IEEE Computational Intelligence Magazine - February 2020 - 29

The second experiment is to predict the user count associated with a mobile station, which is cast as a regression problem. We use the Autoregressive Moving Average (ARMA)
model, SVM regression, and random forest regression as the
baseline algorithms. The Mean Square Error (MSE) values are
shown in Table I. The MSE of our proposed method outperforms our algorithms.
IV. Data-Driven Wireless System Design

In this section, we present two case studies, including video
adaptation and video prefetching, in which the mobile intelligence mined from wireless big data are utilized for system
design and optimization.

where j (t ) is the bitrate choice, ur (t ) is the averaged QoS of all
the mobile users within a base station, and d is the threshold
for QoS guarantee. We will predict the value of ur (t ) and then
utilize a heuristic method to increase or decrease the bitrate.
Game Theory Based Spectrum Allocation
Given the bitrate version j (t ) determined by the network
operator, we will utilize the game theory method for spectrum
allocation to solve the competition problem. In particular, due
to the limited spectrum channels, the selection of a specific
spectrum channel of one user may lead to interference to other
mobile users. Thus, each mobile user aims to maximize its individual utility. Given other users' spectrum selection strategies
a -n, user n aims to find a strategy a n that maximizes the utility:

A. Learning Based Video Adaptation System

DNN Based QoS Prediction and Bitrate Adjustment
The overall QoS of mobile users depends on several coupled
factors and can be estimated by our DNN based method. Within a time interval, there will be a set of mobile users requesting
spectrum channels for video consumption. The number of
mobile users, the spectrum allocation method, and the selected
bitrate version determine the QoS of all the mobile users. These
three factors are dynamic and coupled, leading to the bias of
QoS prediction. In this work, we use the DNN based method
to predict the overall QoS by using the historical information
mobile users' video consumption. The bitrate adjustment problem is given as below:
P1: max
s.t.

j (t )

(6)

ur (t ) 2 d,

(7)

P2: max u n(a n, a -n, j (t ), t ).

(8)

an ! M

We extract a subset of trace related to 7 base stations located
in downtown to drive our experiment. A real video sequence

TABLE I Regression model comparison.
ARMA

SVM REGRESSION

RF REGRESSION

OUR METHOD

3.04

3.17

3.01

2.85

Content Server

Base Station

Mobile User

Base Station

Mobile User

(a)
42
Online Learning
AMBA + SA

41
Average PSNR

We study the mobile social video sharing [28] with regards to
the legacy wireless infrastructure as illustrated in Fig. 4(a). The
content server supplies videos to mobile users via base stations.
Each video will be encoded into different resolution-level versions with diverse file size [29], [30]. If the content server
chooses the highest resolution-level version for mobile users,
they may achieve the best QoS under the good channel condition, e.g., fewer mobile users within the location of interest
[31], [32]. However, if there are more users connected to the
content server, the resolution-level should be decreased to
guarantee the fluent playback. As there are limited channels for
mobile users, the network operator should allocate mobile
users to different channels to reduce the interference. The challenges lie on three aspects: 1) The number of mobile users
within a region is changing quickly, which will affect the
bitrate selection for the content server. 2) Different mobile
users will consume a different amount of videos. Thus, the
occupied time of the channel allocated to the corresponding
mobile users is diverse. and 3) The channel allocation and
bitrate adjustment are coupled. For instance, a channel with
more mobile users, but less total video demand, can still enjoy a
satisfactory QoS.
We adopt a two-level optimization method to jointly solve
the spectrum and bitrate adjustment problem.

40
39
38
37
0.3

0.4

0.5
0.6
0.7
0.8
0.9
Number of Users (% Dataset)

1.0

(b)
FIGURE 4 Learning based video adaptation system: a) system framework; b) performance comparison with the baseline algorithm.

FEBRUARY 2020 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE

IEEE Computational Intelligence Magazine - February 2020

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - February 2020