samples in the training dataset are assumed to be independent of each other [34].
For a more insightful discussion on Bayes’ theory, please refer to [33, 35–38].
1.2 Unsupervised Learning 9
1.1.7 Hidden Markov Models (HMM)
HMM is one kind of Markov models. Markov models are widely used in randomly
dynamic environments which obey the memoryless property. The memoryless
property of Markov models means that the conditional probability distribution of
future states only relates to the value of the current state and is independent of all
previous states [39, 40]. There are other Markov models, such as Markov Chains
(MC). The main difference between HMM and other models is that HMM is often
applied in environments where system states are partially visible or not visible
at all.
1.2 Unsupervised Learning
In contrast to supervised learning, an unsupervised learning algorithm is given a

set of inputs without labels (i.e., there is no output). Basically, an unsupervised
learning algorithm aims to find patterns, structures, or knowledge in unlabeled
data by clustering sample data into different groups according to the similarity
between them. The unsupervised learning techniques are widely used in clustering
and data aggregation [2, 41]. In the following, we will give a detailed repre-
sentation of widely-used unsupervised learning algorithms, such as k-means and
self-organizing map.
1.2.1 k-Means
The k-means algorithm is a popular unsupervised learning algorithm, which is

used to recognize a set of unlabeled data into different clusters. To implement the
k-means algorithm, only two parameters (i.e., the initial dataset and the desired
number of clusters) are needed. If the desired number of clusters is k, the steps
to resolve node clustering problem by using k-means algorithm are: (a) initialize
k cluster centroids by randomly choosing k nodes; (b) use a distance function to
label each node with the closest centroid; (c) assign new centroids according to the
current node memberships and (d) stop the algorithm if the convergence condition
is valid, otherwise go back to step (b). An example procedure of k-means algorithm
is shown in Fig. 1.7. For a more insightful discussion on k-means, please refer to
[2, 42].
Fig. 1.7 Example of k-means algorithm, for k = 2. (a) Randomly choosing two data points as two
centroids; (b) label each node with the closest centroid, resulting that node A and B are a class,
node C, D and E are another class; (c) assign new centroids; (d) label each node with the closest
centroid again, resulting that node A, B and C are a class, node D and E are another class; (e) the
algorithm is converged
1.2.2 Self-Organizing Map (SOM)
SOM, also known as Self-Organizing Feature Map (SOFM), is one of the most
popular unsupervised neural network models. SOM is often applied to perform
dimensionality reduction and data clustering. In general, SOM has two layers, an
input layer and a map layer. When SOM is used to perform data clustering, the
number of neurons in the map layer is equal to the desired number of clusters.
Each neuron has a weight vector. The steps to resolve data clustering problem by
using SOM algorithm are: (a) initialize the weight vector of each neuron in the
map layer; (b) choose a data sample from the training dataset; (c) use a distance
function to calculate the similarity between the input data sample and all weight
vectors. The neuron whose weight vector has the highest similarity is called the Best
Matching Unit (BMU). The SOM algorithm is based on competitive learning, which
means that there is only one BMU each time. (d) The neighborhood of the BMU
References 11
is calculated. (e) The weight vectors of the neurons in the BMU’s neighborhood
(including the BMU itself) are adjusted towards the input data sample. (f) Stop the
algorithm if the convergence condition is valid, otherwise go back to step (b). For a
more insightful discussion on SOM, please refer to [43, 44].
1.3 Semi-supervised Learning
Semi-supervised learning [45, 46] is a type of learning which uses both labeled
and unlabeled data. Semi-supervised learning is useful for a few reasons. First,
in many real-world applications, the acquisition of labeled data is expensive and
difficult while acquiring a large amount of unlabeled data is relatively easy and
cheap. Second, effective use of unlabeled data during the training process actually
tends to improve the performance of the trained model. In order to make the best
use of unlabeled data, assumptions have to be hold in semi-supervised learning, such
as smoothness assumption, cluster assumption, low-density separation assumption,
and manifold assumption. Pseudo Labeling [47, 48] is a simple and efficient semi-
supervised learning technique. The main idea of Pseudo Labeling is simple. Firstly,
use the labeled data to train a model. Then, use the trained model to predict pseudo
labels of the unlabeled data. Finally, combine the labeled data and the newly pseudo-
labeled data to train the model again. There are other semi-supervised learning
methods, such as Expectation Maximization (EM), co-training, transductive SVM
and graph-based methods. Different methods rely on different assumptions [49].
For example, EM builds on cluster assumption, transductive SVM builds on low-
density separation assumption, while graph-based methods build on the manifold
assumption.
References
1. S. B. Kotsiantis, I. Zaharakis, and P. Pintelas, “Supervised machine learning: A review of clas-

sification techniques,” Emerging Artificial Intelligence Applications in Computer Engineering,
vol. 160, pp. 3–24, 2007.
2. J. Friedman, T. Hastie, and R. Tibshirani, The Elements of Statistical Learning. Springer Series
in Statistics New York, 2001, vol. 1.
3. T. Cover and P. Hart, “Nearest neighbor pattern classification,” IEEE Trans. Information
Theory, vol. 13, no. 1, pp. 21–27, Jan. 1967.
4. L. Breiman, J. Friedman, C. J. Stone, and R. A. Olshen, Classification and Regression Trees.
CRC Press, 1984.
5. J. Han, J. Pei, and M. Kamber, Data Mining: Concepts and Techniques. Elsevier, 2011.
6. J. R. Quinlan, “Induction of decision trees,” Machine Learning, vol. 1, no. 1, pp. 81–106, 1986.
7. S. Karatsiolis and C. N. Schizas, “Region based support vector machine algorithm for medical
diagnosis on Pima Indian Diabetes dataset,” in Proc. IEEE BIBE’12, Larnaca, Cyprus, Nov.
2012, pp. 139–144.
8. W. R. Burrows, M. Benjamin, S. Beauchamp, E. R. Lord, D. McCollor, and B. Thomson,

“CART decision-tree statistical analysis and prediction of summer season maximum surface
ozone for the Vancouver, Montreal, and Atlantic regions of Canada,” Journal of Applied
Meteorology, vol. 34, no. 8, pp. 1848–1862, 1995.
9. A. Kumar, P. Bhatia, A. Goel, and S. Kole, “Implementation and comparison of decision tree
based algorithms,” International Journal of Innovations & Advancement in Computer Science,
vol. 4, pp. 190–196, May. 2015.
10. L. Breiman, “Random forests,” Machine Learning, vol. 45, no. 1, pp. 5–32, 2001.
11. S. Haykin, Neural Networks: A Comprehensive Foundation. Prentice Hall PTR, 1994.
12. S. Haykin and N. Network, “A comprehensive foundation,” Neural Networks, vol. 2, no. 2004,
p. 41, 2004.
13. K. Lee, D. Booth, and P. Alam, “A comparison of supervised and unsupervised neural networks
in predicting bankruptcy of Korean firms,” Expert Systems with Applications, vol. 29, no. 1,
pp. 1–16, 2005.
14. S. Timotheou, “The random neural network: A survey,” The Computer Journal, vol. 53, no. 3,
pp. 251–267, March 2010.
15. S. Basterrech and G. Rubino, “A tutorial about random neural networks in supervised learning,”
arXiv preprint arXiv:1609.04846, 2016.
16. H. Bakirciouglu and T. Koccak, “Survey of random neural network applications,” European
Journal of Operational Research, vol. 126, no. 2, pp. 319–330, 2000.
17. Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, p. 436, 2015.
18. J. Baker, “Artificial neural networks and deep learning,” Feb. 2015. [Online]. Available: http://
lancs.ac.uk/~bakerj1/pdfs/ANNs/Artificial_neural_networks-report.pdf
19. J. Schmidhuber, “Deep learning in neural networks: An overview,” Neural Networks, vol. 61,
pp. 85–117, 2015.
20. G. Pandey and A. Dukkipati, “Learning by stretching deep networks,” in International
Conference on Machine Learning, 2014, pp. 1719–1727.
21. A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional
neural networks,” in Advances in Neural Information Processing Systems, 2012, pp. 1097–
1105.
22. C. Li, Y. Wu, X. Yuan, Z. Sun, W. Wang, X. Li, and L. Gong, “Detection and defense
of DDoS attack-based on deep learning in OpenFlow-based SDN,” International Journal of
Communication Systems, 2018.
23. T. Mikolov, M. Karafiát, L. Burget, J. Černockỳ, and S. Khudanpur, “Recurrent neural
network based language model,” in Eleventh Annual Conference of the International Speech
Communication Association, 2010.
24. H. Sak, A. Senior, and F. Beaufays, “Long short-term memory recurrent neural network
architectures for large scale acoustic modeling,” in Fifteenth Annual Conference of the
International Speech Communication Association, 2014.
25. S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural Computation, vol. 9,
no. 8, pp. 1735–1780, 1997.
26. X. Li and X. Wu, “Constructing long short-term memory based deep recurrent neural networks
for large vocabulary speech recognition,” in Proc. IEEE ICASSP’15, Brisbane, QLD, Australia,
April 2015, pp. 4520–4524.
27. V. N. Vapnik and V. Vapnik, Statistical Learning Theory. Wiley New York, 1998, vol. 1.
28. B. Yekkehkhany, A. Safari, S. Homayouni, and M. Hasanlou, “A comparison study of different
kernel functions for SVM-based classification of multi-temporal polarimetry SAR data,” The
International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences,
vol. 40, no. 2, p. 281, 2014.
29. A. Patle and D. S. Chouhan, “SVM kernel functions for classification,” in Proc. IEEE
ICATE’13, Mumbai, India, Jan 2013, pp. 1–9.
30. I. Steinwart and A. Christmann, Support Vector Machines. Springer Science & Business Media,
2008.
References 13
31. M. Martínez-Ramón and C. Christodoulou, “Support vector machines for antenna array
processing and electromagnetics,” Synthesis Lectures on Computational Electromagnetics,
vol. 1, no. 1, pp. 1–120, 2005.
32. H. Hu, Y. Wang, and J. Song, “Signal classification based on spectral correlation analysis and
SVM in cognitive radio,” in Proc. IEEE AINA’08, Okinawa, Japan, March. 2008, pp. 883–887.
33. G. E. Box and G. C. Tiao, Bayesian Inference in Statistical Analysis. John Wiley & Sons, 2011,
vol. 40.
34. J. Bakker, “Intelligent traffic classification for detecting DDoS attacks using SDN/OpenFlow,”
Victoria University of Wellington, pp. 1–142, 2017.
35. N. Friedman, D. Geiger, and M. Goldszmidt, “Bayesian network classifiers,” Machine Learn-
ing, vol. 29, no. 2–3, pp. 131–163, 1997.
36. F. V. Jensen, An Introduction to Bayesian Networks. UCL Press London, 1996, vol. 210.
37. D. Heckerman et al., “A tutorial on learning with Bayesian networks,” Nato Asi Series D
Behavioural And Social Sciences, vol. 89, pp. 301–354, 1998.
38. T. D. Nielsen and F. V. Jensen, Bayesian Networks and Decision Graphs. Springer Science &
Business Media, 2009.
39. L. R. Rabiner, “A tutorial on hidden markov models and selected applications in speech
recognition,” Proceedings of the IEEE, vol. 77, no. 2, pp. 257–286, Feb. 1989.
40. P. Holgado, V. A. VILLAGRA, and L. Vazquez, “Real-time multistep attack prediction based
on hidden markov models,” IEEE Trans. Dependable and Secure Computing, vol. PP, no. 99,
pp. 1–1, 2017.
41. E. Alpaydin, Introduction to Machine Learning. MIT Press, 2014.
42. T. Kanungo, D. M. Mount, N. S. Netanyahu, C. D. Piatko, R. Silverman, and A. Y. Wu, “An
efficient k-means clustering algorithm: Analysis and implementation,” IEEE Trans. Pattern
Analysis and Machine Intelligence, vol. 24, no. 7, pp. 881–892, Jul. 2002.
43. T. Kohonen, “The self-organizing map,” Neurocomputing, vol. 21, no. 1–3, pp. 1–6, 1998.
44. M. M. Van Hulle, “Self-organizing maps,” in Handbook of Natural Computing. Springer, 2012,
pp. 585–622.
45. X. Zhu, “Semi-supervised learning literature survey,” Citeseer, pp. 1–59, 2005.
46. X. Zhou and M. Belkin, “Semi-supervised learning,” in Academic Press Library in Signal
Processing. Elsevier, 2014, vol. 1, pp. 1239–1269.
47. D.-H. Lee, “Pseudo-label: The simple and efficient semi-supervised learning method for deep
neural networks,” in Workshop on Challenges in Representation Learning, ICML, vol. 3, 2013,
p. 2.
48. H. Wu and S. Prasad, “Semi-supervised deep learning using Pseudo labels for hyperspectral
image classification,” IEEE Trans. Image Processing, vol. 27, no. 3, pp. 1259–1270, March
2018.
49. O. Chapelle, B. Scholkopf, and A. Zien, “Semi-supervised learning (chapelle, o. et al., eds.;
2006)[book reviews],” IEEE Trans. Neural Networks, vol. 20, no. 3, pp. 542–542, 2009.
Chapter 2
Reinforcement Learning and Deep
Reinforcement Learning
Abstract In order to better understand state-of-the-art reinforcement learning

agent, deep Q-network, a brief review of reinforcement learning and Q-learning are
first described. Then recent advances of deep Q-network are presented, and double
deep Q-network and dueling deep Q-network that go beyond deep Q-network are
also given.
2.1 Reinforcement Learning
Reinforcement learning is an important branch of machine learning, where an agent

learns to take actions that would yield the most reward by interacting with the
environment. Different from supervised learning, reinforcement learning cannot
learn from samples provided by an experienced external supervisor. Instead, it has to
operate based on its own experience despite that it faces with significant uncertainty
about the environment.
Reinforcement learning is defined not characterizing learning methods, but
by characterizing a learning problem. Any method that is suitable for solv-
ing that problem can be considered as a reinforcement learning method [1].
A reinforcement learning problem can be described as an optimal control of a
Markov Decision Process (MDP), however, state space, explicit transition proba-
bility and reward function are not necessarily required [2]. Therefore, reinforce-
ment learning is powerful in handling tough situations that approach real-world
complexity [3].
There are two outstanding features of reinforcement learning: trial-and-error
search and delayed reward. Trial-and-error search means making trade-off between
exploration and exploitation. The agent prefers to exploit the effective actions that
have been tried in the past to produce rewards, but it also has to explore better
new actions that may yield higher rewards in the future. The agent must try various
actions and progressively favor those that earn the most rewards. The other feature of
reinforcement learning is that the agent looks into a big picture, not just considering
the immediate reward, but also the cumulative rewards in the long run, which is
specified as value function.

