Machine Learning Assisted Caching and Adaptive LDPC Coded

Modulation for Next Generation Wireless Communications

Hassan Nooh, Zhikun Zhu and Soon Xin Ng

School of Electronics and Computer Science, University of Southampton, SO17 1BJ, U.K.

http://www.wireless.ecs.soton.ac.uk

Keywords:

Low Density Parity-check Codes, Latent Dirichlet Allocation, Adaptive Modulation and Coding, K-Means

Clustering.

Abstract:

Unmanned Aerial Vehicles (UAVs) constitute a key technology for next generation wireless communications.

Compared to terrestrial communications, wireless systems with low-altitude UAVs are in general faster to

deploy, more ﬂexible and are likely to have better communication channels due to the presence of short-range

Line of Sight (LoS) links. In this contribution, a Latent Dirichlet Allocation (LDA) based machine learning

algorithm was utilized to optimize the content caching of UAVs, while the K-means clustering algorithm

was invoked for optimizing the assignment of mobile users to the UAVs. We further investigated a practical

adaptive Low Density Parity Check (LDPC) coded modulation (ALDPC-CM) scheme for the communication

links between the UAVs and the users. We found that the caching efﬁciency of each UAV can be boosted

from 50% with random caching to above 90% with the employment of LDA. We also found that the proposed

ALDPC-CM scheme is capable of performing closely to the ideal perfect coding based scheme, where the

mean delay of the former is only about 0.05 ms higher than that of the latter, when the UAV system aims to

minimize both the transmission and request delays.

1 INTRODUCTION

Emerging technologies such as Internet of Things

(IoT) and autonomous vehicles are expected to be

commercialized as their performance requirements

are theoretically met by 5G speciﬁcations (Americas,

2018). Various methods of deployment and system ar-

chitectures have been proposed, yet a clear all-round

approach has not been found. Unmanned Aerial Ve-

hicles (UAVs), also referred to as drones, have been

massively employed in various applications during

the past several decades (Valavanis and Vachtsevanos,

2014), especially for military and recreational use.

More speciﬁcally, UAVs have been used mainly for

military purposes due to its high cost. However, con-

tinuous development has made it possible to build

low-cost and light-weight UAVs for civil and com-

mercial applications. UAV based wireless commu-

nication seems to be a promising solution to support

connectivity for users outside the infrastructure cov-

erage (Merwaday and Guvenc, 2015), which may be

caused by disasters, shadowing or overloading. Be-

sides, UAVs can relay the blocked signals from the

base station (BS) to users due to its advantage in mo-

bility and ﬂexibility (Zeng et al., 2016). In cellular

networks, each BS covers and serves a speciﬁc re-

gion. In order to extend the coverage area of the BS,

the concept of Remote Radio Head (RRH) has been

investigated (Chen et al., 2017). RRH is a transceiver

that is connected to a BS via the wireless or wired in-

terface. Therefore, RRH can be regarded as a relay

node between the BS and the users. In this contribu-

tion, UAVs are employed as our RRHs.

On the other hand, each UAV can be equipped

with limited memory to cache useful contents for fu-

ture user requests, in order to improve the Quality

of Service (QoS). This is referred to as mobile edge

caching (Wang et al., 2017) and machine learning al-

gorithms could be employed to predict user requests

and to perform caching updates. We model the user

preference as a multinomial process based on the La-

tent Dirichlet Allocation (LDA) algorithm (Blei et al.,

2003), which depends on the User-Topics probabil-

ity density function (PDF) and the Topic-Words PDF.

The user preference is simulated according to the 20-

Newsgroups dataset (Lang, 1995) and it is learned by

the LDA algorithm. The learned user preferences are

clustered by the K–means algorithm for the users-to-

Nooh, H., Zhu, Z. and Ng, S.

Machine Learning Assisted Caching and Adaptive LDPC Coded Modulation for Next Generation Wireless Communications.

DOI: 10.5220/0009840500670076

In Proceedings of the 17th International Joint Conference on e-Business and Telecommunications (ICETE 2020) - DCNET, OPTICS, SIGMAP and WINSYS, pages 67-76

ISBN: 978-989-758-445-9

UAVs allocation. When the user preference is accu-

rately predicted, useful data can be cached. If a user

request is in the UAV’s cached memory, it could be

sent directly to the user from the UAV, without fur-

ther requests to a remote BS. Thus, the user request

delay is reduced.

Low Density Parity Check (LDPC) codes (Gal-

lager, 1963; Guo, 2005) are powerful forward er-

ror correction schemes that have been widely investi-

gated and are considered in the 5G standard. Adaptive

coding and modulation (L. Hanzo, S. X. Ng and T.

