Journal Pre-proof

Intelligent multi-agent based C-RAN architecture for 5G radio

resource management

Jin Xu , Zbigniew Dziong , Yan Luxin , Zhe Huang , Ping Xu ,

Adnane Cabani

PII: S1389-1286(19)31487-2
DOI: https://doi.org/10.1016/j.comnet.2020.107418
Reference: COMPNW 107418

To appear in: Computer Networks

Received date: 6 November 2019

Revised date: 15 June 2020
Accepted date: 6 July 2020

Please cite this article as: Jin Xu , Zbigniew Dziong , Yan Luxin , Zhe Huang , Ping Xu ,
Adnane Cabani , Intelligent multi-agent based C-RAN architecture for 5G radio resource manage-
ment, Computer Networks (2020), doi: https://doi.org/10.1016/j.comnet.2020.107418

This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition
of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of
record. This version will undergo additional copyediting, typesetting and review before it is published
in its final form, but we are providing this version to give early visibility of the article. Please note that,
during the production process, errors may be discovered which could affect the content, and all legal
disclaimers that apply to the journal pertain.

© 2020 Published by Elsevier B.V.

Intelligent multi-agent based C-RAN architecture
for 5G radio resource management
a,* a b a c d
Jin Xu , Zbigniew Dziong , Yan Luxin , Zhe Huang , Ping Xu , Adnane Cabani
a
École de Technologie Supérieure, University of Quebec, Montreal, QC H3C 1K3, Canada
b
National Key Laboratory of Science and Technology on Multispectral Information Processing, School of Automation, Huazhong University of Science and
Technology, Wuhan 430074, China
c
Wuhan Junhui Technology Development Co., Ltd, Hubei, 430074, China
d
Normandie Univ, UNIROUEN, ESIGELEC, IRSEEM, 76000 Rouen, France
*
Corresponding author.
E-mail address: [email protected] (Jin Xu)

ABSTRACT

5G commercialized mobile networks are expected to arrive in 2020. Key research of 5G is to support massive
applications with spectral efficiency (SE) and low energy consumption. However, how to manage limited
spectrum for such a massive services demand is a challenge in 5G. To cope with the challenge, new network
architectures with dynamic spectrum sharing - are needed. This paper proposes a new intelligent cloud-based
radio access network (ICB-RAN) architecture for dynamic 5G radio resource management. This study also
develops a new multi-agent learning cloud assisted double auction (MLCADA) mechanism for dynamic
spectrum reusability. Finally, experiment results verify the performance of the proposed approach.

Keywords:

C-RAN, cloud computing, 5G network, multi-agent.