16 2 Reinforcement Learning and Deep Reinforcement Learning
Action (at)
State (st)
Agent Environment
State (st+1)
Reward r(st,at)
Fig. 2.1 A basic diagram of a RL system. The agent takes an action according to the current state
and then receives a reward. r(st , at ) denotes the immediate reward that the agent receives after
performing an action at at the state st
Generally, reinforcement learning can be divided into model-free and model-

based reinforcement learning, which is based on whether or not the environment
elements are already known. Model-free reinforcement learning has recently been
successfully applied to handle the deep neural network and value functions [3–
6]. It can learn policies for tough tasks using the raw state representation directly
as the input to the neural networks [7]. Contrastively, model-based reinforcement
learn a model of the system with the help of supervised learning and optimize a
policy under this model [7–9]. Recently, elements of model-based RL have been
incorporated into model-free deep reinforcement learning to accelerate the learning
rate without losing the strengths of model-free learning [7]. One of the model-free
reinforcement learning algorithms is Q-Learning. The most important component
of Q-learning algorithms is a method for properly and efficiently estimating the
Q value. The Q function can be implemented simply by a look-up table, or by
a function approximator, sometimes a nonlinear approximator, such as a neural
network, or even more complicated deep neural network. Q-learning, combined
with deep neural network, is the so-called deep Q-learning.
2.2 Deep Q-Learning
The agent that uses a neural network to represent Q function is called Q-network,
which is denoted as Q(x, a; θ ). The parameter θ stands for the weights of the neural
network, and the Q-network is trained by updating θ at each iteration to approximate
the real Q values. While neural networks allow for great flexibility, they do so
at the cost of stability when it comes to Q-Learning, which is interpreted in [3].
Deep Q-network that uses deep neural networks instead of approximating the Q-
function is proposed recently, and it is proven to be more advantageous with greater
performance and more robust learning [3]. To transform an ordinary Q-network into
a deep Q-network, three improvements have been implemented.
• Replacing ordinary neural networks with advanced multi-layer deep convolu-
tional networks, which utilize hierarchical layers of tiled convolutional filters to
2.3 Beyond Deep Q-Learning 17
exploit the local spatial correlations, and make it possible to extract high-level
features from raw input data [2, 10].
• Utilizing Experience Replay, which stores its interaction experience tuple
e(t) = (x(t), a(t), r(t), x(t + 1)) at time t into a replay memory D(t) =
{e(1), . . . , e(t)}, and then randomly samples batches from the experience pool
to train the deep convolutional network’s parameters rather than directly using
the consecutive samples as in Q-learning. This allows the network to learn from
more various past experiences, and restrains the network from only focusing on
what it is immediately doing.
• Adopting a second network, to generate the target Q values that are used to
calculate the loss for each action during the training procedure. One network
for both estimations of Q values and target Q values would result in falling
into feedback loops between the target and estimated values. Thus, in order to
stabilize the training, the target network’s weights are fixed and periodically
updated.
The deep Q-function is trained towards the target value by minimizing the loss
function L(θ ) at each iteration, and the loss function can be written as
Li (θi ) = E[(yi − Q(x, a; θi )2 )], (2.1)
where yi = r + ε maxa Q(x , a ; θi− ). Here, the weights θi− are updated as θi− =
θi−G , i.e., the weights update every G time steps in the deep Q-networks instead of
θi− = θi−1 .
2.3 Beyond Deep Q-Learning
Since deep Q-learning is first proposed, great efforts have been made for even
greater performance and higher stability. Here, we briefly introduce two recent
improvements: Double DQN [11] and Dueling DQN [4].
2.3.1 Double DQN
In regular DQN, both choosing an action and evaluating the chosen action use the
max over the Q values, which would lead to over-optimistic Q value estimation. To
relieve the over-estimation problem, the target value in double DQN is designed and
updated as
yidouble = r + εQ(x , arg max Q(x , a ; θi ); θi− ), (2.2)

a
18 2 Reinforcement Learning and Deep Reinforcement Learning
where the action choice is decoupled from the target Q-value generation. This
simple trick makes the over-estimation significantly reduced, and the training
procedure runs faster and more reliably.
2.3.2 Dueling DQN
The intuition behind Dueling DQN is that it is not always necessary to estimate the
value of taking each available action. For some states, the choice of action makes no
influence on what happens. Thus, in dueling DQN, the state-action value Q(x, a) is
decomposed into two components as follows,
Q(x, a) = V (x) + A(a). (2.3)
Here, V (x) is the value function, which simply represents how good it is to be
in a given state x. A(a) is the advantage function, which measures the relative
importance of a certain action compared with other actions. After V (x) and A(a)
are separately computed, their values are combined back into a single Q-function at
the final layer. This improvement would lead to better policy evaluation.
Combining both the above two techniques can achieve better performance
and faster training speed. In the following section, the deep Q-network with the
two improvements is utilized to find the optimal policy for a practical network
application.
References
1. R. S. Sutton, “Introduction: The challenge of reinforcement learning,” in Reinforcement

Learning. Springer, 1992, pp. 1–3.
2. H. Y. Ong, K. Chavez, and A. Hong, “Distributed deep Q-learning,” arXiv preprint
arXiv:1508.04186, 2015.
3. V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves,
M. Riedmiller, A. K. Fidjeland, G. Ostrovski et al., “Human-level control through deep
reinforcement learning,” Nature, vol. 518, no. 7540, pp. 529–533, 2015.
4. Z. Wang, N. de Freitas, and M. Lanctot, “Dueling network architectures for deep reinforcement
learning,” arXiv preprint arXiv:1511.06581, 2015.
5. M. Hausknecht and P. Stone, “Deep reinforcement learning in parameterized action space,”
arXiv preprint arXiv:1511.04143, 2015.
6. T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa, D. Silver, and D. Wierstra,
“Continuous control with deep reinforcement learning,” arXiv preprint arXiv:1509.02971,
2015.
7. S. Gu, T. Lillicrap, I. Sutskever, and S. Levine, “Continuous deep Q-learning with model-based
acceleration,” arXiv preprint arXiv:1603.00748, 2016.
8. S. Levine, C. Finn, T. Darrell, and P. Abbeel, “End-to-end training of deep visuomotor
policies,” Journal of Machine Learning Research, vol. 17, no. 39, pp. 1–40, 2016.
References 19
9. M. Deisenroth and C. E. Rasmussen, “Pilco: A model-based and data-efficient approach to

policy search,” in Proceedings of the 28th International Conference on machine learning
(ICML-11), 2011, pp. 465–472.
10. V. Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wierstra, and M. Riedmiller,
“Playing atari with deep reinforcement learning,” arXiv preprint arXiv:1312.5602, 2013.
11. H. Van Hasselt, A. Guez, and D. Silver, “Deep reinforcement learning with double q-learning,”
CoRR, abs/1509.06461, 2015.
Chapter 3
Deep Reinforcement Learning
for Interference Alignment Wireless
Networks
Abstract Both caching and interference alignment (IA) are promising techniques
for next generation wireless networks. Nevertheless, most existing works on cache-
enabled IA wireless networks assume that the channel is invariant, which is
unrealistic considering the time-varying nature of practical wireless environments.
In this chapter, we consider realistic time-varying channels. Specifically, the channel
is formulated as a finite-state Markov channel (FSMC). The complexity of the
system is very high when we consider realistic FSMC models. Therefore, in
this chapter, we propose a novel deep reinforcement learning approach, which
is an advanced reinforcement learning algorithm that uses deep Q network to
approximate the Q value-action function. We use Google TensorFlow to implement
deep reinforcement learning in this chapter to obtain the optimal IA user selection
policy in cache-enabled opportunistic IA wireless networks. Simulation results are
presented to show that the performance of cache-enabled opportunistic IA networks
in terms of the network’s sum rate and energy efficiency can be significantly
improved by using the proposed approach.
3.1 Introduction
Recently, wireless proactive caching has attracted great attentions from both
academia and industry [1, 2]. By effectively reducing the duplicate content transmis-
sions in networks, caching has been recognized as one of the promising techniques
for future wireless networks to improve spectral efficiency, shorten latency, and
reduce energy consumption [3–5]. Based on the global traffic features, a few popular
contents are requested by many users during a short time span, which accounts for
most of the traffic load. Therefore, proactively caching the popular contents can
remove the heavy burden of the backhaul links.
Another new technology called interference alignment (IA) has been studied
extensively as a revolutionary technique to tackle the interference issue in wireless
networks [6]. IA exploits the cooperation of transmitters to design the precoding
matrices, and thus eliminates the interferences. IA can benefit mobile cellular
networks [7]. Due to the large number of users in cellular networks, multiuser

22 3 Deep Reinforcement Learning for Interference Alignment Wireless Networks
diversity has been studied in conjunction with IA, called opportunistic IA, which
further improves the network performance [8–11].
Jointly considering these two important technologies, caching and IA, can be
beneficial in IA-based wireless networks [12, 13]. The implementation of IA
requires the channel state information (CSI) exchange among transmitters, which
usually relies on the backhaul link. The limited capacity of backhaul link has
significant impacts on the performance of IA [14]. Caching can relieve the traffic
loads of backhaul links, thus the saved capacity can be used for CSI exchange in
IA. In [12], the authors investigate the benefits of caching and IA in the context of
multiple-input and multiple-output (MIMO) interference channels, and maximize
the average transmission rate by optimizing the number of the active transceiver
pairs. In [13], it is shown that by properly placing the content in the transmitters’
caches, the IA gain can be increased.
Although some excellent works have been done on caching and IA, most of these
previous works assume that the channel is block-fading channel or invariant channel,
where the estimated CSI of the current time instant is simply taken as the predicted
CSI for the next time instant. Considering the time-varying nature of wireless
environments, this kind of memoryless channel assumption is not realistic[15],
especially in vehicular environments due to high mobility [16]. In addition, it is
difficult to obtain the perfect CSI due to channel estimation errors, communication
latency and backhaul link constraints [17, 18].
In this chapter, we consider realistic time-varying channels, and propose a novel
deep reinforcement learning approach in cache-enabled opportunistic IA wireless
networks. Deep reinforcement learning is a kind of machine learning, which is a
powerful tool to process wireless data[19]. The distinct features of this chapter are
as follows.
• Cache-enabled opportunistic IA is studied under the condition of time-varying
channel coefficients. The channel is formulated as a finite-state Markov channel
(FSMC), which has been widely accepted in the literature to characterize
the correlation structure of the fading process [20–22]. Considering FSMC
models may enable significant performance improvements over the schemes with
memoryless channel models.
• The complexity of the system is very high when we consider realistic FSMC
models. Therefore, we propose a novel deep reinforcement learning approach in
this chapter. Deep reinforcement learning is an advanced reinforcement learning
algorithm that uses deep Q network to approximate the Q value-action function
[23], and it has been used in wireless networks to improve the performance[24,
25]. Deep reinforcement learning is used in this chapter to obtain the optimal IA
user selection policy in cache-enabled opportunistic IA wireless networks.
• We use Google TensorFlow to implement deep reinforcement learning. The visu-
alization of the deep Q network model is presented. Simulation results with dif-
ferent system parameters are presented to show the effectiveness of the proposed
scheme. It is illustrated that the performance of cache-enabled opportunistic IA
3.2 System Model 23
networks in terms of the network’s sum rate and energy efficiency can be signif-
icantly improved by using the proposed deep reinforcement learning approach.
The rest of this chapter is organized as follows. In Sect. 3.2, the system
model is presented. The deep reinforcement learning algorithm is presented in
Sect. 3.3. In Sect. 3.3, the cache-enabled opportunistic IA network is formulated as
a deep reinforcement learning process. Simulation results are discussed in Sect. 3.4.
Finally, conclusions are presented in Sect. 3.5.
3.2 System Model
In this section, we describe the model of IA, followed by the time-varying channel.
Then, cache-equipped transmitters are described.
3.2.1 Interference Alignment
IA is a revolutionary interference management technique, which theoretically

enables the network’s sum rate grow linearly with the cooperative transmitter and
receiver pairs. That is to say, each user can obtain the capacity 12 log(SNR) +
o(log(SNR)), which has nothing to do with the interferences [26]. Actually, SNR
plays a crucial role in determining the IA results. Cadambe and Jafar pointed out that
IA performs better at very high SNR, and suffers from low quality at moderate SNR
levels [26]. Meanwhile higher and higher SNR is required to approach IA network’s
theoretical maximum sum rate as the number of IA users increases [6]. Thus, there
exist competitions among users for accessing to IA network.
Consider a K-user MIMO interference channel. Nt[k] and Nr[k] antennas are
equipped at the kth transmitter and receiver, respectively. The degree of freedom
(DoF) of the kth user is denoted as d [k] . The received signal at the kth receiver can
be written as
y[k](t) = U[k]† (t)H[kk] (t)V[k] (t)x[k](t)

K (3.1)
+ U[k]† (t)H[kj ] (t)V[j ] (t)x[j ](t)+U[k]†(t)z[k](t),
j =1,j =k
where the first term at the right side represents the expected signal, and the other
two terms mean the inter-user interference and noise, respectively. H[kj ] (t) is the
[j ]
Nr[k] × Nt matrix of channel coefficients from the j th transmitter to the kth
receiver over the time slot t. Each element of H[kj ] (t) is independent and identically
distributed (i.i.d) complex Gaussian random variable, with zero mean and unit
variance. V[k] (t) and U[k] (t) are the unitary Nt[k] × d [k] precoding matrix and
Nr[k] × d [k] interference suppression matrix of the kth user, respectively. x[k] (t)
and z[k] (t) are the transmitted signal vector of d [k] dimensions and the N [k] × 1
Fig. 3.1 System model of a cache-enabled opportunistic IA wireless network
additive white Gaussian noise (AWGN) vector whose elements have zero mean and
σ 2 variance at the kth receiver, respectively.
The interference can be perfectly eliminated only when the following conditions
can be satisfied
U[k]† (t)H[kj ] (t)V[j ] (t) = 0, ∀j = k, (3.2)

rank U[k]† (t)H[kk] (t)V[k] (t) = d [k] . (3.3)
Under this assumption, the received signal at the kth receiver can be rewritten as
y[k] (t) = U[k]† (t)H[kk] (t)V[k] (t)x[k](t) + U[k]† (t)z[k] (t). (3.4)
To meet Condition (3), the global CSI is required at each transmitter. Each
transmitter can estimate its local CSI (i.e., the direct link), but the CSI of other links
can only be obtained by CSI share with other transmitters via the backhaul link[12].
Thus, in IA network, the backhaul link is more than a pipeline for connecting with
Internet. The limited capacity should be made optimum use of. The recent advances
focus on the benefits of edge caching, which is capable to decrease the data transfer
and leave more capacity for CSI share. The detail is described in the following
subsection. In this chapter, we assume the total backhaul link capacity of all the
users is Ctotal , and the CSI estimation is perfect with no errors and no time delay.
3.2.2 Cache-Equipped Transmitters
In the era of explosive information, the vast amount of content makes it impossible
for all of them gain popularity. As a matter of fact, only a small fraction becomes
extensively popular. That means certain content may be requested over and
3.3 Problem Formulation 25
over during a short time span, which gives rise to the network congestion and
transmission delay. We assume that each transmitter is equipped with a cache unit
that has certain amount of storage space. The stored content may follow a certain
popularity distribution.
For consistency, the cache of each transmitter stores the same content, usually
the web content, and thus alleviating the backhaul burden and shorten delay time.
In [27], the authors survey on the existing methods for predicting the popularity
of different types of web content. Specifically, they show that different types of
content follow different popularity distributions. For example, the popularity growth
of online videos complies with power-law or exponential distributions, that of the
online news can be represented by power-law or log-normal distributions, etc. Based
on the content popularity distribution and cache size, cache hit probability Phit
and cache miss probability Pmiss can be derived[12]. In this chapter, the specific
popularity distribution is not the focus, and we just concentrate on two states,
whether the requested content is within the cache or not. We describe the two states
as Λ = {0, 1}, where 0 means the requested content is not within the cache, and 1
indicates it is within the cache.
In this chapter, we consider an MIMO interference network with limited
backhaul capacity and caches equipped at the transmitter side, as illustrated in
Fig. 3.1. There is a central scheduler who is responsible for collecting the channel
state and cache status from each user, scheduling the users and allocating the limited
resources. All the users are connected to the central scheduler via a backhaul link
for CSI exchange and Internet connection, and the total capacity is limited.
When we consider realistic FSMC models, the complexity of the system is very
high, which makes it difficult to be solved using traditional approaches. In this
chapter, we used a novel reinforcement learning approach to solve the optimization
problem in cache-enabled opportunistic IA wireless networks.
3.3 Problem Formulation
In this section, we first formulate the realistic time-varying channels as finite-state

Markov channels (FSMC), and then we demonstrate how to formulate the cache-
enabled IA network optimization problem as a deep Q-learning process, which can
determine the optimal policy for IA user grouping.
3.3.1 Time-Varying IA-Based Channels
We model the realistic time-varying channels as FSMCs, which is an effective

method to characterize the fading nature of wireless channels[20]. Specifically, the
first-order FSMC is used in this chapter.
Here we consider a multi-user interference network with one DoF for each user,
and the received signal to interference and noise ratio (SINR) is an important
parameter that can be used to reflect the quality of a wireless channel. In an IA-
based wireless network, if the interference is completely eliminated, the received
SINR of the kth user at time instant t can be derived as follows,
|h[kk] (t)|2 P [k]

SINR[k] (t) = , (3.5)
σ2
where h[kk] (t) = u[k]† (t)H[kk] (t)v[k] (t). On the other hand, if the interference is not
perfectly eliminated, the received SINR is
|h[kk] (t)|2 P [k]

SINR[k] (t) = [kj ] |2 P [j ] + σ 2
, (3.6)
j =k |h
where h[kj ] (t) = u[k]† (t)H[kj ] (t)v[j ] (t).