Keller, 2005) is another attractive transmission tech-

nology, where a high-rate channel code and a high-

order modulation scheme are employed, for increas-

ing the transmission rate, when the channel quality

is good. By contrast, a low-rate channel code and a

low-order modulation scheme are employed, for im-

proving the transmission reliability, when the chan-

nel quality is poor. In this contribution, an Adaptive

LDPC Coded Modulation (ALDPC-CM) scheme was

investigated and utilized for the UAV-user communi-

cation link. The ALDPC-CM scheme could provide

a near-capacity transmission rate for a given channel

Signal-to-Noise-Ratio (SNR). Hence, a communica-

tion link with high SNR could lead to a lower trans-

mission period (or transmission delay).

The rest of this paper is organized as follows. The

caching model is investigated in Section 2, while our

system model is detailed in Section 3. Our simulation

results are discussed in Section 4, while our conclu-

sions are summarized in Section 5.

2 CACHING MODEL

In this section, the Latent Dirichlet Allocation algo-

rithm is outlined and the user preference model is pre-

sented.

2.1 Latent Dirichlet Allocation

Latent Dirichlet Allocation (LDA) is a generative

probabilistic model (Blei et al., 2003) that character-

izes the document generating process with the graph-

ical model. The grey parameters shown in Fig. 1 are

latent variables and cannot be observed. The remain-

ing variables could be observed with the input data.

LDA algorithm iteratively simulates the generating

process and estimates the latent variables. The gen-

erated data is then evaluated using a cost function.

Additionally, the latent variables are iteratively opti-

mized based on the cost function. Once the model

has converged, it can be used to predict the content of

future user requests.

The LDA model is able to reveal the topic compo-

sition of a document and the word probability that is

related to each topic. It classiﬁes the new document

into different classes and predicts the new words. It

is useful for our system since users that share the

same interest may be gathering around the same ge-

ographical location under speciﬁc application scenar-

ios. LDA is a Bag-of-Words (BoW) model that ne-

glects the order of words in the document. In our case,

the order of user requests is not important. Hence,

the LDA algorithm can be employed to perform user

request prediction. More explicitly, LDA is com-

posed of document generation and parameter estima-

tion, where its graphical model is shown in Fig. 1.

The meaning of symbols used in Fig. 1 are presented

in Table 1. In our context, a ‘document’ represents

the ‘content requested by a user’.

We assume that there are a total of |T | = N

top-

ics, e.g. cars, movies, weather, news, etc, and they

are controlled by Dirichlet distribution Dir(α). For

each user, their interest topics can be regarded as inde-

pendent identical distributed (i.i.d) random variables

sampled from Dir(α). For example, Bob (Bob ∈ C )

is interested in 40% of news, 10% of cars, 50% of

movies, and 0% for the rest. Then, we can com-

pute the User-Topics distribution of Bob, θ

(d)

. On

the other hand, we have Topic-Contents distribution

of each topic φ

for t ∈ T . These distributions are con-

trolled by another Dirichlet distribution Dir(β). Thus,

we can sample from θ

and φ

to generate a content

(d)

for Bob. We continue to do this until we have

generated all N

contents for each d ∈ C . Likewise,

this model can generate contents for each user. In gen-

eral, the probability that a word blank w is ﬁlled by a

term t is given by:

p(w = t) =

∑

(k)

(w = t|z = k)θ

(d)

(z = k),

(1)

where

∑

(d)

(z = k) = 1.

2.2 User Preference Model

To simulate the user request behaviour and to gener-

ate the related distributions, we performed an LDA

clustering over a text dataset called 20-Newsgroups

dataset, which was ﬁrst introduced in (Lang, 1995).

It is a popular dataset for experiments in text appli-

cations, which is composed of 20 groups of news.

Fig. 2(a) shows the main topics of the dataset. We

need to preprocess the dataset to ﬁlter out the stop

words before the LDA algorithm is employed. Then,

we extract n

= 1000 signiﬁcant words from the

dataset to compose a dictionary. Fig. 2(b) shows

the preprocessing of the dataset. Hence, once we

DCNET 2020 - 11th International Conference on Data Communication Networking

Dirichlet Dirichlet

Cat

(d)

t ∈ T

d ∈ C

i ∈ 1, 2, ..., N

(d)

Figure 1: Graphical notation of the LDA algorithm. Each box in the diagram is a ‘for’ loop. From left to right: Dir(α) is

a Dirichlet distribution with parameter α, it generates the topic distribution for each document; θ

(d)

is sampled from Dir(α)

and it represents the topics associated with each document. We then sample from θ

(d)

to get a topic τ

(d)

for the word w

(d)

From right to left: the words associated with each topic φ

is sampled from another Dirichlet distribution with parameter β;

we can then generate the word w

(d)

according to the sampled topic τ

(d)

and the word distribution φ

. Note that the grey circle

represents the variables that are latent, N

is the word count of the d-th document. C is the category of the documents and T

is the set of topics.

