![](https://csdnimg.cn/release/download_crawler_static/86287466/bg1.jpg)
arXiv:1809.07857v1 [cs.NI] 19 Sep 2018
IEEE NETWORK MAGAZINE, VOL. XX, NO. YY, MONTH XXXX 1
In-Edge AI: Intelligentizing Mob ile Edge
Computing, Caching and Communication by
Federated Learning
Xiaofei Wang
1
, Yiwe n Han
1
, Chenyang Wang
1
, Qiyang Zhao
2
, Xu Chen
3
, Min C hen
4∗
1
Tianjin Key Laboratory of Advanced Networking (TANK), School of Computer Scienc e and Technology, Tianjin University, Tianjin, China.
2
Huawei Technology, Shenzhen, P. R. C hina.
3
School of Data and Computer Science, Sun Yat-sen University, Guangdong, Guangz hou, P. R. China.
4
School of Computer Science and Technology, Hua zhong University of Science and Tec hnology, Wuhan, Hubei, P. R. China.
*Corresponding author.
Abstract—Recently, along with the rapid development of
mobile communication technology, edge computing theory and
techniques have been attracting more and more attentions
from global researchers and engineers, which can significantly
bridge the capacity of cloud and requirement of devices by the
network edges, and thus can accelerate the content deliveries
and improve the quality of mobile services. In order to bring
more intelligence to the edge systems, compared to traditional
optimization methodology, and driven by the current deep
learning techniques, we propose to integrate the Deep Reinforce-
ment Learning techniq ues and Federated Learning framework
with the mobile edge systems, for optimizing th e mobile edge
computing, caching and communication. And thus, we design
the “In-Edge AI” framework in order to intelligently utilize
the collaboration among devices and edge nodes to exchange
the learning parameters for a better trainin g and inference
of the models, and thus to carry out dynamic system-level
optimization and application-level enhancement while reducing
the unnecessary system communication load. “In-Edge AI” is
evaluated and p roved to have near-optimal performance but
relatively low overhead of learning, while th e system is cognitive
and adaptive to the mobile communication systems. Finally, we
discuss several related challenges and opportunities for unveiling
a promising upcoming future of “In-Edge AI”.
Index Terms—Mobile Edge Computing, Artificial Intelligence,
Deep Learning
I. INTRODUCTION
With the increasing quantity and quality of rich multimedia
services over mobile networks, there has been a huge increase
in the traffic and computation for mobile users and devices
over recent years, which imposes a huge amount of workload
on today’s already- c ongested backbo ne networks and the mo-
bile networks.
Naturally, the emerging idea of Mobile Edge Computing
(MEC) is proposed [1] as a novel paradigm for easing the bur-
den of backbone n etworks by pushing the comp utation/storage
resources to the proximity of the User Equipments (UEs).
On the other hand, MEC circ umvents the long propagation
delays introduced by transmitting data from mobile devices
to re mote cloud computing inf rastructures, and hence is able
to support latency-critical mobile and Internet of Things
(IoT) applications. Specifically, edge nodes, i.e., base stations
equippe d with computation/storage capability, could de al with
the computation and content requests o f UEs, and conse-
quently this scheme improves the Quality-of-Service (QoS)
of Mobile Network Operators (MNOs) and the Quality-of-
Experience (QoE) of UEs an d relieves the load of backbone
networks, and pressure of clouds (data centers) [2].
Fulfilling th e requirement of QoE of UEs is not a trivial
even by virtue of M EC. The key and difficult po int lies in that
the computation o ffloading requires wireless data transmission
and might bring about the congestion of wireless channels, and
hence raises the decision making or optimization problem on
the whole communication and computation integrated system,
i.e., how to jointly allocate communication resources and
computation resources of edge nodes.
Seve ral pioneer works have been proposed and realize quite
good results in their assuming settings based o n convex opti-
mization, game theory an d so on [3] [4] . Nevertheless, con-
sidering the particular use cases in MEC, these optimiz ation
methods may suffer fr om the following issues: 1) Uncertain
Inputs: they assume the that some key in formation factors
are g iven as inputs, but actually some of them are difficult to
obtain due to variant wireless channels and privacy policies;
2) Dynamic Conditions: dynamics of the integrated commu-
nication and computation system are not well addressed; 3)
Temporal Isolation: most of them do no t consider the long-
term effect of current decision on re source allocation except
for Lyapunov optimization, viz., in a highly time-varying
MEC sy stem, most of proposed optimization algorithms is
optimal or close-to-optimal only for a snapsh ot of the system.
In a word, the pro blem existed in the resource allocation
optimization of the MEC system is “lack of intelligence” .
In view of the increasing complexity of mobile networks,
e.g., a typical 5G node is exp ected to have 2000 or more
configurable parameters, a recent new trend is to optimize
wireless communication by Artificial Intelligence (AI) tech-
niques [4] [5], include but not limited to the application of
AI for Physical Layer (PHY), Data Link Layer, and traffic
control [6]. Particularly, related studies on e dge comp uting
and caching, su c h as [7] [8], have shown that reinforcement
learning [9] (inclu de Deep Q-Learning [10] in Deep Rein -
forcement Learning) has the potential to be effective in joint
resource management. But [7] is based on Q-Learn ing, whic h
is not f easible for the practical ME C system where the state-