As we consider opportunistic IA networks with user scheduling, directly mod-
eling the received SINR of a certain user as a Markov random variable is not
appropriate. Due to the relationship of the received SINR and the channel coeffi-
cient, we model the channel coefficient after using IA, |h[kj ] |2 , as a Markov random
variable. Here, |h[kj ] (t)|2 is newly modelled because of its unknown distribution.
We partition and quantize the range of |h[kj ] |2 into H levels. Each level corresponds
to a state of the Markov channel. This channel information is included in the
system state, which will be discussed in detail in the following subsection. We
consider T time slots over a period of wireless communications. Let us denote
t ∈ {0, 1, 2, . . . , T − 1} as the time instant, and the channel coefficient varies from
one state to another state when one time slot elapses.
3.3.2 Formulation of the Network’s Optimization Problem
In our system, there are L candidates that want to join in the IA network to
communicate wirelessly. We assume that the IA network size is always smaller than
the number of candidates, which is in accordance with the fact that a large number of
users expect wireless communications anytime and anywhere. As aforementioned,
the value of SNR affects the performance of interference alignment, and the candi-
dates who occupy the better channels are more advantageous for accessing to the IA
network. Therefore, we make an action at each time slot to decide which candidates
are the optimal users for constructing an IA network based on their current states.
Here, a central scheduler is responsible for acquiring each candidate’s CSI and
cache status, then it assembles the collected information into a system state. Next,
the controller sends the system state to the agent, i.e., the deep Q network, and then
the deep Q network feeds back the optimal action arg maxπ Q∗ (x, a) for the current
time instant. After obtaining the action, the central scheduler will send a bit to
inform the users to be active or not, and the corresponding precoding vector will be
sent to each active transmitter. The system will transfer to a new state after an action
is performed, and the rewards can be obtained according to the reward function.
Algorithm 1 Deep reinforcement learning algorithm in cache-enabled IA networks

Initialization.
Initialize replay memory.

Initialize the Q network parameters with θ.
Initialize the target Q network parameters with θ − = θ.
For episode k = 1, . . . , K do:
Initialize the beginning state x.
For t = 1, 2, 3 . . . , T do:
Choose a random probability p.

Choose a(t) as,
if p ≥ ε,
a ∗ (t) = arg maxa Q(x, a, θ),
otherwise,
randomly choose a solution a(t) = a ∗ (t),
Execute a(t) in the system. Observe the reward r(t), and next state x(t + 1). Store the
experience (x(t), a(t), r(t), x(t + 1)) in replay memory.
Get mini-batch of samples (x(t), a(t), ζ (t), x(t + 1)) from the replay memory.
Perform a gradient descent step on (y − Q(x, a, θ)2 ) with respect to θ.
Return the value of parameters θ in the deep Q network.
End for
End for
Inside the deep Q network, the replay memory stores the agent’s experience of
each time slot. The Q network parameter θ is updated at every time instant with
samples from the replay memory. The target Q network parameter θ − is copied from
the Q network every N time instants. The ε-greedy policy is utilized to balance the
exploration and exploitation, i.e., to balance the reward maximization based on the
knowledge already known with trying new actions to obtain knowledge unknown.
The training algorithm of the deep Q network is described in Algorithm 1.
In order to obtain the optimal policy, it is necessary to identify the actions, states
and reward functions in our deep Q learning model, which will be described in the
next following subsections.
System State
The current system state x(t) is jointly determined by the states of L candidates.
The system state at time slot t is defined as,
x(t) = {|h[11] (t)|2 , |h[12] (t)|2 , . . . , |h[kj ] (t)|2 , . . . , |h[LL] (t)|2 ,

(3.7)
c1 (t), c2 (t), . . . , cl (t), . . . , cL (t)}.
Here, there are two components in the state: the channel coefficient after being
processed by IA technique |h[kj ] (t)|2 and the cache state ci (t). The cache state
cl (t) ∈ Λ = {0, 1}, the index l means the lth candidate, and l = 1, 2, . . . , L.
The number of possible system states can be very large. Due to the curse of
dimensionality, it is difficult for traditional approaches to handle our problem.
Fortunately, deep Q network is capable of successfully learning directly from high-
dimensional inputs[23], thus it is proper to be used in our system.
System Action
In the system, the central scheduler has to decide which candidates to be set active,
and the corresponding resources will be allocated to the active users.
The current composite action a(t) is denoted by
a(t) = {a1 (t), a2 (t), . . . , aL (t)}, (3.8)
where al (t) represents the control of the lth candidate, i.e., al (t) = 0 means the
candidate l is passive (not selected) at time slot t, and al (t) = 1 means it is
active (selected). Due to the constraint of IA, the condition L l=1 al (t) ≥ 3 must
be satisfied.
Reward Function
The system reward represents the optimization objective, and we take the objective
to maximize the IA network’s throughput, and the reward function of the lth
candidate is defined as,
rl (t) =
⎧ ⎛ ⎞
⎪
⎪
⎪
⎪ ⎜ h[ll] (t) 2 P [l] xl ⎟
⎪
⎪ ⎜ ⎟
⎪
⎪ a (t) log ⎜ 1 + ⎟,
⎪
⎪
l 2
⎝ L 2 ⎠
⎪
⎪ [lj ]
aj (t) h (t) P xj + σ [j ]
⎪
⎪
2
⎪
⎪ j =1,j =l
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪ if cl (t) = 1,
⎪
⎪
⎪
⎪ ⎧ ⎡ ⎤
⎪
⎪
⎪
⎪ ⎪
⎪
⎨ ⎪
⎨⎢ ⎥ (3.9)
⎢ 1 L
⎥
⎪ al (t) min ⎢ (Ctotal − Cc ai (t))⎥ ,
⎪
⎪ ⎪
⎪ ⎣ L ⎦
⎪
⎪ ⎪
⎩ a (t) i=1
⎪
⎪ i
⎪
⎪ ⎛ i=1 ⎞⎫
⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ ⎜ [ll] 2 [l] ⎟ ⎪
⎬
⎪
⎪ ⎜ h (t) P xl ⎟
⎪
⎪ log ⎜ 1 + ⎟
⎪
⎪ 2
⎝ 2 ⎠⎪
⎪
L
⎪
⎪
⎪
⎪ aj (t) h[lj ] (t) P [j ] xj + σ 2 ⎭⎪
⎪
⎪ j =1,j =l
⎪
⎪
⎪
⎪
⎪
⎩
if cl (t) = 0,

L
where ai (t) ≥ 3 is a condition for IA network, Ctotal is the total capacity of
i=1
the backhaul link, and Cc is the reserved capacity allocated to each active user to
share CSI with other active users. For the lth candidate, if the requested content is
not in the local cache, it can only acquire the content through the backhaul link,
and equal capacity (the total capacity minus the total capacity for CSI exchange) is
allocated among the active users. If the requested content is within the cache, the
lth candidate can get the maximum rate that an IA user can achieve. We assume the
realistic IA situations that the interference cannot be perfectly eliminated, i.e., the
interference leakage from other non-direct channels remains. Thus, the reward of
the lth candidate is determined by the integrated system’s state, including the states
of both the direct and non-direct links.
The immediate system reward is the sum of all the candidates’ immediate
L
rewards, i.e., r(t) = rn (t). The central scheduler gets r(t) in state x(t) when
n=1
action a(t) is performed in time slot t. The goal of using deep Q network into our
system model is to find a selection policy that maximizes the discounted cumulative
rewards during the communication period T , and the cumulative reward can be
expressed as
T −1

R = max E t
r(t) , (3.10)
π
t=0
where t approaches to zero when t is large enough. In practice, a threshold for

terminating the process can be set.
The Effects of Imperfect CSI
In (3.7) and (3.9), we assume perfect CSI. However, in practical networks, the
implementation of IA requires the CSI exchange among transmitters on the
backhaul link, which will cause the delay of CSI. The limited capacity of backhaul
link can have dramatic effects on the performance of IA due to the delayed CSI.
Suppose that H(t) is the matrix consisting of the accurate channel coefficients at
time instant t. The delayed channel matrix is denoted as follows.
H(t − τ ) is the delayed channel matrix when CSI is delayed by τ duration. The
relation between the delayed CSI and the current CSI can be modeled as

H(t) = ρH(t − τ ) + +
1 − ρ 2 δ = ρ H(t) 1 − ρ 2 δ, (3.11)
is the matrix of the delayed channel coefficients at the time instant t. δ

where H(t)

has the same distribution with H(t) and H(t). ρ is the normalized autocorrelation
function of a fading channel with motion at a constant velocity, and 0 ≤ ρ ≤ 1. It
can be seen that ρ = 1 corresponds to perfect CSI, whereas ρ = 0 represents no
CSI. ρ is defined as follows.
∗!
E (H)ij H
ij
ρ=" # . (3.12)
2 ! 2 $

E (H)ij E H ij
The value of ρ depends on the time scale of channel variation, which can be defined
by coherence time [28], and length of the delay τ . When the channel is under
Rayleigh fading, ρ can be derived as follows [29]
ρ = J0 (2πfd τ ), (3.13)
where J0 (.) is the zeroth order Bessel function of the first kind, and fd is the Doppler
frequency, which reflects the velocity of the transceivers.
Derivation of u and v
Here we leverage the conventional iterative interference alignment[26] to derive

the precoding vector v and decoding vector u for each user, which utilizes the
reciprocity of wireless channels and aims at minimizing the total interference
leakage at receivers.
In the forward direction of iterations, the total interference leakage at the lth
receiver caused by other users can be written as
!
I [l] = Tr u[l]† Q[l] u[l] , (3.14)
where Tr[A] denotes the trace operator of matrix A, and
L
Q[l]= P [i] H[li] v[i] v[i]† H[li]† . (3.15)
i=1,i=l
In the reverse direction of iterations, the total interference leakage at the ith
receiver (i.e., the ith transmitter in the original direction) can be denoted as
− −[i] !
←
−
I [i] = Tr ←−[i]† ←
u Q [i] ←
u , (3.16)
where
←
−[i]
L
←− −[l] ← ←−
Q = P [l] H [il] ←
v v−[l]† H [il]† . (3.17)
l=1,l=i
In (3.16) and (3.17), ← −[l] = v[l] , and ←

v−[l] = u[l] , ←
u
−
H [li] = H[il]† .
3.4 Simulation Results and Discussions 31
In the original channel, the decoding vector u[l] of the lth user, which can
minimize the total interference leakage, can be updated as

u[l] = vi Q[l] . (3.18)
where vi (X) means the eigenvector corresponding to the ith smallest eigenvalue of
matrix X.
In the reverse channel, the precoding vector of the lth user is set as ←
v−[l] = u[l] .
←
− [l]
The reverse decoding vector u can be obtained as
−
←−[l] = v ←
u [l]
i Q . (3.19)
Then the reverse decoding vector ← −[l] is regarded as the precoding vector in
u
←
−
the original channel, i.e., v[l] = u [l] . This process iterates in such a way until
convergence, and the process is summarized in Algorithm 2.
3.4 Simulation Results and Discussions
Computer simulations are carried out to demonstrate the performance of the pro-
posed deep reinforcement learning approach to the optimization of cache-enabled
opportunistic IA wireless networks. We use TensorFlow [30] in our simulations
Algorithm 2 The conventional iterative interference alignment

Initialize the Nt[l] × 1 precoding vector v[l] arbitrarily.
Execute the iteration
Calculate the interference covariance matrix Q[l] for the lth receiver based on (3.15), l =
1, 2, . . . , L.
Compute the decoding vector u[l] for the lth receiver based on (3.18), l = 1, 2, . . . , L.
Reverse the communication direction. Set the reversed precoding vector as the original decoding
vector, i.e., ←
v−[l] = u[l] , l = 1, 2, . . . , L.
←
−
For the reversed communication direction, calculate matrix Q [k] for the new lth receiver based
on (3.17), l = 1, 2, . . . , L.
Compute the reversed decoding vector ←

−[l] for the new lth receiver based on (3.19), l =
u
1, 2, . . . , L.
Reverse the communication direction. Set the original precoding vector as the reversed decoding
vector, i.e., v[l] = ←
−[l] , l = 1, 2, . . . , L.
u
Continue till Convergence
Obtain the Interference alignment solutions as v[l] and u[l] , l = 1, 2, . . . , L.

to implement deep reinforcement learning. In this section, we first introduce

TensorFlow, followed by simulation settings. Then, simulation results are discussed.
3.4.1 TensorFlow
TensorFlow is an open source software library for expressing machine learning

algorithms, and an implementation for executing such algorithms. TensorFlow has
attracted great interests from both academia and industry for a variety of applica-
tions, such as speech recognition, Gmail, Google Photos, and search [30]. Originally
developed by the Google Brain team for Google’s research and production purposes,
TensorFlow was later released under the Apache 2.0 open source license in 2015.
TensorFlow is Google Brain’s second generation machine learning system to replace
its predecessor DistBelief for the implementation and deployment of large-scale
machine learning models.
TensorFlow provides a Python API, as well as a less documented C++ API.
The reference implementation of TensorFlow runs on single devices. Nevertheless,
TensorFlow can run on multiple CPUs and GPUs for fast execution. A wide variety
of hardware platforms can be used for TensorFlow, which takes computations and
maps them onto different hardware platforms, ranging from mobile device platforms
(e.g., Android and iOS) to modest-sized systems using single machines containing
one or many GPU cards to large-scale systems running hundreds and thousands of
GPUs.
TensorFlow is used to transform an ordinary Q-Network into a deep Q-Network
(DQN) by making the following improvements:
(1) Extending a simple neural network to a multi-layer convolutional net-
work. We utilize the tf.contrib.layers.convolution2d function to eas-
ily create a convolutional layer as follows: convolution_layer =
tf.contrib.layers.convolution2d (in, num_out, k_size, stride, padding),
where num_out is the number of filters applied to the previous layer, k_size
refers to how large a window used to slide over the previous layer, stride is the
number of points skipped as we slide the window across the layer, and padding
indicates if the window just slides over the bottom layer or adds padding around
it to ensure that the convolutional layer has the same dimensions as the previous
layer. Please refer to the Tensorflow documentation [31] for more information.
(2) Implementing Experience Replay, which will allow our network to train itself
using the stored memories from its experience. By keeping the experiences, the
network can learn from a more varied sets of past experiences. We use a tuple
of < state, action, reward, nextstate > to store these experiences. In our
DQN, a class is used to handle storing and retrieving memories.
(3) Utilizing a second “target” network, which is used to compute the target Q-
values during the training procedure. The reason why two networks are used is
as follows. The Q-networks values shift at every step of training, and the value
estimations can easily spiral out of control, if we adjust our network values by
using a constantly shifting set of values. Consequently, there will be feedback
loops between the target and estimated Q-values, and the network can become
destabilized. To address this issue, we fix the target networks weights, and they
are periodically updated to the primary Q-networks values. By doing this, we
can have a training process in a more stable manner.
3.4.2 Simulation Settings
In our simulations, we used a GPU-based server, which has 4 Nvidia GPUs with
version GTX TITAN. The CPU is Intel Xeno E5-2683 v3 with 128G memory. The
software environment we utilized is TensorFlow 0.12.1 with Python 2.7 on Ubuntu
14.04 LTS.
For performance comparison, our proposed OIA scheme is compared with four
other schemes as
1. The same proposed scheme without (w.o.) caching.
2. An existing selection scheme without (w.o.) caching [32], in which invariant
channels are assumed, i.e., the estimated channel coefficient of the current time
instant is simply taken as the predicted channel coefficient of the next time
instant.
3. An existing selection scheme with (w.) caching [12], in which invariant channels
are assumed, and it schedules the transceiver pairs to maximize the network
throughput.
4. An existing scheme with (w.) caching [33], in which invariant channels are
assumed, and power allocation strategy is performed aiming at maximizing the
network throughput, however, no user selection is included.
In the simulations, we consider a cache-enabled opportunistic IA network, in
which L = 5 candidates want to access to. Due to the feasibility of IA[34], i.e.,
Nt + Nr ≥ d(L + 1). Here, we set DoF d to be 1, and assume that each candidate
is equipped with three antennas at both the transmitter node and the receiver node,
i.e. Nt = Nr = 3. Other setup parameters are: the bandwidth B = 10 MHz per
transmitter, the total backhaul capacity Ctotal = 60 Mb/s, and the reserved capacity
for sharing CSI Cc = 2 Mb/s per active user. The noise power σ 2 is set to be 0.1 mW
throughout the simulation. The normalized autocorrelation value ρ is set to be 0.99
in this chapter.
Based on the definition of system states, we quantize and uniformly partition
the channel coefficients |h[kj ] |2 into 10 levels with 9 boundary values as 10−6 , 0.3,
0.6, 0.9, 1.2, 1.5, 1.8, 2.1, 2.4. Uniformly setting the boundary values of FSMC is
widely used in the literature (e.g., [35]) for simplicity. Nevertheless, the boundary
values for a particular environment under certain criterion should be optimized for
better performance [35]. We assume that the channel state transition probability is
identical for all the candidates. In one simulation scenario, the transition probability
.....
.....
.....
.....
.....
Fig. 3.2 Visualization of the proposed deep reinforcement learning algorithm using TensorBoard
of remaining in the same state is set to be 0.489, and the probability of transition to
the adjacent state to be twice that of transition to a nonadjacent state. The channel
state transition matrix is shown on the top of the next page. We change the channel
state transition probability in other simulation scenarios.
⎛ ⎞
0.489 0.256 0.128 0.064 0.032 0.016 0.008 0.004 0.002 0.001
⎜ 0.001 0.489 0.256 0.128 0.064 0.032 0.016 0.008 0.004 0.002 ⎟
⎜ ⎟
⎜ 0.002 0.001 0.489 0.256 0.128 0.064 0.032 0.016 0.008 0.004 ⎟
⎜ ⎟
⎜ 0.004 0.002 0.001 0.016 0.008 ⎟
⎜ 0.489 0.256 0.128 0.064 0.032 ⎟
⎜ ⎟
⎜ 0.008 0.004 0.002 0.001 0.489 0.256 0.128 0.064 0.032 0.016 ⎟
Pchannel =⎜ ⎟.
⎜ 0.016 0.008 0.004 0.002 0.001 0.489 0.256 0.128 0.064 0.032 ⎟
⎜ ⎟
⎜ 0.032 0.016 0.008 0.004 0.002 0.001 0.489 0.256 0.128 0.064 ⎟
⎜ ⎟
⎜ 0.064 0.032 0.016 0.008 0.004 0.002 0.001 0.489 0.256 0.128 ⎟
⎜ ⎟
⎝ 0.128 0.064 0.032 0.016 0.008 0.004 0.002 0.001 0.489 0.256 ⎠
0.256 0.128 0.064 0.032 0.016 0.008 0.004 0.002 0.001 0.489
The cache at each transmitter includes two states: existence and nonexistence of
the requested content. The cache state transition probability matrix is set to be
% &
0.6 0.4
Pcache = .
0.4 0.6
Table 3.1 Parameter values used in the simulations