Table 1: Symbol meaning of the LDA algorithm shown in Fig. 1.

Symbols Contents Generation Documents Generation

C User groups Document categories

d ∈ C User Document

i ∈ 1,2,...,N

i-th content blank of d i-th word blank of d

Dir(α) Generate User-Topics distributions Generate Document-Topics distributions

Dir(β) Generate Topic-Contents distributions Generate Topic-Words distributions

(d)

User d’s User-Topics distribution Document d’s Document-Topics distribution

(d)

i-th content blank’s topic of d i-th word blank’s topic of d

Topic t’s Topic-Contents PDF Topic t’s Topic-Words PDF

(d)

i-th content of d i-th word of d

W Contents Space, w ∈ W Dictionary, w ∈ W

have ﬁnished the transform, we should have a n

×n

sparse matrix, which is 11314 × 1000 for this speciﬁc

case. The sparse matrix is then input to the LDA al-

gorithm.

We consider N

= 100 users, N

= 4 topics and

= 100 words (contents), as well as N

= 4 UAVs

to provide services to the users. The parameters α and

β for Dirichlet distribution are generated based on the

20-Newsgroups dataset. We assign a Topic–Contents

PDF for each topic, and a User–Topics PDF for each

user. Fig. 3(a) shows the Topic–Contents PDFs. It

is clear from Fig. 3(a) that different topics share dis-

tinct top ﬁve words, which are marked with red points.

Fig. 3(b) shows the topic distribution for the ﬁrst 20

users. As can be seen from Fig. 3(b), some users have

a dominant topic amongst the four topics.

The preferred contents by each user may be

different. The user preference is determined by

the User-Topic Probability Density Function (UT-

PDF) and Topic-Content Probability Density Func-

tion (TCPDF). For N

= 100 users we have 200 PDFs

as parameters to characterize the user preference.

Again, these parameters are generated according to

the 20-Newsgroup dataset. Fig. 4 shows the user pref-

erence with their dominant topics.

K-means clustering is then performed to match

each user to a UAV, in order to optimize the sys-

tem QoS, which is characterized mathematically as

a function of delay (τ

, i ∈ 1,2,...,N

). The delay

function for the ith user is determined by two terms,

namely the time to transmit the content to the user

from the UAV (τ

) and time to request the content (un-

available at the UAV) from the remote BS (τ

). For a

given request S

, the transmit delay τ

and request de-

lay τ

are given by Eq. (2) and Eq. (3), respectively.

= f (SNR,S

), (2)



0 S

∈ C

1 S

/∈ C

d ∈ 1, 2,..., N

, (3)

where C

is the contents cached in UAV d, which is

serving user i. For example, the request delay equals

to zero if the UAV has the user request in its memory.

Hence, the system delay for all users can be computed

as:

τ =

∑

i=1



+ τ



, (4)

which is a function of the receive SNR and the re-

quested content. Further analysis on this is carried

out in Section 4.2.

Machine Learning Assisted Caching and Adaptive LDPC Coded Modulation for Next Generation Wireless Communications

20 Newsgroups

Religon

Cryptog

raphy

Medicine

Electronics

Space

MAC_ hardware

Windows X

Graphics

Windows misc

Autos

PC_ hardware

Hockey

Motorcycles

Baseball

Rec Computer

Science

Politics

talk

Misc

Mideast

Guns

Misc

Christian

Atheism

Forsale

(a) The groups of the 20 Newsgroups dataset. The dataset

is composed of 20 newsgroups and they belong to 6 gen-

eral topics. There are a total of n

= 11314 documents

in the dataset. Besides, some of the groups are highly

related to each other (i.e., hardware for PC and MAC).

(b) The preprocessing of the input documents. We ex-

tract the ﬁrst 1000 high frequency words as our dictio-

nary. Then, we use this dictionary to transform each doc-

ument into a sparse vector, where the value of its n-th

element represents the number of repetitions of the n-th

word in the dictionary.

Figure 2: Composition and preprocessing of the 20-

Newsgroup dataset.

