![](bg3.png)
suggest to correlate the so called Cellular Quality In-
dex (CQI)
1
with the achieved throughput rate at dif-
ferent locations of a map. The CQI is a parameter de-
rived from the different measured signal strength pa-
rameters (Reference Signal Received Power - RSRP,
Reference Signal Received Quality - RSRQ and Re-
ceived Signal Strength Indication - RSSI) to indicate
the overall channel quality. While driving a vehicle
Kelch et al. periodically executed the download of a
large file via the HTTP protocol using the TCP trans-
port layer protocol to fully utilize the present channel
and to measure the maximum achievable bandwidth.
The Channel Quality Indicator (CQI) was collected in
parallel by polling it each second from the used cellu-
lar modem via an AT-command.
In a similar way (Lu et al., 2015) send data for a
consecutive time of 60 seconds via the UDP protocol
to obtain active measurements of the currently present
HSDPA cell. They used Nexus 5 smartphones as mea-
suring devices. The CQI values were retrieved with a
high precision timely resolution of 2-8 ms by connect-
ing those smartphones to the so called Qualcomm eX-
tensible Diagnostic Monitor (QXDM)
2
framework.
As a result of their investigation both authors
suggest the Channel Quality Indicator to be a good
value to predict the achievable throughput. They state
this as possible, as an exact mapping of the Chan-
nel Quality Indicator onto the reserved resources and
the used modulation scheme of the celltower is pos-
sible. Those two parameters than directly specify the
granted available bandwidth. Such an exact mapping
however is not possible any more in the more ad-
vanced 4G/LTE networks
1
, which are investigated in
our work. Furthermore the achievable network speeds
in 4G networks are far larger than the speeds obtained
by Kelch et al. and Lu et al., being limited in the peak
to only 7,2 Mbit/s. This makes the predictability more
difficult as larger variations in the achievable through-
put rate are possible. We argue that although we ex-
pect the CQI as well to be an important parameter
in the throughput prediction, it should not be consid-
ered as sufficiently enough. Thus we investigate fur-
ther parameters (see Section 3) in our personal work,
which describe the overall network quality and can
be stored within the connectivity map as well. The
work of (Lu et al., 2015) further relies on the highly
accurate CQI values obtained via the QXDM toolset.
This hinders an actual large scale deployment, as such
software requires expensive licenses and might not be
easily deployable on low cost hardware, as used in our
work.
(P
¨
ogel and Wolf, 2012) also investigate the pos-
1
www.sharetechnote.com/html/Handbook LTE CQI.html
2
Diagnostic Monitor. http://goo.gl/ibV7g1
sibilities of a connectivity map to predict certain net-
work parameters in the vehicular context. This in-
cludes the next cell tower, to which a future handover
of the current connection will be performed, as well
as the future to be experienced bandwidth. In a fur-
ther work (P
¨
ogel and Wolf, 2015) this gathered infor-
mation is than used to improve a variety of network
services like adaptive video streaming and the han-
dover between different network technologies (2G,
3G networks). Similiar to (Kelch et al., 2013), (P
¨
ogel
and Wolf, 2012) collected active measurement data
from a productive HSDPA network by performing
drive tests. Furthermore both approaches only rely on
historical collected data without taking into consid-
eration currently measured quality parameters of the
moving vehicle. (P
¨
ogel and Wolf, 2012) stated this as
a future work in their research, however did not fur-
ther investigate it in (P
¨
ogel and Wolf, 2015).
In conclusion of this section it can be summarized
that the connectivity map is capable to improve the
overall network experience of vehicles by leveraging
its historic data to plan future network transmissions
accordingly. However due to the high timely fluctu-
ation of the overall network quality as described by
(Kamakaris and Nickerson, 2005), we argue that the
data stored within the connectivity map should not
be considered sufficient enough. In addition instanta-
neous measurements of the currently experienced net-
work quality should be taken into account, too. A va-
riety of contributions, which only rely on those instant
values, are presented in the following Section 2.2.
2.2 Online Throughput Estimation
As one of the earlier contributions within the field of
instantaneous throughput prediction (Xu et al., 2013)
develop PROTEUS, a system interface, which col-
lects instantaneous network performance parameters,
such as throughput, loss rate and one way delay of 3G
networks to predict the future network performance.
The authors did not investigate the performance of
their approach regarding the more advanced 4G/LTE
networks as it is the case in our contribution. PRO-
TEUS relies on Regression trees to enable the im-
provement of services such as VoIP, video confer-
encing and online gaming. To realize its future net-
work predictions PROTEUS only relies upon the last
20 seconds of experienced network performance and
completely avoids any form of previous offline train-
ing.
(Liu and Lee, 2015) addressed this gap by rely-
ing on previously driven trace data of 3G networks
to train their online throughput estimation approach.
The estimator itself relied upon 60 average through-
Cellular Bandwidth Prediction for Highly Automated Driving
123