1. Introduction
It is expected in the next decade, mobile data traffic will increase more than 1,000 times together with significant growth of the
number of smartphone and tablet users. Already the number of mobile devices is bigger than the total world population. Therefore, to
meet growing user demands, mobile network operators are expected to increase capacity for the current generation of wireless
network. The fifth generation (5G) networks are expected to arrive in 2020 and they will provide a wide variety of high rate services
(i.e., downlink 300 Mbps and uplink 60 Mbps).
The 5G platform is seen as a global standard for a unification of the existing standards, e.g., Long Term Evolution-Advanced
(LTEA), High Speed Packet Access (HSPA), and Wireless Fidelity (WiFi). Some of the emerging features and trends of 5G networks
are: multi-tier dense heterogeneous networks (HetNets) [1], device-to-device (D2D) and machine-to-machine (M2M)
communications, densification of the heterogeneous base stations [2], cloud based radio access networks [3], integrated use of
multiple radio access technologies (RATs) [4], wireless network virtualization [5] [6], and massive MIMO [7]. Also, future wireless
networks are being devised with the vision of heterogeneity. Various multiple RATs in different RANs, including 3GPP families (e.g.,
UMTS, HSPA, LTE) and IEEE ones (e.g., WiFi, WiMAX), are being integrated and jointly managed. 5G networks should provide a
better quality of experience (QoE) for users that include better quality of service (QoS) [8], longer battery lifetime [9], and lower cost.
To achieve this objective, various RATs will cooperate to cost-efficiently meet user needs and preferences.
In future 5G heterogeneous wireless networks, users expect to be always connected and served through a RAT [10] that fulfills their
requirements. The 5G networks will have a multi-tier architecture consisting of macro cells, different types of licensed small cells and
D2D capability to serve users with different QoS requirements in a spectrum efficient way. However, how to meet such massive
service demands with limited spectrum resource is a challenge. Normally, there are three main solutions [11]: increasing the small
RAN deployment; efficient spectrum utilization; offering more radio spectrum. However, increasing the small RAN and offering
more radio spectrum are costly solution. Because the radio frequency spectrum is a costly and finite resource, and to increase the
wireless networks infrastructure is also expensive [12]. More efficient spectrum allocation is a solution for improve spectrum
utilization of next generation wireless network. The key to meeting the future networks’ requirements on spectrum, power and cost
1
efficiency is to use spectrum efficiently. Therefore, increasing spectrum utilization is an economic and efficient way. Dynamic
spectrum access can efficiently improve spectrum utilization [13]. Incorporating dynamic spectrum management (DSM) functionality
with the RANs’ attributes can be an attractive approach to the considered. However, traditional radio resource management
technology cannot manage the increasing demands from both users and providers side. Therefore, in recent years, the research is
focused on studying different RATs and on trying to provide the best service experience to mobile users with modern technologies
such as cloud computing, mobile access, and seamless service. There are also several works trying to apply distributed resource
allocation and management to 5G heterogeneous networks. But managing a wireless system and offering access services to consumers
is a highly complex task that requires technological, regulatory, as well as economic insights. In particular, given a set of exogenous
constraints, the objective is to make sure that resources are utilized efficiently so that user demand can be met at reasonable costs with
required QoS. Furthermore, cloud computing and wireless virtualization technologies also are expected to be widely used in future
networks. Unfortunately, traditional RAN is not flexible and can hardly keep up with the new challenges.
The rest of the paper is organized as follows: In section 2, we review the background and related work. The architecture of
ICB-RAN is defined in section 3. Section 4 introduces the intelligent spectrum sharing. Section 5 describes the problem formulation
and auction design goals. Section 6 introduces the MLCADA mechanism. Section 7 presents the spectrum auction algorithm.
Simulation results and discussions are presented in section 8. Finally, we summarize our work in section 9.

2. Background
Cloud based RAN (C-RAN) proposed by China Mobile Ltd. (CMCC) [14] is an active research topic with many research studies
carried out in recent years. C-RAN offers new architecture characteristics and radio access technologies for the future wireless
network. It is expected to cope with many new challenges for mass end-user’s services growing. C-RAN architecture has been
proposed as a cost-efficient way of deploying small cells [15]. Figure 1 depicts the conceptual architecture of C-RAN.

Mobile Backhaul
Network

BBU Pool BBU 2 BBU n

BBU 1 BBU 3

Fronthaul

RRH n
RRH 2

RRH 1 RRH 3

User
Fig. 1. The architecture of C-RAN.