3 SYSTEM MODEL

3.1 Rician Fading Channel

The Rician fading channel models a dominant (non-

variant) component amongst the multipath fading

channel components (Goldsmith, 2005). This dom-

inant component can be used to model the Line-of-

Sight (LoS) link between the UAV and its user. The

Rician PDF is given by:

p(r) =

exp



−

+ r

2σ



, (5)

where A is the peak amplitude of the dominant signal,

is the zero-order modiﬁed Bessel function of the

ﬁrst kind, K =

2σ

is the Rician factor representing

the ratio of the dominant signal power to the multipath

0 20 40 60 80 100

0.1

0.2

0.3

0.4

0.5

Topic: 1

PDF

p(Top

)=0.98503

0 20 40 60 80 100

0.1

0.2

0.3

0.4

0.5

Topic: 2

PDF

p(Top

)=0.64332

0 20 40 60 80 100

0.1

0.2

0.3

0.4

0.5

Topic: 3

PDF

p(Top

)=0.94064

0 20 40 60 80 100

0.1

0.2

0.3

0.4

0.5

Topic: 4

PDF

p(Top

)=0.20008

Topic-word PDF

Words/ No.

Probability

(a) Words distribution for all four topics. The top ﬁve

frequent words are highlighted in red.

1 2 3 4

0.2

0.4

0.6

0.8

User: 1

1 2 3 4

0.2

0.4

0.6

0.8

User: 2

1 2 3 4

0.2

0.4

0.6

0.8

User: 3

1 2 3 4

0.2

0.4

0.6

0.8

User: 4

1 2 3 4

0.2

0.4

0.6

0.8

User: 5

1 2 3 4

0.2

0.4

0.6

0.8

User: 6

Probability

(b) Topic distribution for the ﬁrst 6 users.

Figure 3: Generated users Topic–Contents PDFs and User–

Topics PDFs. We assume n

= 4 topics and n

= 100 con-

tents.

variance. As K → 0 Eq. (5) approaches the Rayleigh

PDF:

p(r) =

exp



−

2σ



, (6)

which is the commonly used model to describe the

statistical time varying nature of the multipath fading

channel without a dominant component. In a similar

manner, as K → ∞ The Rician PDF approaches the

Gaussian distribution.

3.2 LDPC Codes

Low Density Parity Check (LDPC) codes (Gallager,

1963; D. J. C Mackay, and R. M. Neal, 1997) be-

long to the family of linear block codes, which are

deﬁned by a parity check matrix having M rows and N

columns. The column and row weights are low com-

pared to the dimension M and N of the parity check

matrix H. Fig. 5 shows the bipartite graph (M. R. Tan-

ner, 1981) representation of the parity check matrix.

DCNET 2020 - 11th International Conference on Data Communication Networking

0 50 100 150 200 250 300

X (m)

100

150

200

250

300

Y (m)

Users with associated dominant topics

Topic: 1

Topic: 2

Topic: 3

Topic: 4

Drone

0 50 100

Drone 1

0.1

0.2

0.3

0.4

0.5

Probability

0 50 100

Drone 2

0.05

0.1

0.15

0.2

Probability

0 50 100

Drone 3

0.1

0.2

0.3

0.4

Probability

0 50 100

Drone 4

0.01

0.02

0.03

0.04

0.05

0.06

Probability

Figure 4: The primary model with user preference. N

= 4 UAVs serving N

= 100 users within a 300m×300m area.

This graph constitutes of two types of nodes, namely

the message nodes, each of which corresponds to a

column of the parity check matrix, and the check

nodes, each of which corresponds to a row of the ma-

trix. There are lines connecting these two types of

nodes, and each connection corresponds to a non-zero

entry in the parity check matrix of Fig. 5. For exam-

ple, the non-zero entry at the bottom right corner of

the parity check matrix in Fig. 5 corresponds to the

connection between the 6

node on the left and the

node on the right.

N Message Nodes

1 1

0 0 0

1 1 0

0 0

(N−K) Parity Check Nodes

N−K

Figure 5: Bipartite graph representation of the parity check

matrix.

The number of information bits encoded by an

LDPC code is denoted by K = N − M, yielding a cod-

ing rate:

(7)

The modulation rate for an M-ary modulation is given

by:

= log

(M) . (8)

is the number of modulated bits per symbol, e.g.

for BPSK, QPSK, and 8PSK, we have R

equals to 1,

2 and 3 bits, respectively. Hence, the overall rate of

the coded modulation scheme (or information bits per

modulated symbol) can be computed as:

R = R

·R

·log

(M) . (9)

3.3 Adaptive LDPC Coded Modulation

Fig. 6 shows the Bit Error Ratio (BER) versus SNR

performance of an LDPC-coded 8PSK scheme, when

communicating over Rician fading channels. Table 2

depicts the corresponding parameters. The size of the

H is chosen to maintain a code rate of R

= 0.5. As

seen from Fig. 6, as the Rician factor K increases, the

BER performance improves because the channel be-

comes more Gaussian like. When K = 0, the channel

becomes a harsh Rayleigh fading channel. We have

chosen a Rician factor of K = 3 for the rest of our

simulations.