Parameter Value Description
Mini-batch size 8 How many experience cases are used for each training
step
Update frequency 4 The frequency to perform a training step
Experience replay buffer 50,000 Training cases are randomly sampled from this number
size of the most recently experiences
Pre-training steps 10,000 How many steps of random actions are executed before
the leaning begins, and the resulting experience are
stored to populate the experience replay buffer
Total training steps 500,000 How many steps are used to train the network model
Discount factor 0.99 Discount factor used on the Q-function
Learning rate 0.0001 The learning rate used by AdamOptimizer
Initial exploration 1 The starting chance of random action in the ε-greedy
exploration
Final exploration 0.1 The final chance of random action in the ε-greedy
exploration
Anneling steps 10,000 How many training steps used to reduce ε from its
starting value to the final value
Target network update 0.001 The rate to update target Q-network towards primary
rate Q-network
The detailed parameters in the proposed deep reinforcement learning algorithm

are listed in Table 3.1. The visualization of the deep Q network model is presented
in Fig. 3.2 using TensorBoard, which is a build-in module of TensorFlow. From the
graph, it can be clearly visualized that double deep Q-networks are utilized, four
convolutional layers are adopted in each deep Q-network, and the advantage and
value functions are separately computed. The network models are saved and can be
loaded to further train or test the models. In the simulations, some parameters are
changed to study the effects of these parameters.
3.4.3 Simulation Results and Discussions
Figure 3.3 shows the convergence performance of the proposed scheme with
different learning rates in the deep reinforcement learning algorithm. As we can
see from the expression of the system state, the number of the possible states is
2 2
2L ×H L , where 2L is for the cache status and H L is for the status of L2 equivalent
channel coefficients |h[kj ] (t)|2 with k, j = 1, 2, . . . , L. In addition to the number of
possible states, the complexity of the deep Q learning algorithm depends on many
other factors, including the number of actions, the state transition probability, and
the rewards. Moreover, since deep Q learning utilizes deep learning to approximate
the Q function, the convergence time is affected by many elements, such as the
number of convolutional layers, the learning rate, the batch size, and some other
parameters in the training process. From Fig. 3.3, we can observe that the average
sum rate of the proposed scheme is very low at the beginning of the learning process.
With the increase of the number of the episode, the average sum rate increases
until it reaches a relatively stable value, which is around 300 Mbps in Fig. 3.3.
This shows the convergence performance of the proposed scheme. We can also
observe that the learning rate in the AdamOptimizer has effects on the convergence
performance in Fig. 3.3. Specifically, the convergence is faster when the learning
rate is 0.0001 compared to the case when the learning rate is 0.00001. However,
a larger learning rate will result in local optimum instead of global optimum.
Therefore, an appropriate learning rate should be chosen for a specific problem.
In the rest of the simulations, we choose the learning rate of 0.0001.
Figure 3.4 shows the effects of the mini-batch size for each gradient update in the
deep reinforcement learning algorithm on the convergence performance. The param-
eter of mini-batch size determines how many experience cases are used for each
training step. We can see from Fig. 3.4 that the convergence is faster when the mini-
batch size is 8 compared to the cases when the mini-batch size is 32 and 64. Similar
to the learning rate parameter, an appropriate mini-batch size should be chosen for a
specific problem. In the rest of the simulations, we choose the mini-batch size of 8.
Figure 3.5 shows the network’s average sum rate with different state-transition
probabilities of staying in the same state. Our proposed schemes with (w.) and
without (w.o.) cache are compared with an existing selection scheme w. cache
[12] and an existing selection scheme w.o. cache [32], where invariant channels are
310
300
290
280
Sum Rate (Mbps)
Learning rate 0.0001

Learning rate 0.00001
270
260
250
240
230
220
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Episode × 104
Fig. 3.3 Convergence performance with different learning rates

310
300
290
280
Sum Rate (Mbps)
270
Mini-batch size 8
260 Mini-batch size 32
Mini-batch size 64
250
240
230
220
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
4
Episode ×10
Fig. 3.4 Convergence performance with different batch sizes
assumed for both existing schemes. It can be seen that the proposed OIA w. cache
scheme can achieve the highest sum rate compared to the other three schemes. This
is because the channel is time-varying, and the proposed scheme can obtain the
optimal IA user selection policy in the realistic time-varying channel environment
using the deep reinforcement learning algorithm. For both the cases w. cache and
w.o. cache, we can observe that the performance of the existing selection method
is getting closer to the proposed OIA as the transition probability increases, and
this method performs the same when the channel remains absolutely static, i.e., the
transition probability that the channel will be in the same state is 1.
Figure 3.6 shows the network’s average sum rate with different average SNR
values, where the SNR is defined as 10log10 (P [k] /σ 2 ) dB, and the same SNR
definition applies to other figures in this chapter. From this figure, we can observe
that the average sum rate of the proposed OIA scheme w. cache can outperform the
other four schemes with different average SNR values. This is because the existing
selection scheme assumes time-invariant channels, and the estimated CSI of the
current time instant is simply taken as the predicted CSI of the next time instant. In
addition, the proposed OIA scheme w.o. cache does not take advantages of caching
to relieve the traffic loads of backhaul links, thus has worse performance compared
to the proposed scheme w. cache. Note that the existing user selection scheme w.
cache performs similarly with the existing scheme w.o. user selection due to the
fact that they both considers caching and aims at maximizing the network sum
rate, however, one exploits power allocation strategy, and the other one utilizes user
selection strategy.
350
Proposed OIA scheme w. cache
Existing user selection scheme w. cache
Proposed OIA scheme w.o. cache
Existing user selection scheme w.o. cache
300
Sum Rate (Mbps)
250
200
150
100
0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
Transition Probability
Fig. 3.5 Comparison of the average sum rate with different transition probabilities
Figure 3.7 shows the effects of different ρ values. The average SNR is 30 dB.
As defined in (3.12), ρ is the normalized autocorrelation function of a fading
channel with motion at a constant velocity. From Fig. 3.7, we can see that the
network’s average sum rate increases with the increase of ρ in different schemes.
This is because a higher ρ value means more accurate CSI, which will result in
higher average sum rate in the network. In comparison, the proposed OIA with
cache scheme has better performance compared to two other schemes, because
the proposed scheme considers the time-varying channels and takes advantage of
caching. Figure 3.8 shows the effects of different ρ values when the average SNR
is 10 dB. We observe the same trend in Fig. 3.8 with lower average sum rate due to
the low SNR value compared to Fig. 3.7.
Figure 3.9 shows the effects of different values of the total backhaul capacity. The
average SNR is 30 dB. From Fig. 3.9, we can see that the network’s average sum rate
increases with the increase of total backhaul capacity in different schemes. This is
because a higher value of the total backhaul capacity will provide more capacity for
CSI exchange among the transmitters in interference alignment wireless networks,
which will result in higher average sum rate in the network. Nevertheless, the effect
of the total backhaul capacity is not very significant in the proposed scheme w.
cache, because the proposed scheme takes advantages of caching, and the more
backhaul capacity is saved for CSI exchange. Figure 3.10 shows the effects of
350
Existing scheme w.o. user selection
Existing user selection scheme w. cache
300 Proposed OIA scheme w.o. cache
250
Sum Rate (Mbps)
200
150
100
50
0
0 5 10 15 20 25 30
SNR (dB)
Fig. 3.6 Comparison of the average sum rate with different SNR values
different values of the total backhaul capacity when the average SNR is 10 dB. We
can see the same trend in Fig. 3.10 with lower average sum rate due to the low SNR
value compared to Fig. 3.9.
Figure 3.11 shows the energy efficiency of proposed scheme and other compared
schemes. Except the existing w.o. user selection scheme [33] that exploits power
allocation, all the other four schemes are of equal power allocation. The power
consumption of circuit blocks is set to be 210 mW. The transmitted power of one
user is 100 mW, where for the power allocation scheme, the total power of the
network is 500 mW for a 5-user IA network. The energy efficiency of the power
amplifier is set to 90%. From this figure, we can observe that the existing scheme
w.o. user selection that uses power allocation technique performs better than other
existing schemes due to effective power allocation. Nevertheless, it still performs
worse than the proposed scheme due to the underlying assumption of time-invariant
channels used in the existing scheme w.o. user selection that uses power allocation
technique.
350
Proposed OIA with cache scheme

300 Proposed OIA without cache scheme
Existing selection scheme without cache
250
SumRate(Mbps)
200
150
100
50
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
ρ
Fig. 3.7 Comparison of the average sum rate with different ρ values (average SNR = 30 dB)
150
Proposed OIA with cache scheme

Proposed OIA without cache scheme
Existing selection scheme without cache
100
Sum Rate (Mbps)
50
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
ρ
Fig. 3.8 Comparison of the average sum rate with different ρ values (average SNR = 10 dB)
500
450 Existing user selection scheme w. cache
400
350
Sum Rate (Mbps)
300
250
200
150
100
50
40 45 50 55 60 65 70 75 80 85 90
C (Mbps)
total
Fig. 3.9 Comparison of the average sum rate with different values of the total backhaul capacity
(average SNR = 30 dB)
200
ProposedOIA scheme w. cache
Existingscheme w.o. user selection
180 Existing user selectionscheme w. cache
ProposedOIA scheme w.o. cache
Existing user selectionscheme w.o. cache
160
140
Sum Rate (Mbps)
120
100
80
60
40
20
40 45 50 55 60 65 70 75 80 85 90
C (Mbps)
total
Fig. 3.10 Comparison of the average sum rate with different values of the total backhaul capacity
(average SNR = 10 dB)
40
35 Existing user selection scheme w. cache
30
Energy Efficiency(bits/Hz/Joule)
25
20
15
10
0
0 5 10 15 20 25 30
SNR(dB)
Fig. 3.11 Energy efficiency comparison with different SNR values
3.5 Conclusions and Future Work
In this chapter, we studied cache-enabled opportunistic IA under the condition of

time-varying channel coefficients. The introduction of caching can save the limited
backhaul capacity that can be utilized for CSI exchange among the transmitters
in interference alignment wireless networks. The system complexity is very high
when we model the time-varying channel as a finite-state Markov channel. Thus,
we exploited the recent advances, and formulated the optimization of the cache-
enabled opportunistic interference alignment network as a deep reinforcement
learning problem. A central scheduler is responsible for collecting the CSI from
each candidate, and then sends the integral system state to the deep Q network
to derive the optimal policy for user selection. Simulation results were presented
to show that deep reinforcement learning is an effective approach to solve the
optimization problem in cache-enabled opportunistic IA wireless networks. It was
demonstrated that the performance of cache-enabled opportunistic IA networks
can be significantly improved by using the proposed deep reinforcement learning
approach. Nevertheless, some parameters, such as learning rate and mini-batch size,
should be carefully chosen in the algorithm. Future work is in progress to consider
wireless network virtualization in the proposed framework to further improve the
network performance.
References 43
References
1. G. Paschos, E. Bastug, I. Land, G. Caire, and M. Debbah, “Wireless caching: technical