In the future 5G network, there will have a massive variety of devices with different characteristics and needs. Some of the most
mentioned use cases of 5G are IoT and Mobile Broadband. All these new use cases require different types of features of the network,
such as mobility, charging, security, policy control, latency, reliability, etc. In traditional RANs architecture, the radio units are
situated together with baseband units (BBUs). The C-RAN concept split Base Station (BS) to BBUs and remote radio heads (RRHs).
It moves the BBUs to a high-performance hosting datacenter with DSP processors and provides high-bandwidth optical fiber
front-haul transport network linking to the RRHs. Therefore, BBUs and RRHs are separated from each other in C-RAN. The BBUs
are located at central cell sites, and the RRHs are located at cell sites. These separately located units are connected by one port circuit
front-haul interface. Normally each RRH has two antennas and one LTE channel with 20MHz bandwidth while the capacity of the
front-haul interface is 2.45Gbps. C-RAN is more flexible and efficient than RAN. Currently, many operators use C-RAN architecture.
Compared to 4G, C-RAN in 5G can offer:
• High radio speed of 20Gbps/UE - New RAT, mmWave, massive multiple-input multiple-output (MIMO), massive aggregation.
• Low latency - Tectile Internet, autonomous driving, remote-controlled machine.
• Massive devices connectivity- IoT devices connection.
C-RAN should also improve the spectrum utilization with new types of spectrum sharing models, such as dynamic spectrum
sharing. However, how to manage the spectrum requirements with massive devices connectivity, low latency, enormous capacity and
high reliability is a challenge for the future 5G system. Intelligent radio resource management can play a significant role in addressing
this challenge. In this paper, we propose ICB-RAN architecture with an intelligent mechanism for future heterogeneous wireless
network dynamic spectrum sharing and offer a prototyping model.
Multi-agent systems (MAS) is a technology where an autonomous intelligent system incorporates several agents. Agents interact
with the environment and each other to optimize the system utility. Within a multi-agent system, the agent’s decision-making process
decomposes a complex and large system into several smaller agent systems and agents coordinate and communicate with each other to
make decisions. Application of multi-agent systems to the optimization problem can be significantly more efficient in comparison
with traditional approaches. Agent communication and learning are important features of MAS applications. Different agents have
different intelligence levels depending on their role and responsibilities for learning entity. Therefore, MAS based applications are
especially suited for complex large-scale systems with complicated relationships among multiple entities, such as 5G or multiple radio
access systems.
Future 5G wireless networks will be composed of different RATs, such as LTE, GSM, WiMAX and UMTS, to share spectrum. For
example, the 3.5 GHz band is allowed to be shared between military and commercial systems for small cell (Wi-Fi, LTE based)
services. For the wireless service providers are impossible to deploy all standards. Duo to the radio frequency spectrum is a costly and
finite resource, spectrum sharing among operators to offer high-speed access is a good option to enlarge available spectrum utility.
There are some questions of how to share spectrum. For example, how to divide spectrum and revenues into multiple subsystems that
managed by potentially competing providers. And how to share spectrum in different allocation units, such as IEEE 802.11 channels
and LTE resource frame’s blocks.
Spectrum auction is widely used and effective solution for wireless networks spectrum allocation. In [16], Xiang designed a truthful
and discriminatory pricing auction. Dong [17] proposed a combinatorial auction considering spectrum time–frequency division
reusability. Nevertheless, these auction mechanisms mentioned above are based on single channel. True-MCSA [18] extended
TRUST work in multi-channel auctions, which is spectrum reusability and satisfy the economic properties. In [19], Chouayakh
developed a truthful auction mechanism for dynamic allocation of LSA spectrum blocks for 5G. Devi [20] proposed a double auction
multi-winner framework for spectrum allocation in cognitive radio networks. However, it remains challenging to efficiently solve the
spectrum sharing problem when the number of RRHs is large. To cope with this challenge, we propose a cloud computing based
intelligent double auction approach. Through multi-agent optimization, the unused portion of spectrum is shared by spectrum dynamic
intelligent auction.