-6

-5

-4

-3

-2

-1

2 3 4 5 6 7 8 9 10

BER

SNR (dB)

K=0

K=1

K=3

K=10

Figure 6: BER versus SNR performance of LDPC-coded

8PSK scheme, when communicating over Rician fading

channels having a Rician factor of K = {0,1, 3, 10}.

Machine Learning Assisted Caching and Adaptive LDPC Coded Modulation for Next Generation Wireless Communications

Table 2: LDPC coding parameters used in Fig. 6.

Parameter Value

Maximum decoding iterations 15

Maximum Frame Size 100 000 bits

Column Weight 3

H Matrix Size 2400 × 1200

In order for a range of data rates to be supported

in the UAV system the rate can be manipulated such

that each modulation supports a throughput value pro-

gressively higher than the previous scheme, as out-

lined below. More speciﬁcally, the proposed ALDPC-

CM scheme supports 8 transmission modes as de-

tailed in Table 3. The BER versus SNR performance

of these LDPC coded modulation schemes have been

simulated when communicating over AWGN channel

(Fig. 7), Rician fading channel (Fig. 8) and Rayleigh

fading channels (Fig. 9).

Table 3: Parameters of ALDPC-CM modes.

Modulation

Coding

Rate

Parity Check

Matrix Size

(K × N)

Throughput

(bits/symbol)

BPSK 0.5 2100 × 4200 0.5

4QAM 0.5 2100 × 4200 1

8PSK 2/3 2100 × 3150 2

16QAM 3/4 2100 × 2800 3

32QAM 4/5 2100 × 2625 4

64QAM 5/6 2100 × 2520 5

128QAM 6/7 2100 × 2450 6

256QAM 7/8 2100 × 2400 7

-6

-5

-4

-3

-2

-1

-5 0 5 10 15 20 25 30 35 40

BER

SNR (dB)

BPSK R=1/2

4QAM R=1/2

8PSK R=2/3

16QAM R=3/4

32QAM R=4/5

64QAM R=5/6

128QAM R=6/7

256QAM R=7/8

Figure 7: BER performance of LDPC coded modulation

when communicating over AWGN channels K = ∞.

Based on the BER curves in Fig. 7, Fig. 8 and

Fig. 9, we extract the SNR thresholds for each of

these LDPC coded modulation schemes at a target

BER of 10

−4

, which are tabulated in in Table 4. The

results displayed on Table 4 are inline with our ex-

pectations based on our prior knowledge of the na-

ture of the respective channels. The AWGN channel

-6

-5

-4

-3

-2

-1

-5 0 5 10 15 20 25 30 35 40

BER

SNR (dB)

BPSK R=1/2

4QAM R=1/2

8PSK R=3/4

16QAM R=3/4

32QAM R=4/5

64QAM R=5/6

128QAM R=6/7

256QAM R=7/8

Figure 8: BER performance of LDPC coded modulation

when communicating over Rician fading channels K = 3.

-6

-5

-4

-3

-2

-1

0 10 20 30 40

BER

SNR (dB)

BPSK R=1/2

4QAM R=1/2

8PSK R=2/3

16QAM R=3/4

32QAM R=4/5

64QAM R=5/6

128QAM R=6/7

256QAM R=7/8

Figure 9: BER performance of LDPC coded modulation

when communicating over Rayleigh fading channels K = 0.

Table 4: SNR thresholds at BER = 10

−4

for various LDPC

CM schemes. The perfect coding scheme is based on the

channel capacity of the 256QAM scheme.

Modulation

Perfect

Coding

(dB)

AWGN

(dB)

Rician,

K=3 (dB)

Rayleigh

(dB)

BPSK -3 -0.899 -0.236 1.029

4QAM 1 2.092 2.775 4.043

8PSK 6 7.538 8.810 10.829

16QAM 9 10.806 12.332 14.847

32QAM 13 14.564 16.080 18.778

64QAM 16 17.568 19.346 22.413

128QAM 20 20.858 21.418 25.983

256QAM 23 24.026 25.556 29.200

is the best case for all the LDPC coded modulation

schemes followed by the Rician fading channel and

the Rayleigh fading channel. It is also worth pointing

out that the Rician fading channel (K = 3) is closer to

the ideal case of the AWGN channel than the worst

case (Rayleigh fading channel) across all schemes.

Fig. 10 shows a visual presentation of the data pre-

sented in Tables 3 and 4. As seen in Fig. 10, when ac-

tual coding (which has a discrete coding rate) is used,

DCNET 2020 - 11th International Conference on Data Communication Networking

the throughput has a staircase like curve. This is be-

cause when the received SNR is between two thresh-

olds, a lower coded modulation mode is activated. For

example, when the received SNR is above 12.332 dB

but below 16.080 dB (see Table 4), LDPC-16QAM

will be invoked when communicating over the Ri-

cian fading channels. Hence, a constant throughput

of 3 bits/symbol (see Table 3) will be yielded for

a received SNR that is above 12.332 dB but below

16.080 dB.