misconceptions and business barriers,” IEEE Comm. Mag., vol. 54, no. 8, pp. 16–22, Aug.
2016.
2. C. Liang, F. R. Yu, and X. Zhang, “Information-centric network function virtualization over
5G mobile wireless networks,” IEEE Network, vol. 29, no. 3, pp. 68–74, May 2015.
3. X. Wang, M. Chen, T. Taleb, A. Ksentini, and V. C. M. Leung, “Cache in the air: exploiting
content caching and delivery techniques for 5G systems,” IEEE Commun. Mag., vol. 52, no. 2,
pp. 131–139, Feb. 2014.
4. C. Fang, F. R. Yu, T. Huang, J. Liu, and Y. Liu, “A survey of green information-centric
networking: Research issues and challenges,” IEEE Comm. Surveys Tutorials, vol. 17, no. 3,
pp. 1455–1472, Thirdquarter 2015.
5. D. Liu, B. Chen, C. Yang, and A. F. Molisch, “Caching at the wireless edge: design aspects,
challenges, and future directions,” IEEE Commun. Mag., vol. 54, no. 9, pp. 22–28, 2016.
6. V. R. Cadambe and S. A. Jafar, “Interference alignment and degrees of freedom of the K-user
interference channel,” IEEE Trans. Inform. Theory, vol. 54, no. 8, pp. 3425–3441, Aug. 2008.
7. C. Suh and D. Tse, “Interference alignment for cellular networks,” in Proc. 46th Annual
Allerton Conf. on Commun., Control, and Computing,. Monticello, IL, Sep. 2008, pp. 1037–
1044.
8. S. M. Perlaza, N. Fawaz, S. Lasaulce, and M. Debbah, “From spectrum pooling to space
pooling: opportunistic interference alignment in MIMO cognitive networks,” IEEE Trans.
Signal Proc., vol. 58, no. 7, pp. 3728–3741, 2010.
9. X. Li, N. Zhao, Y. Sun, and F. R. Yu, “Interference alignment based on antenna selection
with imperfect channel state information in cognitive radio networks,” IEEE Trans. Veh. Tech.,
vol. 65, no. 7, pp. 5497–5511, July 2016.
10. B. C. Jung and W.-Y. Shin, “Opportunistic interference alignment for interference-limited
cellular TDD uplink,” IEEE Commun. Lett., vol. 15, no. 2, pp. 148–150, 2011.
11. Y. He, H. Yin, and N. Zhao, “Multiuser-diversity-based interference alignment in cognitive
radio networks,” AEU-Int. J. Electron. C, vol. 70, no. 5, pp. 617–628, 2016.
12. M. Deghel, E. Baştuğ, M. Assaad, and M. Debbah, “On the benefits of edge caching for MIMO
interference alignment,” in Proc. IEEE SPAWC, 2015, pp. 655–659.
13. M. A. Maddah-Ali and U. Niesen, “Cache-aided interference channels,” in Proc. IEEE ISIT,
2015, pp. 809–813.
14. O. E. Ayach, S. W. Peters, and R. W. Heath, “The practical challenges of interference
alignment,” IEEE Wirel. Commun., vol. 20, no. 1, pp. 35–42, Feb. 2013.
15. J. Yang, A. K. Khandani, and N. Tin, “Statistical decision making in adaptive modulation and
coding for 3G wireless systems,” IEEE Trans. Veh. Technol., vol. 54, no. 6, pp. 2066–2073,
2005.
16. A. Z. Ghanavati, U. Pareek, S. Muhaidat, and D. Lee, “On the performance of imperfect
channel estimation for vehicular ad-hoc networks,” in Proc. IEEE VTC’10Fall, Sept. 2010,
pp. 1–5.
17. R. Xie, F. R. Yu, and H. Ji, “Dynamic resource allocation for heterogeneous services in
cognitive radio networks with imperfect channel sensing,” IEEE Trans. Veh. Tech., vol. 61,
pp. 770–780, Feb. 2012.
18. Y. Cai, F. R. Yu, C. Liang, B. Sun, and Q. Yan, “Software defined device-to-device (D2D)
communications in virtual wireless networks with imperfect network state information (NSI),”
IEEE Trans. Veh. Tech., no. 9, pp. 7349–7360, Sept. 2016.
19. Y. He, F. R. Yu, N. Zhao, H. Yin, H. Yao, and R. C. Qiu, “Big data analytics in mobile cellular
networks,” IEEE Access, vol. 4, pp. 1985–1996, Mar. 2016.
20. Y. Wei, F. R. Yu, and M. Song, “Distributed optimal relay selection in wireless cooperative
networks with finite-state Markov channels,” IEEE Trans. Veh. Technol., vol. 59, no. 5,
pp. 2149–2158, 2010.
21. H. S. Wang and P.-C. Chang, “On verifying the first-order Markovian assumption for a
Rayleigh fading channel model,” IEEE Trans. Veh. Technol., vol. 45, no. 2, pp. 353–357, 1996.
22. C. Luo, F. R. Yu, H. Ji, and V. C. M. Leung, “Cross-layer design for TCP performance
improvement in cognitive radio networks,” IEEE Trans. Veh. Tech., vol. 59, no. 5, pp. 2485–
2495, June 2010.
23. V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves,
M. Riedmiller, A. K. Fidjeland, G. Ostrovski et al., “Human-level control through deep
reinforcement learning,” Nature, vol. 518, no. 7540, pp. 529–533, 2015.
24. Y. He, C. Liang, F. R. Yu, N. Zhao, and H. Yin, “Optimization of cache-enabled opportunistic
interference alignment wireless networks: A big data deep reinforcement learning approach,”
in Proc. IEEE ICC’17, Paris, France, June 2017.
25. Y. He, F. R. Yu, N. Zhao, V. C. M. Leung, and H. Yin, “Software-defined networks with
mobile edge computing and caching for smart cities: A big data deep reinforcement learning
approach,” IEEE Commun. Mag., vol. 55, no. 12, Dec. 2017.
26. K. Gomadam, V. R. Cadambe, and S. A. Jafar, “A distributed numerical approach to inter-
ference alignment and applications to wireless interference networks,” IEEE Trans. Inform.
Theory, vol. 57, no. 6, pp. 3309–3322, 2011.
27. A. Tatar, M. D. de Amorim, S. Fdida, and P. Antoniadis, “A survey on predicting the popularity
of web content,” Springer J. Internet Services and Applications, vol. 5, no. 1, p. 8, 2014.
28. D. Tse and P. Viswanath, Fundamentals of Wireless Communication. Cambridge, U.K.:
Cambridge Univ. Press, 2005.
29. R. H. Clarke, “A statistical theory of mobile radio reception,” Bell Syst. Tech. J., vol. 47, no. 6,
pp. 957–1000, Jul.-Aug. 1968.
30. M. Abadi, A. Agarwal et al., “Tensorflow: Large-scale machine learning on heterogeneous
systems,” arXiv:1603.04467, Nov. 2015.
31. “Tensorflow.org,” https://www.tensorflow.org/.
32. N. Zhao, F. R. Yu, H. Sun, and M. Li, “Adaptive power allocation schemes for spectrum
sharing in interference alignment (IA)-based cognitive radio networks,” IEEE Trans. Veh.
Tech., vol. 65, no. 5, pp. 3700–3714, May 2016.
33. F. Cheng, Y. Yu, Z. Zhao, N. Zhao, Y. Chen, and H. Lin, “Power allocation for cache-aided
small-cell networks with limited backhaul,” IEEE Access, vol. 5, pp. 1272–1283, 2017.
34. C. M. Yetis, T. Gou, S. A. Jafar, and A. H. Kayran, “On feasibility of interference alignment in
MIMO interference networks,” IEEE Trans. Signal Proc., vol. 58, no. 9, pp. 4771–4782, 2010.
35. H. S. Wang and N. Moayeri, “Finite-state Markov channel-a useful model for radio communi-
cation channels,” IEEE Trans. Veh. Tech., vol. 44, no. 1, pp. 163–171, Feb. 1995.
Chapter 4
Deep Reinforcement Learning for Mobile
Social Networks
Abstract Social networks have continuously been expanding and trying to be

innovative. The recent advances of computing, caching, and communication (3C)
can have significant impacts on mobile social networks (MSNs). MSNs can leverage
these new paradigms to provide a new mechanism for users to share resources (e.g.,
information, computation-based services). In this chapter, we exploit the intrinsic
nature of social networks, i.e., the trust formed through social relationships among
users, to enable users to share resources under the framework of 3C. Specifically,
we consider the mobile edge computing (MEC), in-network caching and device-
to-device (D2D) communications. When considering the trust-based MSNs with
MEC, caching and D2D, we apply a novel deep reinforcement learning approach
to automatically make a decision for optimally allocating the network resources.
The decision is made purely through observing the network’s states, rather than any
handcrafted or explicit control rules, which makes it adaptive to variable network
conditions. Google TensorFlow is used to implement the proposed deep Q-learning
approach. Simulation results with different network parameters are presented to
show the effectiveness of the proposed scheme.
4.1 Introduction
Mobile social networks (MSNs) have been developing rapidly, which can provide
a variety of social services and applications to mobile users [1]. Millions of mobile
users interact with each other in MSNs, and MSNs will become one of the most
important networking paradigms in future wireless mobile networks [2]. MSNs have
always been trying to be innovative and leverage new technologies. The recently
proposed integrated framework of computing, caching and communication (3C)
can have positive impacts on MSNs. The integration of the 3C framework and
MSNs can help create a new mechanism to share resources among users. The
available resources that can be shared extend beyond information, for instance, the
storage devices and computing capability contributed by users can also be shared.
Additionally, the MSNs have an inherent advantage, i.e., the trust of mobile users
is naturally formed through the social relationships and interactions. This makes

46 4 Deep Reinforcement Learning for Mobile Social Networks
the resource sharing between users more reliable. In this chapter, we consider the
specific 3C framework with mobile edge computing (MEC), in-network caching,
and device-to-device (D2D) communications.
With MEC, computation resources are placed at the edge of wireless mobile
networks in physical proximity to mobile users [3]. Compared to traditional mobile
cloud computing, MEC can provide faster interactive response by low-latency
connections. Therefore, MEC has been envisioned as a promising technique to offer
agile and ubiquitous computation augmenting mobile services and applications,
including social services and applications [4, 5]. Another promising technology is
in-network caching, which can effectively reduce the duplicate content transmission
in the network. Recent studies of applying the in-network caching technique in
MSNs show that traffic loads and latency can be significantly reduced by caching
contents in MSNs [2]. In addition, D2D communications can be beneficial to MSNs
as well [6]. With D2D communications, users in close proximity can directly
communicate with each other via D2D links, instead of accessing base stations
(BSs) exclusively. When it comes to the content-centric MSNs, in spite of the
smaller-sized storage (compared to that of BSs), the ubiquitous caching capability
residing in mobile devices cannot be neglected due to their ubiquitous in-network
distribution and ever-increasing storage size [6, 7].
Although some works have been done on applying recent advances of MEC, in-
network caching and D2D to improve the performance of MSNs, the knowledge
of social relationships among users in MSNs is largely ignored in these new
paradigms to improve the reliability and efficiency of resource sharing and delivery
in MSNs. In fact, social relationships have been investigated to enhance wireless
networking. For example, Zhang et al. [8] exploit social ties to achieve a reliable
D2D communication, which can improve packet transmission and reduce the
workload on the network infrastructures. Pan et al. [9] utilize social information
as an essential element for designing forwarding algorithms in ad hoc networks.
Social relationships also play an important role for routing in wireless environments
[10].
Furthermore, the dynamic nature of the available resources has not been paid
much attention in the existing literature. To fill in this gap, in this chapter, we
study trust-based social networks with recent advances of MEC, caching and D2D.
Considering the integrated network, the allocation of resources for subscribed users
is complicated, especially when the conditions of the network resources are varying
with time. Therefore, we utilized a novel deep reinforcement learning approach to
automatically achieve the resource allocation tasks.
In the following, we first review some of the existing excellent works that
focus on efficient management and allocation schemes of computing, caching and
communication resources. Some of the proposed schemes only consider one of these
three types of resources. On the other hand, in some other works, two or three types
of resources are jointly considered to effectively and efficiently achieve optimal
performance metrics. Then, we discuss our contributions in this chapter.
4.1 Introduction 47
4.1.1 Related Works
With the increasing popularity of compute intensive applications, such as augmented

reality, interactive video games, etc., MEC is becoming a very promising paradigm
that enables the possibility of offloading the computation tasks from resource-
limited mobile devices to more powerful edge servers. MEC can bring various
benefits, where minimizing energy consumption and minimizing latency are two
major optimization objectives [11]. In the literature, various resource allocation
approaches have been proposed to solve the formulated optimization problems.
For example, Zhang et al. [12] propose an energy-efficient computation offloading
scheme that jointly optimizes radio resource allocation and offloading to minimize
the energy consumption of the offloading system with the latency constraint. In their
scheme, the mobile devices are first classified into three types and then wireless
channels are allocated to mobile devices based on their priority iteratively. By doing
this, the optimization problem can be solved in polynomial complexity. Liu et
al. [13] design an optimal computation task scheduling policy for MEC systems.
They first analyze the average delay of each task and average power consumption
at the mobile device side under the proposed scheduling policy using Markov
chain theory. Then they formulate a delay minimization problem with the power
constraint. An efficient one-dimensional search algorithm is used to derive the
optimal offloading policy. In [14], the authors jointly optimize the radio resources
and computational resources to minimize the total energy expenditure, where an
iterative algorithm based on successive convex approximation technique is adopted
to solve the formulated non-convex optimization problem. In [15], a distributed
game theoretic approach is proposed to solve the NP-hard efficient computation
offloading problem. Chen et al. [16] choose a model-free RL technique to solve the
optimal traffic offloading strategy for heterogeneous cellular networks with dynamic
traffic.
When caching technology is integrated into mobile networks, where to cache,
what to cache, and how to cache are the major factors that make a significant
influence on the system performance [11]. In [17], the authors investigate the
proactive storage allocation problem over BSs in cellular networks, which is proved
to be NP-hard. To get a low-complexity solution, a heuristic method is utilized and
the convergence is proved by strict theoretic deduction. A caching-enabled D2D
communication scheme is designed in [18], where the caching strategy is optimized
by jointly considering users’ social relationships and common interests with the
constraint of hit ratio, delay and caching capacity. In fact, contents with the most
popularity should be cached within limited caching capacity. Since the content
popularity is varying with time, learning based caching policies are proposed in [19].
The authors formulate the distributed caching over small BSs (SBSs) as a reinforce-
ment learning problem, and with the help of coded caching, the complicated caching
problem is solved by being reduced to a convex optimization problem. In [20], Qiao
et al. consider the mobility pattern of mobile users, and formulate the mobility-
aware caching problem as an optimization problem to maximize the caching utility.
The problem is solved by using a polynomial-time heuristic solution. He et al.

exploit deep reinforcement learning to address the resource allocation problems in
cache-enabled interference alignment networks [21, 22].
The comprehensive resource allocation schemes for efficient integration of com-
puting, caching and communication resources to achieve the optimal performance
have been developed, even though not yet been widely investigated. In [23], Zhou et
al. design a novel information-centric heterogeneous network framework, and with
the virtualization technology, communication, computing and caching resources are
shared among users. They formulate the virtual resource allocation strategy as a
joint optimization problem, which is a non-convex NP hard problem. An alternating
direction method of multiplier (ADMM) approach is used to solve the problem
by simplifying and relaxing the non-convex problem into a convex one. ADMM
method is also used in [24] to achieve the optimal resource allocation strategy in
wireless networks with MEC and in-network caching being jointly considered. He
et al. propose an integrated framework that can enable dynamic orchestration of
networking, caching and computing resources, where the dynamics are modeled as
finite-state Markov chains [25, 26]. The authors use the deep Q-leaning approach
to solve the complex problems with large numbers of system states and actions.
Research on this topic of resource allocation strategies for integrated systems of
computing, caching and communication will continue with the evolving of each of
these three technologies.
4.1.2 Contributions
The main contributions of this chapter are summarized as follows.

• In this chapter, we consider trust-based social networks specifically with MEC,
in-network caching and D2D communications under the umbrella of a 3C
framework. An optimization problem is formulated to maximize the network
operator’s utility with comprehensive considerations of trust values, computation
capabilities, wireless channel qualities, and the cache status of all the available
BSs and D2D nodes.
• To be more realistic, we consider that the network conditions (i.e., trust value,
computation capability, wireless channels, and cache status) are varying with
time, and the dynamics of the computing capabilities, cache status and wireless
channel conditions are modeled as Markov chains. In the meanwhile, the trust
values are derived from both direct and indirect observations using Bayesian
inference and Dempster-Shafer theory, respectively.
• The complexity of the integrated network is very high when we jointly consider
the dynamic trust values, computation capabilities, wireless channel conditions
and the cache status, and it is extremely difficult to solve the formulated
optimization problem. In this chapter, we exploit deep Q-learning based resource
allocation strategy to solve the optimization problem without any explicit
assumptions or simplifications.
4.2 System Model 49
• Google TensorFlow is used to implement the proposed deep Q-learning

approach. Simulation results with different system parameters are presented
to show the effectiveness of the proposed scheme. It is illustrated that the
performance of MSNs can be significantly improved with the proposed approach.
The remainder of this chapter is outlined as follows. We describe the system
model in Sect. 4.2, which includes system description, network model, communi-
cation model, caching model and computing model. Next, the social trust scheme
with uncertain reasoning is presented in Sect. 4.3. Section 4.4 formulates this system
as a deep reinforcement learning problem. Simulation results are presented and
discussed in Sect. 4.5. Finally, we give the conclusions and future work in Sect. 4.6.
4.2 System Model
In this section, we first present the system description. Next, the network
model, communication model, cache model and computing model are presented,
respectively.
4.2.1 System Description
MSNs have been developed rapidly to provide a variety of social services and
applications to mobile users, focusing not only on the behaviors but also on the
social needs of the users [1]. Compared to conventional mobile wireless networks,
mobile users in MSNs do not always contact remote servers to request contents.
Instead, mobile users in MSNs can directly obtain contents from each other within
a community based on their social ties [8, 27].
In MSNs, huge amounts of information-rich data will be exchanged by mobile
users. Although cloud computing is powerful, it is difficult for data centers to
provide mobile users with low-latency services. To address these issues, MEC
has been proposed to deploy computing resources closer to end users, which
can efficiently improve the quality of service (QoS) for applications that require
intensive computations and low latency [4, 28]. MEC applies the concept of cloud
computing in network edge nodes, to facilitate services and applications, including
mobile social services and applications.
In addition, due to user mobility and poor-quality wireless radio links, it is
challenging to deliver huge amounts of data using the traditional client-server
approach in MSNs [2]. Recent advances of in-network caching can be extended
to MSNs to address this issue, which can efficiently reduce the duplicate content
transmission in networks. This innovative content-centric approach has been studied
in MSNs, which natively privilege the information (e.g., trusted information in a
specific proximity of an event and a specific time period) rather than the node
identity. In addition, with in-network caching, mobility and sporadic connectivity
issues can be effectively addressed in MSNs.
Fig. 4.1 A mobile social network with edge computing, caching and device-to-device (D2D)
communications
Moreover, with D2D communications, users in close proximity can directly