3. Architecture of ICB-RAN
The overview of ICB-RAN architecture is shown in Fig. 2. It contains four layers: physical radio access layer, intelligent cloud
computing resource pool layer, 5G core network layer and applications layer. In the first layer, there are several RRHs. These low
energy consumption RRHs are not different from C-RAN’s that shown in Fig.1. They are represented by cognitive agents in the
decentralized virtual BBU pool to meet user demand. Communications techniques inside the surrounding wireless area networks are
UMTS, GSM, LTE, WiMAX, and so on. However, the second intelligent cloud computing resource pool layer is different from
C-RAN’s. The virtual BBU pool in ICB-RAN is intelligent radio access controller, which use cloud-computing based multi-agent
systems processing techniques to allocate spectrum resource between RRHs and BBUs. BBUs are virtualized with cognitive agents
help. The provider-specific BBUs that have been deployed are replaced by standard servers. With intelligent virtualization technology
help, the standard servers which equipped with general-purpose processors (GPPs) process RAN functions and services fast.
Cognitive agent’s interaction is a primary aspect of the proposed architecture. The communication bandwidth needed for agent’s
interaction is not a critical issue because it occurs in the cloud computing platform. The cognitive agents in cloud optimize system
performance by collecting the statistics from the virtual BBUs. They also contribute to the dynamic spectrum allocation across RRHs
by exchanging information among neighbors. Through cognitive agents with intelligent virtual BBU pool corresponding, the desired
RRHs share spectrum resources from other RRHs in the neighbor.
From the system level view, the physical multiple RRHs are located within a geographic area connected with virtual BBUs. They
are mapped into a virtual intelligent decision network in the cloud, where the decision makers are cognitive agents. Each cognitive
agent in cloud corresponds to a virtual BBU particularly. Specifically, the cognitive agent reports corresponding to BBU’s individual
information (e.g., preference, location, traffic load) to the cloud. At the same time, agents interact with other agents within the cloud.
These interactions include requirements and sharing information of spectrum. Cognitive agents act as radio resource coordinators with
each other in the cloud to deal with resource management. The cognitive agents utilize cloud computing based intelligent double
auction optimization approach to obtain the best dynamic spectrum sharing decision results. The intelligent double auction scheme is
discussed in next section. Then cognitive agents distribute the optimized results to spectrum users by virtual BBUs and RRHs.

Applications Layer

Physical Processor
5G Core Network
Layer

Intelligent Access Cognitive Agent 3

Controller Cognitive Agent 2

Intelligent
Cloud Cognitive Agent 1 Cognitive Agent n
Computing
Resource
Pool Layer

Virtual BBU Virtual BBU 2

Virtual BBU n
Pool

Virtual BBU 1 Virtual BBU 3

Fronthaul

RRH 2 RRH n
Physical Radio
Access Layer RRH 1 RRH 3
User

Fig. 2. ICB-RAN architecture.

Through cognitive agent optimization, all users’ spectrum requirements are processed in the virtual BBU pool. With intelligent
cloud computing processing technology, wireless service providers achieve high diversity gains and offer value-added services to
their users at the same time. Inter interference of RRHs is restrained in the virtual BBU pool. Through multi-agent control, they also
improve signal communications quality and SE performance of ICB-RAN. The activated RRHs adaptive to traffic volume are to
improve SE performance. The impact of interactive interference is globally balanced through multi-agent control to meet traffic load
demand intelligent. User’s authentication functions such as traffic demand and mobility management are implemented in cognitive
agents. The cognitive agents dynamically balance all spectrum resources for the purpose of accommodating more users with the least
amount of cost. For example, all RRHs are not always in heavy traffic load. A part of potential RRHs will be into sleep mode when
their traffic load is low. Thick and fast RRHs will be active to meet the huge spectrum resources demands when the massive traffic
load comes.
We consider each RRH consisting of a set of W = {1,…, W}, access points (APs). Each AP serves several users. There are set of N
={1,…, N}, channels available for the APs. It is assumed that each RRH chooses Ln channels for data transmission of the users. The
number Ln can be regarded as the load of each AP, which is determined by the number of active users’ traffic demands. The choice of
channels by APn as Pn = {c1, c2, . . . , cLn}, ci ∈ N , ∀ 1 ≤ i ≤ Ln. The number of possible chosen channel profiles of APn is:

( ) ( )
( )
(1)