-20 -10 0 10 20 30 40

SNR (dB)

Capacity (bits/symbol)

Perfect Coding with Discrete Coding Rate

Perfect Coding with Continuous Coding Rate

ALDPC-CM under Rician fading channel, K = 3

ALDPC-CM under Rayleigh fading channel

ALDPC-CM under AWGN channel

Shannon Capacity (Ideal)

Figure 10: Capacity or throughput versus SNR curves of the

ALDPC-CM scheme in comparison with theoretical lim-

its. The Shannon capacity and the two ‘Perfect Coding’

schemes are based on AWGN channels.

4 RESULTS AND DISCUSSIONS

A simulator was created to investigate the overall sys-

tem performance of the proposed UAV based scheme,

incorporating machine learning assisted caching and

ALDPC-CM aided transmission. The mobile users

are served by UAVs, while the UAVs are connected

wirelessly to a Base Station (BS).

The user preferences are estimated by the LDA

algorithm, which classiﬁes the information (word by

word) into a set of topics. The users are then clus-

tered into N

groups and each user is allocated to a

UAV. This last step is carried out using the K–means

clustering algorithm. Three user-UAV allocation cri-

teria were considered:

1. max-snr: allocate user to the UAV that would

give the highest received SNR for minimizing the

transmission delay.

2. caching-efﬁciency: allocate user to the UAV that

is likely to have the requested content, for mini-

mizing the request delay.

3. min-delay: allocate user to the UAV that would

minimize the overall delay (the sum of both trans-

mission and request delay).

4.1 Caching Efﬁciency

In this section, we investigate the performance of

LDA based caching technique in comparison to a ran-

dom caching benchmark scheme. For most scenar-

ios, only a limited amount of historical user requests

can be accessed. Besides, it is also important to pro-

vide immediate QoS to users once the connection is

established. Hence, the system delay under limited

historical requests is a signiﬁcant performance indi-

cator. Here, we assume that each user is allocated to

a UAV that has a link with the highest received SNR

(i.e. max-snr criteria). Then, we only examine the ef-

fect of different caching strategies. A perfect coding

scheme with continuous coding rate is assumed in this

subsection.

The simulation results for LDA based caching and

random caching with different historical request sizes

are shown in Fig. 11(a), when the UAV’s memory

size is 50% of the total data (i.e. |C

| = 50). It

should be noticed that the x-axis in Fig. 11(a), which

is the size of the historical requests, represents the

number of past requests from each user. Not surpris-

ingly the caching efﬁciency is around 50% for the ran-

dom caching strategy. As shown in Fig. 11(a), the

LDA based approach achieves above 90% caching ef-

ﬁciency after only 5 historical requests.

Fig. 11(b) shows the inﬂuence of both memory

size and the number of historical requests. It can

be observed from Fig. 11(b) that the memory size

is more signiﬁcant than the number of historical re-

quests, when considering the caching efﬁciency. This

is because the LDA algorithm learns user preferences

relatively fast when the memory size is big, as shown

in Fig. 11(a). We have shown that the LDA based

caching scheme is signiﬁcantly more efﬁcient com-

pared to the random caching method. The conver-

gence speed of the LDA algorithm is also fast, espe-

cially when the memory size is big. Having consid-

ered the ideal case with perfect coding here, we will

now investigate the performance of the ALDPC-CM

based scheme in the following section.

4.2 Overall Delay

The simulation parameters used in this section are

given in Table 5. The ALDPC-CM scheme supports

the 8 transmission modes detailed in Table 3, which

have SNR thresholds shown in Table 4. Table 6 shows

the throughput achieved for each coded modulation

mode. Additionally, if the received SNR is lower than

the SNR threshold of the lowest mode (BPSK), then

no transmission (No Tx) will happen. By contrast,

when the received SNR is higher than the highest

Machine Learning Assisted Caching and Adaptive LDPC Coded Modulation for Next Generation Wireless Communications

1 5 10 15 20 25 30

Size of History Request (Each User)

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Target Rate

Request Target Rate Under Various Methods

Cache (LDA)

Cache (Count)

(a) Caching efﬁciency simulation of LDA based and ran-

dom caching strategies under varying size of historical

requests. The system has n

= 100 users, n

= 4 UAVs,

= 100 contents, and n

= 4 topics. A total of |C

| = 50

contents can be cached for each UAV.