communicate with each other via D2D links, instead of accessing BSs exclusively.
As a promising approach to offload traffic from BSs, D2D communications can
enable the sharing of radio connectivity and direct information delivery between
two close users [29, 30].
Security is always an important aspect in wireless applications and services [31,
32]. Trust-based security schemes are important detection-based approaches in
MSNs. The definition of trust in MSNs is similar to that in sociology, where trust is
interpreted as degrees of the belief that an entity will carry out tasks as expected [33].
In this chapter, we present a framework for trust-based social networks specifically
with MEC, in-network caching and D2D communications, which is depicted in
Fig. 4.1. Under the framework, we illustrate a video content request task as an
example, which will be elaborated in the following network model.
4.2.2 Network Model
In this chapter, we consider the scenario with K BSs, and M mobile users that are
considered as transmitters in the potential D2D communications. The network is
4.2 System Model 51
operated by a central controller. Assume that there are L subscribed mobile users
issuing video requests to the network operator, which is responsible for assigning
a proper content provider to each requester, i.e., associating with one BS or setting
up a one-to-one D2D communication. The sets are denoted as CK = {1, 2, . . . , K},
CM = {1, 2, . . . , M} and L = {1, 2, . . . , L}, respectively. All the K BSs, the
M D2D transmitters and the L mobile users are equipped with both caching and
computing capabilities. The unicast D2D communication is available only when the
mobile user’s requested video is stored in the transmitter’s cache. In addition, the
transmitters can also provision mobile edge computing if they have spare computing
capacity left at the considered moment.
In our system model, we assume that the lth mobile user, denoted as sl , issues
a video request. First, the network controller checks all'the video providers, i.e.,
all the BSs and D2D transmitters. We denote C = CK CM be the set of video
providers, and let ci (1 ≤ i ≤ K) be the ith BS and cj (K + 1 ≤ j ≤ K + M) be the
j th transmitter. Each video provider’s cache is labeled by a content indicator which
indicates whether or not the requested video is being stored.
If the requested video is being stored at any of the video provider’s caches, and
the versions match up as well, the network controller will built up a communication
between mobile user sl and the optimal video provider. However, if all the stored
videos’ versions mismatch with the requested one, video transcoding should be
performed at either mobile user sl itself or at the video provider’s side. On the
other hand, If the requested video is not being stored at any of the caches, it has
to associate with a proper BS.
Note that in this work, we use video version to represent video specification
(e.g., H.263, H.264, MPEG2, or MPEG4). With the rapid growth of mobile services,
increasing volumes of videos are played by mobile devices. Consequently, service
providers often need to transcode the video contents into different specifications
(e.g., bit rate, resolution, quality, etc.) with different QoS (e.g., delay) for heteroge-
neous mobile devices, networks, and user preferences.
4.2.3 Communication Model

c
Denote hslp as the channel gain between mobile user sl and content provider cp , ∀l ∈
L , ∀p ∈ C . According to the Shannon theorem, the achievable rate of mobile user
sl when associating with provider cp can be expressed as
( c
)
c pT hslp
rslp = B log2 1+ , (4.1)
No B
where B denotes the bandwidth allocated to each mobile user, pT is the equally
transmitted power, and No is the noise spectral density.
c
Here, we consider the realistic wireless channels, and actually hslp is a continuous
random variable. Such assumption is intractable for analysis [34], and therefore
we model the wireless channels as the finite-state Markov channels (FSMC),

which may achieve performance improvement compared with the traditional
c
assumption of static channels. In our model, the channel gain hslp is discretized
c c
and quantized into H levels: H0 , if h∗0 hslp < h∗1 ; H1 , if h∗1 hslp < h∗2 ;
cp
. . . ; HH −1 , if hsl h∗H −1 . The boundary values for a particular environment
under certain criterion should be optimized for better performance [35]. However,
for simplicity, uniformly setting the boundary values of FSMC is widely used in
the literature (e.g., [35]). In this chapter, all the boundary values from h∗1 to h∗H −1
are increasing in the same distance with each other. Each level corresponds to a
state of the Markov chain, and therefore a H -element state space is formed. Due
c c c
to the relationship between hslp and rslp , for convenience we use rslp to describe a
state of the wireless channel. We consider the dynamic process, and the channel
c c
state realization of rslp at time instant t is denoted as Υsl p (t). Based on a certain
cp
transitional probability, Υsl (t) varies from one state to another as one time slot
c
is gone. The transitional probability of Υsl p (t) from one state js to another state
ks is denoted as ψjs ks (t). The H × H channel state transitional probability matrix
between mobile user sl and provider cp is shown as:
c * +
Ψsl p (t) = ψjs ks (t) H ×H , (4.2)

c c
where ψjs ks (t) = Pr Υsl p (t + 1) = ks | Υsl p (t) = js .
4.2.4 Cache Model
Assume that there are totally I video contents in the network that mobile users
can request for. The set of the contents is denoted as I = {1, 2, . . . , I }, which is
ranked by popularity. Here, we consider finite cache capacity, i.e., each provider
caches some of the I video contents, and refreshes the contents periodically. We
consider that mobile user sl has a video request vi , i ∈ I . After receiving the
request message, the network controller checks every provider’s content indicator
for video vi , i.e., xcvpi , ∀p ∈ C . The content distribution indicator xcvpi = 1 means
that video vi is being stored in the cache of provider p; otherwise xcvpi = 0. Here, xcvpi
is viewed as a random variable, representing the state of the cache, and is modeled
using a two-state (i.e., state 0 and 1) Markov chain [36]. The transitional probability
matrix of the state xcvpi is defined as:
Γcvpi (t) = [δas bs (t)]2×2 , (4.3)

where δas bs (t) = Pr xcvpi (t + 1) = bs | xcvpi (t) = as , and as , bs ∈ {0, 1}.
When the cache capacity is assumed to be finite, the transitional probability
matrix of the cache state can be derived based on different cache refreshment
strategies. An important one is the least recently used (LRU) cache refreshment
policy. The transitional probability matrix can be obtained using the following
4.2 System Model 53
Markov chain flow matrix [36],

⎡ ⎤
−γi 0 ··· 0 0 γi
⎢ζi + μi −β − μi ··· 0 0 γi ⎥
⎢ ⎥
⎢ . .. .. .. .. .. ⎥
⎢ . ⎥
Λi = ⎢ . . . . . . ⎥. (4.4)
⎢ .. ⎥
⎢ μ 0 . ζi −β − μi γi ⎥
⎣ i ⎦
..
μi 0 . 0 ζi −ζi − μi
Here ζi = β −γi , and γi represents the average request rate for the ith popular video
vi that can be denoted as
β
γi (t) = α , (4.5)
ρi
where the request for vi is assumed to arrive as a Poisson process with rate β [36],
and the probability follows a Zipf-like distribution.
Therefore, the probability of
requesting content vi is 1/ρi α , where ρ = Ii=1 1/i α and α is the Zipf slope with
0 < α < 1. The video content is refreshed periodically, and the lifetime of any
video content vi follows an exponential distribution with mean 1/μi .
4.2.5 Computing Model
On the condition that the versions of the requested video at all the providers’ caches
do not match with the requested version, the associated provider has to extract the
current video content and construct a computation task based on the input data
information, as well as the number of CPU cycles. In the computation model, we
construct the computation task as Qvsli = {osvli , qsvli }. The first parameter osvli repre-
sents the size of the requested-version video, and the second parameter qsvli indicates
the required number of CPU cycles that is needed to finish the computation task.
The computation capability of provider cp assigned to mobile user sl is denoted
c
as fsl p , which can be measured by the number of CPU cycles per second [15, 23].
In our network, multiple mobile users may associate with the same provider and
share the computing device simultaneously, which would result in the fact that the
computation capability for provider cp is not exactly known at the next time instant.
c
Therefore, the computation capability fsl p is modelled as a random variable, and
equally divided into N discrete levels denoted by E = {E0 , E1 , . . . , EN −1 }. The
c c
realization of the random variable fsl p is denoted as Fslp at time slot t. We model
the transition of the computation capability level of a provider as a Markov process.
c
The transitional probability matrix of Fslp can be expressed as:
c * +
Θslp (t) = ιxs ys (t) N ×N , (4.6)

c c
where ιxs ys (t) = Pr Fslp (t + 1) = ys | Fslp (t) = xs , and xs , ys ∈ E .
The consumed time for executing the task Qvsli at provider cp can be obtained
v
qsli
as tsl ,cp = cp . Furthermore, the rate of computation, i.e., the number of bits
Fsl (t)
computed per second, can be given
c
comp comp osvli comp fsl p (t)osvli
rsl ,cp (t) = asl ,cp (t) = asl ,cp (t) c , (4.7)
tsl ,cp qslp
comp
where asl ,cp (t) ∈ {0, 1} represents whether or not the computation task is decided
to be performed at provider cp .
4.3 Social Trust Scheme with Uncertain Reasoning
In this section, we will derive how to obtain the trust values of mobile users. We
evaluate the trustworthiness of a mobile user by a real number T r ranging from 0 to
1. In our model, the trust value T r is jointly determined based on direct observations
and indirect observations. The direct observation trust of a mobile user is defined as
the estimated degree of trustworthiness from its directly-connected mobile users
based on their past experiences. However, the subjective evaluation from direct
connections may be prejudiced, and therefore in order to be more objective and
impartial, we also consider the rating of trust from other indirectly-connected mobile
users. Here, we denote the trust value from direct observations as T r D and the trust
value from indirect observations as T r I nD . By combining these two trust values, we
can obtain a more accurately estimated trust value of a mobile user as
T r = ωT r D + (1 − ω)T r I nD (4.8)
where ω is the weight coefficient to adjust the weightiness between direct and
indirect observations, and 0 ≤ ω ≤ 1.
The trust evaluation procedure in our model can be visually explained as Fig. 4.2.
In the following, we will discuss how to obtain the trust evaluation from direct
Fig. 4.2 Social trust

evaluation with both direction
observation and indirect
observation
4.3 Social Trust Scheme with Uncertain Reasoning 55
observations and indirect observations by using the Bayesian inference approach

and Dempster-Shafer theory [37], respectively.
4.3.1 Trust Evaluation from Direct Observations
In the direct observations, consider that an observing mobile user can overhear the
data forwarded by the observed mobile user, and identify the observed mobile user’s
malicious behaviors, such as discarding or modifying some of the original data.
Through multiple observations of the observed mobile user’s behavior, the
observing mobile user can evaluate the trust value by exploiting Bayesian infer-
ence [38], which is a method of statistical inference using the Bayes’ theorem to
update the probability for a hypothesis with more evidence becomes available.
Under the Bayesian framework, we model the trust of a mobile user as a
continuous random variable, denoted as Φ, where φ takes values from 0 to 1. We
assume that Φ follows a beta distribution [39], i.e., Φ ∼ Beta(a, b), which is
defined as follows with parameters a and b
φ a−1 (1 − φ)b−1
Beta(a, b) = , 1 , (4.9)
0 φ a−1 (1 − φ)b−1 dφ
for 0 ≤ φ ≤ 1. Here, since Φ is assumed to obey a beta distribution, the trust value
can be represented by the two parameters a and b.
We summarize our belief of the trust Φ in a probability distribution iteratively as
more observations are available. Assume that the prior probability density function
(pdf) at the (t − 1)th observation is known. Then, according to the Bayes theorem,
the posterior distribution at the tth observation can be obtained with the pdf as
ft (xt |φ, yt )ft−1 (φ)
ft (φ) = , 1 , (4.10)
0 ft (xt |φ, yt )ft−1 (φ) dφ
where xt and yt are the number of data packets that has been forwarded correctly and
the number of packets received by the observed mobile user at the tth observation,
respectively; ft (xt |φ, yt ) is the likelihood function, which follows a binomial
distribution as
% &
yt xt
ft (xt |φ, yt ) = φ (1 − φ)yt −xt . (4.11)
xt
In Bayesian inference, the beta distribution is the conjugate prior probability

distribution for the binomial distribution [38, 39]. Since the likelihood function
ft (xt |φ, yt ) follows a binomial distribution, the priori distribution ft−1 (φ) is
definitely assumed to follow a beta distribution, which reflects what is already
known about the distribution of Φ at the (t − 1)th observation. Given that the priori
distribution ft−1 (φ) follows a beta distribution, the posterior distribution ft (φ) also
follows a beta distribution. Particularly, if ft−1 (φ) ∼ Beta(at−1 , bt−1 ) and xt , yt
from the tth observation are also given, then
ft (φ) ∼ Beta(at−1 + xt , bt−1 + yt − xt ), t ≥ 1. (4.12)
At the very beginning of the observation, there is no evidence about the distribution
of Φ, thus we assume that it follows a uniform distribution, i.e., f0 (φ) ∼ Beta(1, 1).
Summarily, we can say that ft (φ) follows Beta(at , bt ) with parameters
at = at−1 + xt , bt = bt−1 + yt − xt ,
a0 = 1, b0 = 1. (4.13)
The trust value can be represented with the mathematical expectation of beta
distribution as
at
Et [Φ] = . (4.14)
at + bt
Intuitively, the trust value of a mobile user is 0.5 at the beginning, and updated
continuously by using the follow-up observations.
From the above discussion, we can see that past experiences play an important
role in the Bayesian inference. In fact, recent behaviors of a mobile user should
weight more in the trust evaluation. Here, we introduce a punishment factor for
reputation fading, which gives more weights on misbehaviors in the Bayesian
inference. The trust evaluation formula in (4.14) is revised as follows:
at
Et [Φ] = , (4.15)
at + τ bt
where τ is the punishment factor, and τ ≥ 1. With the increment of τ , the trust value
declines quickly.
The punishment factor makes the trust evaluation more realistic and reliable.
First, if a mobile user once behaves maliciously, compared with those who have no
bad records, its trust value will be lowered much more. Second, the trust value will
not recover quickly even if he behaves well recently because of the constraint of the
punishment factor. This helps distinguish the malicious mobile users quickly and
prevent them disrupting others’ trust evaluation. Based on the above deduction, the
trust value from direct observation, T r D , is defined as:
T r D = Et [Φ]. (4.16)
4.3.2 Trust Evaluation from Indirect Observations
Apart from the direct observations, indirect observations from other mobile users
are also very important in assessing the trustworthiness of an observed mobile user.
Considering the indirect observations helps mitigate the situation that an observed
mobile user is loyal to one mobile user but cheating on others. Assume that an
4.3 Social Trust Scheme with Uncertain Reasoning 57
observing mobile user collects observations from several other mobile users (also
called subsidiary observing mobile user), and combines the collected evidence into
a decision about the observed mobile user’s trust value. However, these subsidiary
observing mobile users may be untrustworthy or the evidence offered by them is
unreliable.
The Dempster-Shafer theory can be used as an effective way to handle the
uncertainty issue and combine the evidence from multiple subsidiary observers [40].
The core of this theory is based on two ideas: the degrees of belief about a
proposition can be obtained from multiple subjective probabilities of a related
theme, and these degrees of belief can be combined together under the condition
that they are from independent evidence [37]. In the indirect observation, we assume
that there are more than one subsidiary observing mobile users and the evidence
provided by them is mutually independent.
Belief Function
As an introduction to the belief function, suppose that each subsidiary observing

mobile user has two states, i.e., trustworthy and untrustworthy with probabilities p1
and 1 − p1 respectively. Assume that mobile user 1 claims that the observed mobile
user B is trustworthy. If mobile user 1 is trustworthy, its statement is regarded as true;
however, if mobile user 1 is untrustworthy, its statement is not necessarily untrue.
We think that mobile user 1 provides p1 degree of belief in the observed mobile
user’s trustworthiness (hypothesis H ), 0 degree of belief in the untrustworthiness
(hypothesis H ), and 1 − p1 degree in the uncertainty, i.e., either trustworthy or
untrustworthy (hypothesis U ). Each hypothesis is assigned with a basic probability
value m(H ), m(H ), m(U ) taking values from the interval [0, 1], respectively. In
our scheme, we assign the basic probability value with the direct observation trust
value. For example, if mobile user 1 believes that mobile user B is trustworthy, then
the basic probability value for each possible hypothesis is:
m1 (H ) = TA1
D
,
m1 (H ) = 0,
m1 (U ) = 1 − TA1
D
, (4.17)
D represents the trust value of mobile user 1 from the direct observation of
where TA1
mobile user B.
Oppositely, if mobile user 1 thinks that user B is untrustworthy, the basic
probability value will be:
m1 (H ) = 0,
m1 (H ) = TA1
D
,
m1 (U ) = 1 − TA1
D
. (4.18)
Concerning the formal definition of the belief function, let Ω be the universe of
discourse, i.e., the set representing all possible states of the considered system.
Ω = {trustworthy, untrustworthy}. The power set 2Ω refers to the set of

all subsets of Ω, here 2Ω = {∅, {trustworthy}, {untrustworthy}, Ω}. Any
hypothesis Ai refers to a subset of 2Ω , and is mapped into a basic probability
m(Ai ) that reflects the proportion of total belief assigned
to hypothesis Ai . These
two conditions should be satisfied: m(∅) = 0 and Ai ⊆Ω m(Ai ) = 1. For any
hypothesis B, the belief function is defined as

beli(B) = m(Ai ), (4.19)
Ai ⊆B
which represents the strength of the evidence that supports hypothesis B’s
provability.
Dempster’s Rule of Combining Belief Functions
Based on the definition of belief function, the Dempster-Shafer theory combines