Moreover, new features for 5G or newly released RAN functions/services will be easily installed/removed in ICB-RAN through
simply upgrade open application programming interface (API). Furthermore, this architecture allows sharing computing and spectrum
resources among BBUs, flood traffic in one BBU is easily diverted to other BBUs.
4. Intelligent spectrum sharing
For ICB-RAN, more advanced radio resource allocation and optimization techniques are required than C-RAN due to the powerful
virtual BBU pool. An efficient mechanism for implementing traffic load balancing will be the intelligent spectrum sharing. Resource
optimization significantly enhances the overall system performance and maintains satisfactory QoS and QoE in ICB-RAN. Spectrum
sharing has extraordinary challenges because the large-scale spectrum requirements processing among RRHs and delay-tolerant
packet traffic arrives in an unpredictable and burst mode. Normally, there are three main approaches: equivalent rate constraint,
Lyapunov optimization, and Markov decision processes (MDPs) to deal with the spectrum resource optimization problem. For the
equivalent rate constraint approach, it uses queuing theory or large deviation theory to convert the average time delay constraints into
equivalent average rate constraints. Lyapunov approach optimizes drift-plus-utility function. Bellman equation with stochastic
learning is used in the MDP approach. Unfortunately, these approaches are not real-time and intelligent.
In the intelligent cloud computing resource pool layer, each cognitive agent that corresponds to a virtual BBU acts as a player and
manages the spectrum in a decentralized fashion. Every cognitive agent allows other agents to share its spectrum by bidding and its
infrastructure in an implicit fashion. The cloud acts as a cloud auctioneer-agent (see Fig. 3). The agent who has spectrum to share acts
as a seller agent. The agent who has interesting about the spectrum acts as a buyer agent to compete with other buyer agents by
bidding.
Seller agent and buyer agent are learning agents. They are selfish to learn how they should bid to maximize its user’s expected
utility. Every agent does not care about how their decisions will affect others. The buyer agent can update its bid

Fig. 3. Intelligent double auction.

price in real time within a single auction before the spectrum trade is made, and he neglects the overhead associated with the bid. In
this way, the end user utilizes the agent bidding to get the spectrum resources that he needs. Meanwhile, Spectrum utilization is
improved.

5. Problem formulation
In this section, we first introduce the problem formulation of the multi-channel MLCADA for heterogeneous spectrum exchange.
Then we overview the economic properties of the auction mechanism and state auction design goals.

5.1. Problem formulation

In this paper, we consider the scenario where buyer agents (spectrum demanders) trying to buy spectrum resources from seller
agents (spectrum owners). A single round multi-unit double auction consists of buyer agents and seller agents. We assume that
each buyer agent could request multiple channels and each seller agent contributes at least one channel in a single-round double
auction and has the same maximum power levels. Let { } , denotes the set of buyer agents and
{ } denotes the set of seller agents. In the auction, the winning bidders and the payment are decided by third-party
which acts as the cloud auctioneer-agent. We assume that all bids are private, sealed-bid and collusion-free. Buyer agents and seller
agents submit their bids privately to the cloud auctioneer-agent without any information of others’ bids. All bidders have different
valuations of the channels, the buyer agents’ bids are channel-specific without collusion.
Let denotes the bid of seller . The bid matrix of all seller agents is denoted by ( ), and
( ) denotes the competing bid matrix without ’s bid. Similarly, let denotes the bid of buyer .
( ) denotes the bid matrix of all buyer agents, and ( ) denotes the competing bid
matrix without buyer agent ’s bid. ( ) is the bid vector of , where denote the bid of for seller agent ’s
channel. is the true valuation of for its channel. The true valuation of buyer agent for seller agent ’s channel is . The
valuation vector of buyer agent is ( ). In the auction, the bids of both buyer agents and seller agents may or may
not be equal to their true valuation. For buyer agent , , When ’s channel unavailable to , is zero. The number of
channels that seller agent will offer is (where ), and the number of channels that buyer agent will request is (where
 ). In every auction, the third part auctioneer determines the payment for seller agent and the price that buyer agent
should pay. may or not be equal to .
Therefore, the utility of the seller agent is:

( )
{ (2)

Where ( ) is the number of transaction channels of seller agent .

The utility of the buyer agent is:

( )
{ (3)

Where ( ) is the number of transaction channels of buyer agent .

In our model, we also assume that both seller agents and buyer agents cannot submit untruthful channel prices, and the number of
submitted channels must be truthful.