(b) Caching efﬁciency simulation of LDA based and ran-

dom caching strategies under varying size of historical

requests and a range of memory sizes.

Figure 11: Caching efﬁciency simulation based on perfect

coding scheme that has continuous coding rate.

mode (256QAM), then a throughput of 7 bits/symbol

will be supported, as illustrated in Table 6.

Fig. 12 shows the instantaneous delay and mean

delay of the ALDPC-CM based scheme, for the three

user-UAV allocation criteria considered. The user-

UAV allocation constitutes a trade-off between avail-

able resources. As seen in Fig. 12, minimizing the

transmission delay (max-snr criteria) would give a

lower mean delay compared to minimizing the re-

quest delay (caching-efﬁciency criteria). Hence, the

transmission delay is more dominant than the re-

quest delay in our system. Furthermore, the min-

delay criteria gives the lowest mean delay as ex-

pected. However, this would require additional com-

putation/overhead for calculating the transmission de-

lay and the request delay.

Table 5: Simulation Parameters.

Parameter Value

No. of UAVs 4

No. of users 50

Total contents, n

100

Area served 300 m

Drone/UAV height 30 m

Noise PSD -85 dBm

Signal bandwidth 20 MHz

Carrier Frequency 2 GHz

UAV’s memory size, |C

| 50

UAV-BS Request delay 2 ms

Maximum user speed 15 m/s

Coded Modulation scheme ALDPC-CM

Table 6: Data rates for various LDPC-CM schemes.

Capacity

Interval

(bits/symbol)

Modulation

Coding

Rate

Throughput

(bits/symbol)

[−∞,0.5] No Tx 0 0

[0.5,1] BPSK 0.5 0.5

[1,2] 4QAM 0.5 1

[2,3] 8PSK 2/3 2

[3,4] 16QAM 3/4 3

[4,5] 32QAM 4/5 4

[5,6] 64QAM 5/6 5

[6,7] 128QAM 6/7 6

[7,+∞] 256QAM 7/8 7

Fig. 13 shows the instantaneous delay compari-

son for the ALDPC-CM based transmission scheme

when considering the three user-UAV allocation crite-

ria. The max-snr criteria may be considered as a good

tradeoff in terms of performance and complexity. Ta-

ble 7 depicts the mean delay values associated with

the instantaneous delay plots shown in Fig. 13 un-

der each criterion. The practical ALDPC-CM based

scheme only invoked an additional mean delay of

1.32 − 1.27 = 0.05 ms compared to the ideal perfect

coding based scheme when the min-delay criteria is

considered, as shown in Table 7.

Table 7: Mean delay comparison over 101 frames.

Criterion

Mean Delay

(ms),

ALDPC-CM

Mean Delay

(ms), Perfect

Coding

max-snr 2.02 1.27

min-delay 1.32 1.27

caching-efﬁciency 3.16 2.84

DCNET 2020 - 11th International Conference on Data Communication Networking

0 50 100 150 200 250 300

X (m)

100

150

200

250

300

Y (m)

User Assignment According to Maximum SNR

0 50 100 150 200 250 300

100

150

200

250

300

X (m)

Y (m)

User Assignment According to User Preference

0 50 100 150 200 250 300

X (m)

100

150

200

250

300

Y (m)

User Assignment According to Minimum Delay

SNR: 11.9 dB, Delay: 1.28 ms

Is Cached: 1, Mean Delay: 1.96 ms

SNR: 11.9 dB, Delay: 0.853 ms

Is Cached: 1, Mean Delay: 1.32 ms

SNR: 11.9 dB, Delay 1.28 ms

Is Cached: 1, Mean Delay: 3.08 ms

Figure 12: An instance of the moving user allocation process under three allocation criteria. The four UAVs are represented

by the four ﬁlled-markers, while users allocated to the UAV are denoted by the blank-markers of the same shape and color.

0 10 20 30 40 50 60 70 80 90 100

Frame Index

0.5

1.5

2.5

3.5

4.5

Delay (ms)

Instantaneous Delay with LDPC coding

Maximising SNR

Minimising Delay

Maximising Caching Efficiency

Figure 13: Instantaneous delay comparison for the ALDPC-

CM based transmission scheme when considering the three

user-UAV criteria.

5 CONCLUSIONS AND FUTURE

WORK

In this contribution, we have proposed an LDA based

machine learning technique for predicting user re-

quests, which is then investigated for improving the

caching efﬁciency of UAVs. It was evidenced in

Fig. 11 that the LDA based caching signiﬁcantly out-

performed the random caching method, where above

90% caching efﬁciency can be achieved after a short

training based on 5 historical user requests.

We have also implemented an ALDPC-CM

scheme that can support 8 transmission modes on top

of the no transmission mode, as outlined in Table 6.