multiple users’ belief. Assuming that beli1 (B) and beli2 (B) are two belief functions
over the same universe of discourse, Ω. the orthogonal sum of beli1 (B) and
beli2 (B) can be defined as
beli(B) =beli1 (B) ⊕ beli2 (B)
-
i,j,Ai Aj =B m1 (Ai )m2 (Aj ) (4.20)
= -
,
i,j,Ai Aj =∅ m1 (Ai )m2 (Aj )
where Ai , Aj ⊆ Ω.
In our scenario, the combined degree of belief from mobile user 1 and mobile
user 2 can be calculated as follows [37]:
1
m1 (H ) ⊕ m2 (H ) = [m1 (H )m2 (H ) + m1 (H )m2 (U )
K
+ m1 (U )m2 (H )],
1
m1 (H ) ⊕ m2 (H ) = [m1 (H )m2 (H ) + m1 (H )m2 (U ) (4.21)
K
+ m1 (U )m2 (H )],
1
m1 (U ) ⊕ m2 (U ) = m1 (U )m2 (U ),
K
where
K =m1 (H )m2 (H ) + m1 (H )m2 (U ) + m1 (U )m2 (U )
+ m1 (U )m2 (H ) + m1 (U )m2 (H ) + m1 (H )m2 (H ) (4.22)
+ m1 (H )m2 (U ).
Following the rule of combination of belief, we can combine more degree of

belief from other mobile users. Based on the Dempster-Shafer theory, TABN is
defined as:
I nD
T rAB = m1 (H ) ⊕ m2 (H ) . . . ⊕ mn (H ), (4.23)
where there are totally between mobile user A and mobile user B.
4.4 Problem Formulation
In this section, we formulate an optimization problem of the integrated trust-based

social network with the 3C framework with MEC, caching, and D2D communica-
tions. We assume that a mobile user requests for a video content to the integrated
network. For the network operator, it should decide which BS or a D2D transmitter
is assigned to serve the requesting user, whether or not the video transcoding (i.e.,
MEC) should be performed, and whether or not newly emerged contents should be
cached. The network operator need a comprehensive consideration of many factors,
including: the wireless channel conditions, whether or not the requested content
is stored at the local cache, whether or not the content version is matched up, the
computational capacity, the trustworthiness of a D2D transmitter. Here, we consider
dynamic scenarios, i.e., the available network conditions are varying with time. We
exploit deep Q-learning algorithm to solve the formulated optimization problem.
In order to obtain the optimal policy, identifying the system’s states, actions, and
reward functions is necessarily required, which will be described in more details
below.
4.4.1 System State
The system states for a subscriber sl requesting video vi at time slot t ∈

c
{0, 1, . . . , T − 1} mainly includes five components: the realization Υsl p (t) of the
c c
random variables γsl p (channel state), the realization Fslp (t) of the random variables
cp cp c
fsl (computation capability), the realization Xvi (t) of the random variables xvip
cp cp
(content indicator), the version indicator yvi , and the trust value index T rsl for all
the video providers, i.e., ∀p ∈ C . Consequently, the system state vector can be
described as follows.
* c
χsl (t) = Υscl 1 (t), Υscl 2 (t), . . . , Υsl K+M (t),
c
Fscl1 (t), Fscl2 (t), . . . , FslK+M (t),
c
Xvc1i (t), Xvc2i (t), . . . , XvK+M
i (t), (4.24)
yvi (t),
c +
T rscl1 (t), T rscl2 (t), . . . , T rslK+M (t) .
c
Here, the trust index for all the BSs should be set to 1, i.e., T rslp (t) = 1, ∀cp ∈
CK ; meanwhile, this index for all the D2D transmitters should be less than 1, i.e.,
c
T rslp (t) < 1, ∀cp ∈ CM . yvi (t) denotes the indicator for whether or not the cached
video version matches with the requested version. If yes yvi (t) = 1, if no yvi (t) = 0.
4.4.2 System Action
In our system, the network controller decides which video provider is assigned to
the subscribed user, whether or not the computation offloading (video transcoding)
should be performed, and whether or not the video provider should cache the new
video. The network controller’s composite action for subscriber sl is given by
comp
asl (t) = {acomm
sl (t), asl (t), acache
sl (t)}, (4.25)
comp
where the row vectors acomm
sl (t), asl (t), acache
sl (t) are interpreted in more details
as follows:
• The first row vector acomm
sl (t) is defined as
acomm
sl (t) = [ascomm
l ,c1
(t), ascomm
l ,c2
(t), . . . , ascomm
l ,cK+M
(t)], (4.26)
where ascomml ,cp

(t) is the association indicator, and ascomm
l ,cp
(t) ∈ {0, 1}. If the
subscribed user is associated with video provider cp , asl ,cp (t) = 1, otherwise
comm
ascomm
l ,cp
(t) = 0.
comp
• The second row vector asl (t) is defined as
comp comp comp comp
asl (t) = [asl ,c1 (t), asl ,c2 (t), . . . , asl ,cK+M (t)], (4.27)
comp comp
where asl ,cp (t) indicates the computation offloading decision, and asl ,cp (t) ∈
{0, 1}. If the network controller has decided that the computation task should
comp
be performed at the provider cp ’s device, asl ,cp (t) = 1; on the other hand, if
the computation is decided to be executed on subscriber’s own mobile device,
comp
asl ,cp (t) = 0.
• The third row vector acache
sl (t) is defined as
acache
sl (t) = [ascache
l ,c1
(t), ascache
l ,c2
(t), . . . , ascache
l ,cK+M
(t)], (4.28)
where ascachel ,cp

(t) means the whether or not the video provider cp should cache
the newly-emerged video or the new version, and ascache l ,cp
(t) ∈ {0, 1}. If the
network controlled decides that cp caches the video, asl ,cp (t) = 1, otherwise
cache
ascache
l ,cp
(t) = 0.
4.4.3 Reward Function
In this chapter, we set the system reward to be the total revenue (i.e., utility) of
the network operator. The network operator charges the subscribed user sl for
associating with the video provider, which is denoted as λsl unit price per bps.
On the condition that video transcoding (computation offloading) is decided to
be executed on video provider’s side, the operator can charge for its computing
service, which is defined as νsl unit price per bps. In addition, if the network
controller decides to let the BSs or D2D transmitters cache the new video or new
version, the operator gets a potential revenue on estimated backhaul save, which is
denoted as κsl unit price per Hz.
On the other hand, the operator has to pay for the rented spectrum and backhaul
bandwidth. The unit price for the usage of spectrum is defined as δcp per Hz for
provider cp . Moreover, if video transcoding is performed, a certain amount of
energy will be consumed for the computing, the network operator is obliged to pay
ηcp unit price for per consumed Joule. The cost of caching the video content in the
memory of provider cp is denoted as ςcp unit price per unit space.
The system reward for serving subscribed user sl requesting video vi is defined as
the network operator’s total revenue, and it is formulated as a function of the system
states, and actions. The system actions determine whether or not the reward can be
optimized. Here, we define the reward function for a specific subscribed user sl as:
comp
!
Rsvli (t) = Rscomm
,c
l p
(t) + Rs l ,c p (t) + Rs
cache
,c
l p
(t)
cp ∈C
c c comp
= ascomm
l ,cp
(t) λsl T rslp (t)Υsl p (t)(1 − w comp asl ,cp (t)) − δcp B
cp ∈C

c c
− (1 − Xvpi (t))σ cp Υsl p (t) (4.29)
c comp c
c
Fslp (t)osvli c
+ (1 − yvip )asl ,cp (t)(νsl T rslp (t) − ηcp qslp ecp )
qsvli
cp ∈C
c c
+ ascache
l ,cp
(t)(κsl T rslp (t)Υsl p (t) − ςcp osvli ).
cp ∈C
The above reward function is composed of three terms, i.e., the earnings from
c
communications, computing and caching, respectively. Here, the trust index T rslp (t)
is added to each incoming revenue term. For the first communication earnings,
c c comp
λsl T rslp (t)Υsl p (t)(1 − w comp asl ,cp (t)) denotes the available earnings for providing
the video transmitting service to subscriber sl . Note that if video transcoding is
decided to be performed, the video will be compressed and the corresponding
income will be reduced. δcp B is the cost for consuming the radio spectrum
c c
bandwidth. (1 − Xvpi (t))σ cp Υsl p (t) is the cost for possible consumed backhaul
c
bandwidth: if Xvpi (t) = 1, i.e., the requested video exists in the cache of provider
c
cp , no backhual cost is needed; otherwise Xvpi (t) = 0, the video has to be fetched
from the Internet and backhaul cost is unavoidable. Note that σ cp is a very important
parameter. It is not realistic to build up a D2D communication when the requested
video does not exist in its cache. Thus, σ cp should be set to be an extremely large
c
number for cp , ∀cp ∈ CM . For the second computing earnings, (1 − yvip ) is used
cp v
Fsl (t)osli
to indicate whether or not the version is matching. νsl T rsl (t)cp v represents
qsli
c
the gaining for implementing video transcoding, and ηcp qslp ecp is the expenditure
for the energy consumption, where ecp denotes consumed energy for running one
c
CPU cycle. For the third caching earnings, κsl T rsl (t)cp Υsl p (t) is the estimated
future income of the saved backhaul bandwidth if the new video or new version is
decided to be cached. Here the saved backhaul bandwidth is equal to the wireless
communication rate. At last, ςcp osvli denotes the cost for caching the content.
Rsvli (t) is the system’s immediate reward, i.e., the network operator gets Rsvli (t) at
state χsl (t) when action asl (t) is performed in time slot t. However, a maximum
immediate value is not a guarantee for the maximum long-term future rewards.
Thus, we should also consider a big picture. A future reward with a discount factor
ϕ is reasonable, which can be denoted as
t=T −1
f

Rsl ,vi = max E ϕ t
Rsvli (t) , (4.30)
π
t=0
where π represents a Q-learning policy, i.e., a series of actions where a specific

action will be executed given a specific system state, and ϕ t approximates to be
zero when T is large enough. In fact, a condition for terminating the training process
should be set.
The objective of adopting deep Q-learning into our network model is to help
find an optimization policy that can maximize the cumulated future rewards for the
network operator.
4.5 Simulation Results and Discussions
In this section, we evaluate the performance of proposed scheme using computer

simulations. We use TensorFlow [41] in our simulations to implement deep Q-
learning. For performance comparison, the following four algorithms are presented:
(1). Proposed scheme, which considers MEC, caching, D2D, as well as both
direct observation and indirect observation; (2). Proposed scheme w.o. indirect
observation, which does not consider indirect observation; (3). An existing scheme
w.o. mobile edge computing [42]; and (4). An existing scheme w.o. D2D communi-
cations [2].
4.5.1 Simulation Settings
In the simulations, we consider an MSN consisting of 5 BSs, and 15 D2D

transmitters. The radius of the cell is set to 500 m. Because D2D communications
typically perform within short ranges, we use a clustered-based distribution model
[43], where multiple users are located within one cluster with a radius of 50 m. In the
network, there are 20 subchannels, each of which has a bandwidth of 180 KHz. The
transmit powers of D2D transmitter and BS are 24 and 46 dBm, respectively. The
noise spectral density is −174 dBm/Hz. A loss model 35.3 + 37.6 log(d(m)) [44]
is used. In addition, block fading with a block size of 100 is assumed for channel
fading. Moreover, we assume that there are totally 10 types of contents distributed
in the network, and each content cache state follows the Markov model. We set the
cache state transition probability of staying in the same state in the BS as 0.6 (0.3 in
the D2D transmitter), and set the transition probability from one state to another as
0.4 (0.7 in the D2D transmitter). We further assume a Zipf popularity distribution in
the simulations, with θ = 1.5 [45]. The computation states of MEC servers follow
the Markov model. We assume that the computation state of the MEC server can be
very low, low, medium, high, and very high.
We assume that there are two types of D2D transmitters in the network:
normal nodes, which share the content and perform computing normally, and
compromised nodes, which modify contents maliciously. The BSs are assumed to
be not compromised due to the physical security of BSs. We also assume that the
number of compromised nodes is much less than the total number of nodes in the
network. The attackers are independent. Hence, there is no collusion attack in the
network.
In our simulations, we used a GPU-based server, which has 4 Nvidia GPUs with
version GTX TITAN. The CPU is Intel Xeno E5-2683 v3 with 128G memory. The
software environment we utilized is TensorFlow 0.12.1 with Python 2.7 on Ubuntu
14.04 LTS.
The training process of the deep convolutional networks are shown in Fig. 4.3.
We formulate the 5 BSs and 15 D2D transmitters as the rows, and the 5 levels of
computational capacity as the columns in a grid image. Whether or not the required
Fig. 4.3 The training process of the deep convolution networks

100
Proposed scheme
Proposed scheme w.o. indirecto bservation
95 Existing scheme w.o. mobile edge computing
Existing scheme w.o. D2D communications
90
Total reduced backhaul usage (Mbps)
85
80
75
70
65
60
55
50
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Average size per content (MB)
Fig. 4.4 The total reduced backhaul usage v.s. average size per content
content is in the local cache is characterized by different colors in each small grid.
The initial input image of size 20 × 5 × 3 is firstly resized into 84 × 84 × 3. Through
4 layers of convolutional operations, the output is 512 nodes, and these nodes will
be fully connected to be trained in the deep reinforcement learning.
4.5.2 Simulation Results
The effects of average size per content on the total reduced backhaul usage is
shown in Fig. 4.4. There two malicious D2D transmitters, and there are 4 types
of contents in the system. We can see from Fig. 4.4 that the total reduced backhaul
usage increases with the increase of the average size per content. This is because
a larger content size will increase the fee for caching the content, which induces a
lower gain of caching utility, and thus the system is reluctant to cache the contents
in the network. Compared to the existing schemes without mobile edge computing
and D2D communications, the proposed scheme has larger total reduced backhaul
usage due to the benefits of mobile edge computing and D2D communications. In
addition, without indirect observation, the accuracy of trust evaluation is lower in
the proposed scheme, which induces a lower gain of mobile edge computing and
caching, and lower total reduced backhaul usage.
105
Proposed scheme
Proposed scheme w.o. indirect observation
100 Existing scheme w.o. mobile edge computing
95
Total reduced backhaul usage (Mbps)
90
85
80
75
70
65
60
55
1 2 3 4 5 6 7 8 9 10
Number of content types
Fig. 4.5 The total reduced backhaul usage v.s. the number of content types
Figures 4.5 and 4.6 show the effects of the number of content types on the total
reduced backhaul usage and the total utility, respectively. In these simulations, the
average size per content is 1.8 MB, and there are two malicious D2D transmitters.
We can see from both figures that the reduced backhaul usage and the total utility
decrease as the number of content types increases. This is because more content
types in the system will lead to the scenarios that a requesting mobile user is unlikely
to find the specific content from the BSs and D2D transmitters due to the limited
storage. In addition, more content types will result in reduced the popularity of all
contents according to the Zipf popularity distribution, thus decreasing caching gain.
Figure 4.7 shows the effects of the required numbers of CPU cycles for video
transcoding. From Fig. 4.7, we can see that the required number of CPU cycles has
no effects on the total utility of the existing scheme without mobile edge computing.
This is because the total utility will not be changed with different required numbers
of CPU cycles if mobile edge computing is not available in the system. In the
proposed scheme and the existing scheme without D2D communications, the total
utility decreases exponentially with the increase of the required number of CPU
cycles for video transcoding. This is because a larger number of required number of
CPU cycles will result in a higher consumed computation energy, and consequently
a lower gain of computation utility. Therefore, the total utility decreases with the
required number of CPU cycles for video transcoding.
Next, we study the performance of the proposed social trust scheme. Assume
that, at episode 14K, D2D transmitter 1 changes its maliciousness to 0.6. Our
6500
6000
5500
5000 Proposed scheme

Total utility
Existing scheme w.o. mobile edge computing

4500 Existing scheme w.o. D2D communications
4000
3500
3000
2500
1 2 3 4 5 6 7 8 9 10
Number of content types
Fig. 4.6 The total utility v.s. the number of content types
goal is to observe the accuracy of trust tracking in this scenario where the
malicious behavior of a D2D transmitter changes rapidly. Figure 4.8 shows the
trust tracking of the system using direct observation with Bayesian inference and
the proposed scheme using both direct observation with Bayesian inference and
indirect observation with the Dempster-Shafer theory. We can observe from Fig. 4.8
that only direct observation can result in inaccurate trust values. By contrast, the
proposed scheme can track the trust value accurately with both direct observation
and indirect observation.
The number of malicious D2D transmitters in the network also has a significant
impact on the performance the network. Here, We investigate the system utility
with the number malicious D2D transmitters, from 1 to 5, in a 15 D2D transmitter
environment. The basic parameter is the same as above. Figure 4.9 shows that, as
the number of malicious D2D transmitters increases, the utility drops dramatically.
When the number of malicious D2D transmitters reaches to one third of the total
number of D2D transmitters in the network, the utility decreases to about half of the
utility in the network with 1 malicious nodes. From this figure, we can see that the
network is deeply affected by the number of malicious D2D transmitters.
Fig. 4.7 The total utility v.s. the required number of CPU cycles for video transcoding
1.2
Expected trust of node 1
Evaluated trust of node 1 in the proposed scheme
Evaluated trust of node 1 in the proposed scheme w.o. indirect observation
1
0.8
Trust value
0.6
0.4
0.2
0
1.38 1.39 1.4 1.41 1.42 1.43 1.44 1.45
×104
Episode
Fig. 4.8 The trust value v.s. the episodes