The Rician fading channel having a Rician factor of

K = 3 was invoked for modeling the LoS link between

the UAV and its serving users. A range of simulations

were done for computing the BER versus SNR curves

of the various LDPC-CM schemes as shown in Fig. 7,

Fig. 8 and Fig. 9. An adaptive scheme was then cre-

ated for maintaining a BER of ≤ 10

−4

while increas-

ing the link throughput as the channel SNR improves,

as depicted in Fig. 10.

After the caching technique and the link-level

transmission were implemented, the K–means algo-

rithm was then utilized for allocating all users to

the UAVs. Three allocation criteria were consid-

ered, where the max-snr criterion aims to minimize

the transmission delay, the caching-efﬁciency crite-

rion aims to minimize the request delay and the min-

delay criterion aims to minimize the overall delay.

We found that the min-delay criterion would give

the lowest mean delay, while the max-snr criteria

is a good tradeoff when considering the achievable

performance and required complexity, as shown in

Fig. 13. The proposed practical ALDPC-CM based

scheme performed very close to the ideal perfect cod-

ing scheme, as evidenced in Table 7.

For the future work, other machine learning mod-

els such as the Non-negative Matrix Factorization

(NMF) should be comparatively analyzed against our

LDA-based approach for caching optimization. Other

channel coding schemes such as polar codes, trellis

codes and space-time codes can also be utilized in

the link-level transmission. Deep Learning and Re-

inforcement Learning techniques (Ha et al., 2019),

(Bruno et al., 2014) could also be considered for im-

proving the overall performance of the UAV based

system.

REFERENCES

Americas, G. (2018). 5g ultra-reliable and low-latency com-

munication (urllc) to digitize industries and unearth

new use cases. White paper, 5G Americas.

Blei, D. M., Ng, A. Y., and Jordan, M. I. (2003). Latent

Machine Learning Assisted Caching and Adaptive LDPC Coded Modulation for Next Generation Wireless Communications

dirichlet allocation. Journal of machine Learning re-

search, 3(Jan):993–1022.

Bruno, R., Masaracchia, A., and Passarella, A. (2014).

Robust adaptive modulation and coding (amc) selec-

tion in lte systems using reinforcement learning. In

2014 IEEE 80th Vehicular Technology Conference

(VTC2014-Fall), pages 1–6.

Chen, M., Challita, U., Saad, W., Yin, C., and Debbah, M.

(2017). Machine learning for wireless networks with

artiﬁcial intelligence: A tutorial on neural networks.

arXiv preprint arXiv:1710.02913.

D. J. C Mackay, and R. M. Neal (1997). Near shannon

limit performance of low density parity check codes.

Electronics Letters, 33(6):457–458.

Gallager, R. G. (1963). Low Density Parity Check Codes.

PhD thesis, M.I.T, United States.

Goldsmith, A. (2005). Wireless communications. Cam-

bridge university press.

Guo, F. (2005). Low Density Parity Check Coding. PhD

thesis, University of Southampton, United Kingdom.

Ha, C., You, Y., and Song, H. (2019). Machine learn-

ing model for adaptive modulation of multi-stream in

mimo-ofdm system. IEEE Access, 7:5141–5152.

L. Hanzo, S. X. Ng and T. Keller (2005). Quadra-

ture Amplitude Modulation : From Basics to Adap-

tive Trellis-Coded, Turbo-Equalised and Space-Time

Coded OFDM, CDMA and MC-CDMA Systems.

IEEE Press.

Lang, K. (1995). Newsweeder: Learning to ﬁlter netnews.

In Proceedings of the Twelfth International Confer-

ence on Machine Learning, pages 331–339.

M. R. Tanner (1981). A recursive approach to low complex-

ity codes. IEEE Transactions on Information Theory,

27(5).

Merwaday, A. and Guvenc, I. (2015). Uav assisted hetero-

geneous networks for public safety communications.

In Wireless Communications and Networking Confer-

ence Workshops (WCNCW), 2015 IEEE, pages 329–

334. IEEE.

Valavanis, K. P. and Vachtsevanos, G. J. (2014). Handbook

of Unmanned Aerial Vehicles. Springer Publishing

Company, Incorporated.

Wang, S., Zhang, X., Zhang, Y., Wang, L., Yang, J., and

Wang, W. (2017). A survey on mobile edge networks:

Convergence of computing, caching and communica-

tions. IEEE Access, 5:6757–6779.

Zeng, Y., Zhang, R., and Lim, T. J. (2016). Wire-

less communications with unmanned aerial vehi-

cles: opportunities and challenges. arXiv preprint

arXiv:1602.03602.

DCNET 2020 - 11th International Conference on Data Communication Networking