6000
Proposed scheme
Existing scheme w.o. mobile edge computing
5500
5000
Total utility
4500
4000
3500
1 1.5 2 2.5 3 3.5 4 4.5 5
The number of malicious D2D transmitters
Fig. 4.9 The total utility v.s. the number of malicious D2D transmitters
4.6 Conclusions and Future Work
MSNs have become one of the most important networking paradigms in future
wireless mobile networks. Recent advances of wireless mobile networks can have
significant impacts on the performance of MSNs. In this chapter, we studied recent
advances in mobile edge computing, in-network caching and D2D communications
in MSNs. In addition, we considered the knowledge of social relationships in these
new paradigms to improve the security and efficiency of MSNs. Specifically, we pre-
sented a social trust scheme with both direct observation using Bayesian inference
and indirect observation using the Dempster-Shafer theory. We further proposed a
deep Q-learning approach to study this complicated system. Extensive simulation
results were presented to show the effectiveness of the proposed scheme. Since
energy efficiency is becoming increasingly important in the big data era [46, 47]
Future work is in progress to consider energy-efficient resource allocation strategy
in the proposed integrated framework.
References 69
References
1. N. Vastardis and K. Yang, “Mobile social networks: Architectures, social properties, and key
research challenges,” IEEE Comm. Surveys Tutorials, vol. 15, no. 3, pp. 1355–1371, Third
Quarter 2013.
2. Z. Su and Q. Xu, “Content distribution over content centric mobile social networks in 5G,”
IEEE Comm. Mag., vol. 53, no. 6, pp. 66–72, Jun. 2015.
3. ETSI, “Mobile-edge computing – introductory technical white paper,” ETSI White Paper, Sep.
2014.
4. N. Kumar, S. Zeadally, and J. J. P. C. Rodrigues, “Vehicular delay-tolerant networks for smart
grid data management using mobile edge computing,” IEEE Comm. Magazine, vol. 54, no. 10,
pp. 60–66, Oct. 2016.
5. C. Wang, C. Liang, F. R. Yu, Q. Chen, and L. Tang, “Computation offloading and resource
allocation in wireless cellular networks with mobile edge computing,” IEEE Trans. Wireless
Comm., vol. 16, no. 8, pp. 4924–4938, Aug. 2017.
6. Y. Meng, C. Jiang, H. H. Chen, and Y. Ren, “Cooperative device-to-device communications:
Social networking perspectives,” IEEE Network, vol. 31, no. 3, pp. 38–44, May 2017.
7. K. Wang, F. R. Yu, and H. Li, “Information-centric virtualized cellular networks with device-
to-device (D2D) communications,” IEEE Trans. Veh. Tech., vol. 65, no. 11, pp. 9319–9329,
Nov. 2016.
8. Y. Zhang, E. Pan, L. Song, W. Saad, Z. Dawy, and Z. Han, “Social network aware device-
to-device communication in wireless networks,” IEEE Trans. Wireless Comm., vol. 14, no. 1,
pp. 177–190, Jan. 2015.
9. P. Hui, J. Crowcroft, and E. Yoneki, “Bubble rap: Social-based forwarding in delay-tolerant
networks,” IEEE Trans. Mobile Comput., vol. 10, no. 11, pp. 1576–1589, Nov. 2011.
10. E. Bulut and B. K. Szymanski, “Friendship based routing in delay tolerant mobile social
networks,” in Proc. IEEE GLOBECOM, Miami, FL, USA, Dec. 2010, pp. 1–5.
11. S. Wang, X. Zhang, Y. Zhang, L. Wang, J. Yang, and W. Wang, “A survey on mobile edge
networks: Convergence of computing, caching and communications,” IEEE Access, vol. 5,
pp. 6757–6779, Mar. 2017.
12. K. Zhang, Y. Mao, S. Leng, Q. Zhao, L. Li, X. Peng, L. Pan, S. Maharjan, and Y. Zhang,
“Energy-efficient offloading for mobile edge computing in 5G heterogeneous networks,” IEEE
Access, vol. 4, pp. 5896–5907, Aug. 2016.
13. J. Liu, Y. Mao, J. Zhang, and K. B. Letaief, “Delay-optimal computation task scheduling
for mobile-edge computing systems,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), Barcelona,
Spain, Jul. 2016, pp. 1451–1455.
14. S. Sardellitti, G. Scutari, and S. Barbarossa, “Joint optimization of radio and computational
resources for multicell mobile-edge computing,” IEEE Trans. Signal Inf. Process. over Netw.,
vol. 1, no. 2, pp. 89–103, Jan. 2015.
15. X. Chen, L. Jiao, W. Li, and X. Fu, “Efficient multi-user computation offloading for mobile-
edge cloud computing,” IEEE/ACM Trans. Networking, vol. 24, no. 5, pp. 2795–2808, 2016.
16. X. Chen, J. Wu, Y. Cai, H. Zhang, and T. Chen, “Energy-efficiency oriented traffic offloading
in wireless networks: A brief survey and a learning approach for heterogeneous cellular
networks,” IEEE J. Sel. Areas Commun., vol. 33, no. 4, pp. 627–640, Jan. 2015.
17. J. Gu, W. Wang, A. Huang, and H. Shan, “Proactive storage at caching-enable base stations
in cellular networks,” in Proc. IEEE Int. Symp. on Personal Indoor and Mobile Radio Comm.
(PIMRC), London, UK, Sept. 2013, pp. 1543–1547.
18. B. Bai, L. Wang, Z. Han, W. Chen, and T. Svensson, “Caching based socially-aware D2D
communications in wireless content delivery networks: a hypergraph framework,” IEEE
Wireless Commun., vol. 23, no. 4, pp. 74–81, Aug. 2016.
19. A. Sengupta, S. Amuru, R. Tandon, R. M. Buehrer, and T. C. Clancy, “Learning distributed
caching strategies in small cell networks,” in Proc. 11th Int. Symp. Wireless Commun. Syst.
(ISWCS), Aug. 2014, pp. 917–921.
20. Y. Guan, Y. Xiao, H. Feng, C.-C. Shen, and L. J. Cimini, “Mobicacher: Mobility-aware content
caching in small-cell networks,” in Proc. IEEE GLOBECOM, Austin, TX, USA, Feb. 2014,
pp. 4537–4542.
21. Y. He, C. Liang, F. R. Yu, N. Zhao, and H. Yin, “Optimization of cache-enabled opportunistic
interference alignment wireless networks: A big data deep reinforcement learning approach,”
in Proc. IEEE ICC’17, Paris, France, May 2017, pp. 1–6.
22. Y. He, F. R. Yu, N. Zhao, V. Leung, and H. Yin, “Software-defined networks with mobile edge
computing and caching for smart cities: A big data deep reinforcement learning approach,”
IEEE Commun. Magazine, Sept. 2017.
23. Y. Zhou, F. R. Yu, J. Chen, and Y. Kuo, “Resource allocation for information-centric virtualized
heterogeneous networks with in-network caching and mobile edge computing,” IEEE Trans.
Veh. Tech., vol. 66, no. 12, pp. 11 339–11 351, Dec. 2017.
24. C. Wang, C. Liang, F. R. Yu, Q. Chen, and L. Tang, “Computation offloading and resource
allocation in wireless cellular networks with mobile edge computing,” IEEE Trans. Wireless
Comm., vol. 16, no. 8, pp. 4924–4938, Aug. 2017.
25. Y. He, N. Zhao, and H. Yin, “Integrated networking, caching, and computing for connected
vehicles: A deep reinforcement learning approach,” IEEE Trans. Veh. Tech., vol. 67, no. 1,
pp. 44–55, Jan. 2018.
26. Y. He, F. R. Yu, N. Zhao, V. C. Leung, and H. Yin, “Software-defined networks with
mobile edge computing and caching for smart cities: A big data deep reinforcement learning
approach,” IEEE Commun. Magazine, vol. 55, no. 12, pp. 31–37, Dec. 2017.
27. Z. Su and Q. Xu, “Security-aware resource allocation for mobile social big data: A matching-
coalitional game solution,” IEEE Trans. on Big Data, vol. PP, no. 99, pp. 1–1, 2017.
28. X. Hou, Y. Li, M. Chen, D. Wu, D. Jin, and S. Chen, “Vehicular fog computing: A viewpoint
of vehicles as the infrastructures,” IEEE Trans. Veh. Tech., vol. 65, no. 6, pp. 3860–3873, Jun.
2016.
29. A. Asadi, Q. Wang, and V. Mancuso, “A survey on device-to-device communication in cellular
networks,” IEEE Commun. Surveys Tutorials, vol. 16, no. 4, pp. 1801–1819, Fourth Quarter
2014.
30. C. Yang, J. Li, P. Semasinghe, E. Hossain, S. M. Perlaza, and Z. Han, “Distributed interference
and energy-aware power control for ultra-dense D2D networks: A mean field game,” IEEE
Trans. Wireless Comm., vol. 16, no. 2, pp. 1205–1217, Feb. 2017.
31. K. Hamedani, L. Liu, R. Atat, J. Wu, and Y. Yi, “Reservoir computing meets smart grids: attack
detection using delayed feedback networks,” IEEE Trans. Ind. Inf., vol. 14, no. 2, pp. 734–743,
Feb. 2018.
32. R. Atat, L. Liu, H. Chen, J. Wu, H. Li, and Y. Yi, “Enabling cyber-physical communication
in 5G cellular networks: challenges, spatial spectrum sensing, and cyber-security,” IET Cyber-
Phys. Syst.Theory Appl., vol. 2, no. 1, pp. 49–54, Apr. 2017.
33. I. R. Chen, F. Bao, and J. Guo, “Trust-based service management for social Internet of things
systems,” IEEE Trans. Dependable and Secure Computing, vol. 13, no. 6, pp. 684–696, Nov.
2016.
34. F. Parzysz, M. Di Renzo, and C. Verikoukis, “Power-availability-aware cell association for
energy-harvesting small-cell base stations,” IEEE Trans. Wireless Commun., vol. 16, no. 4,
pp. 2409–2422, Apr. 2017.
35. H. S. Wang and N. Moayeri, “Finite-state Markov channel-a useful model for radio communi-
cation channels,” IEEE Trans. Veh. Tech., vol. 44, no. 1, pp. 163–171, Feb. 1995.
36. H. Gomaa, G. G. Messier, C. Williamson, and R. Davies, “Estimating instantaneous cache hit
ratio using markov chain analysis,” IEEE/ACM Trans. Netw., vol. 21, no. 5, pp. 1472–1483,
Oct. 2013.
37. T. M. Chen and V. Venkataramanan, “Dempster-Shafer theory for intrusion detection in ad hoc
networks,” IEEE Internet Comput., vol. 9, no. 6, pp. 35–41, Nov. 2005.
38. R. Changiz, H. Halabian, F. R. Yu, I. Lambadaris, H. Tang, and C. M. Peter, “Trust
establishment in cooperative wireless networks,” in Proc. IEEE MilCom’10, San Jose, CA,
Nov. 2010, pp. 1074–1079.
References 71
39. C. Zouridaki, B. L. Mark, M. Hejmo, and R. K. Thomas, “Hermes: A quantitative trust

establishment framework for reliable data packet delivery in MANETs,” Journal of Computer
Security, vol. 15, no. 1, pp. 3–38, Jan. 2007.
40. S. Bu, F. R. Yu, P. Liu, P. Manson, and H. Tang, “Distributed combined authentication and
intrusion detection with data fusion in high-security mobile ad hoc networks,” IEEE Trans.
Veh. Tech., vol. 60, no. 3, pp. 1025–1036, Mar. 2011.
41. M. Abadi, A. Agarwal et al., “Tensorflow: Large-scale machine learning on heterogeneous
systems,” arXiv:1603.04467, Nov. 2015.
42. Y. Zhao, W. Song, and Z. Han, “Social-aware data dissemination via device-to-device
communications: Fusing social and mobile networks with incentive constraints,” IEEE Trans.
Services Computing, vol. PP, no. 99, pp. 1–1, 2017.
43. B. Kaufman and B. Aazhang, “Cellular networks with an overlaid device to device network,” in
Proc. Asilomar Conf. Signals, Systems & Computers (ACSSC), Pacific Grove, CA, Oct. 2008,
pp. 1537–1541.
44. D. H. Lee, K. W. Choi, W. S. Jeon, and D. G. Jeong, “Two-stage semi-distributed resource
management for device-to-device communication in cellular networks,” IEEE Trans. Wireless
Commun., vol. 13, no. 4, pp. 1908–1920, Apr. 2014.
45. P. Blasco and D. Gunduz, “Learning-based optimization of cache content in a small cell base
station,” in Proc. IEEE Int. Conf. Commun. (ICC), Sydney, Australia, Jun. 2014, pp. 1897–
1903.
46. J. Wu, S. Guo, J. Li, and D. Zeng, “Big data meet green challenges: big data toward green
applications,” IEEE Systems Journal, vol. 10, no. 3, pp. 888–900, Sept. 2016.
47. J. Wu, S. Guo, J. Li, and D. Zeng, “Big data meet green challenges: Greening big data,” IEEE
Systems Journal, vol. 10, no. 3, pp. 873–887, Sept. 2016.

DRL For WSN Book

Uploaded by

Copyright:

Available Formats

DRL For WSN Book

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

DRL For WSN Book

Uploaded by

Copyright:

Available Formats

SPRINGER BRIEFS IN

ELEC TRIC AL AND COMPUTER ENGINEERING

More information about this series at http://www.springer.com/series/10059

ISSN 2191-8112 ISSN 2191-8120 (electronic)

Library of Congress Control Number: 2018967964

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2019

A Brief Journey Through “Deep Reinforcement Learning

There is a phenomenal burst of research activities in machine learning and wireless

Chapter 2 focuses on reinforcement learning and deep reinforcement learning.

Ottawa, ON, Canada F. Richard Yu

1 Introduction to Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

3.4 Simulation Results and Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

Abstract Machine learning is evolved from a collection of powerful techniques in

1.1 Supervised Learning

Supervised learning is a kind of labelling learning technique. Supervised learning

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2019 1

Training input set

Training output set

Machine learning methods

Decision making phase

Fig. 1.1 The general processing procedure of a machine learning approach

1.1.1 k-Nearest Neighbor (k-NN)

The k-NN is a supervised learning technique, where the classification of a data

1.1.2 Decision Tree (DT)

The DT is one of the classification techniques which performs classification through

Machine learning techniques

Support vector machine

Hidden markov model

Deep reinforcement learning

RL-based game theory

Fig. 1.2 Common machine learning algorithms applied to SDN

1.1.3 Random Forest

Fig. 1.3 Example of k-NN

Input Layer Hidden Layer Output Layer

1.1.4 Neural Network (NN)

A neural network is a computing system made up of a large number of simple

The random NN can be represented as an interconnected network of neurons which

In feed-forward neural networks, the information is transmitted directionally from

1.1.5 Support Vector Machine (SVM)

1.1.6 Bayes’ Theory

Bayes’ theory uses the conditional probability to calculate the probability of an

where E is a new evidence, H is a hypothesis, P (H |E) is the posterior probability

1.1.7 Hidden Markov Models (HMM)

1.2 Unsupervised Learning

In contrast to supervised learning, an unsupervised learning algorithm is given a

The k-means algorithm is a popular unsupervised learning algorithm, which is

1.2.2 Self-Organizing Map (SOM)

1.3 Semi-supervised Learning

1. S. B. Kotsiantis, I. Zaharakis, and P. Pintelas, “Supervised machine learning: A review of clas-

8. W. R. Burrows, M. Benjamin, S. Beauchamp, E. R. Lord, D. McCollor, and B. Thomson,

Abstract In order to better understand state-of-the-art reinforcement learning

2.1 Reinforcement Learning

Reinforcement learning is an important branch of machine learning, where an agent

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2019 15

Generally, reinforcement learning can be divided into model-free and model-

2.2 Deep Q-Learning

Li (θi ) = E[(yi − Q(x, a; θi )2 )], (2.1)

2.3 Beyond Deep Q-Learning

2.3.1 Double DQN

where t approaches to zero when t is large enough. In practice, a threshold for

is the matrix of the delayed channel coefficients at the time instant t. δ