5.2. Economic properties and design goals

First, we define four main economic properties in double multi-channel spectrum auctions:

Definition 1: Truthfulness. If neither seller nor the buyer can obtain a higher utility gain by bidding untruthfully, the double auction
is truthful, i.e., setting or .
Truthfulness is important for our double multi-channel spectrum auction design to avoid market manipulation. In the untruthful
auction, some selfish bidders can obtain more interests by submitting untruthful bids. But when most bids are not truthful in the
auction, the auction result will be heavy overhead and the market itself may collapse. In the truthful auction, there is no market
manipulation in the dominant strategy for bidders. Therefore, the auctioneer can ensure spectrum allocation efficiently and fairness to
buyers.

Definition 2: Ex-post Individual Rationality. A double auction is ex-post individual rational if all sellers and buyers will get
non-negative utility. That means no winning seller is paid less than its ask-price and no winning buyer pays more than its bid price:

(4)

Definition 3: Ex-post Budget Balance. A double auction is ex-post budget balanced if the auctioneer’s profit is non-negative. That
means total revenues of the sellers not less than payments by the buyers.

∑ ∑ (5)

Ex-post budget balance ensures that the auctioneer will never run into deficit.

Definition 4: Economic Efficiency. A double auction is economically efficient if the auctioneer can maximize total utility with
complete all users’ information. According to the impossibility theorem [21], there is no mechanism can achieve all properties in any
double auction at the same time. Therefore, our auction mechanism design aim is to achieve high economic efficiency.

6. MLCADA mechanism
In this section, we introduce the challenges of multi-channel spectrum auction design and multi-agent systems (MAS) at first. Then
we present MLCADA mechanism for heterogeneous spectrum sharing.

6.1. Challenges of auction design

6.1.1. Frequency heterogeneity
Frequency heterogeneity is defined as different frequencies have different transmission ranges. According to the propagation model
recommended by ITU, only the center frequency of one spectrum band will affect the path loss between two nodes. The total path loss
is defined as:

( ) (6)
 Where is the total path loss in decibel (dB), is the frequency of transmission in megahertz (MHz), is the distance in meter (m),
is the distance power loss coefficient and ( ) is the floor loss penetration factor.
In our model, we consider that each spectrum owner has a wide range of frequencies. The bandwidth of each channel is narrow and
different center frequency. Because each channel path loss is different, interference relationships caused by frequency heterogeneity
among the buyers are different.

6.1.2. Truthfulness and market manipulation

VCG [22] and McAfee [23] double auction algorithms are both truthfulness. But truthfulness cannot be ensured if these algorithms
directly applied to heterogeneous spectrum allocation situation. For example, one buyer only can win one item in a McAfee
heterogeneous double auction, but it also can manipulate its bids to achieve higher utility.

6.2. Auction mechanism design

In our scenario, spectrum sellers are represented by seller agents, spectrum buyers are represented by buyer agents. Both seller
agents and buyer agents submit their bids to cloud auctioneer-agent in every bidding round. Spectrum resources (channels) are
maximized utilization by reallocated quickly and efficiently through double auction. No seller agent or buyer agent can obtain a higher
utility gain by cheating. In each bidding round, there is maximum accepted price for each channel. And there is also a
minimum accepted price (reservation price) , which equal to true valuation of the channel. Maximum price is set by
cloud auctioneer-agent. For seller agent or buyer agent, channel has a price range of [ ]. Any seller agent’s ask price
lower than will be ignored by the cloud auctioneer-agent. For the same reason, any buyer agent’s bid price higher than
will be ignored by the cloud auctioneer-agent as well. There is also a time range for channel bidding round.

6.2.1. Seller agent bidding strategy

Seller agent aims at maximizing his profit, so he always tends to submit the highest ask price to cloud auctioneer-agent. For
channel , the bidding round time is , bidding time is decreased from to zero.

 ( ) ( ) ( )

 ( ) ( )( ) 

( ) 

Where ( ) is seller agent ask price for channel at bidding time , which in the bidding time range . Each ask
price action is assigned with a probability ( ). It presents how channel is important for anticipated profit to the seller agent at
bidding time , and ( ) . is reward function, equals to the number of channels with satisfied QoS requirements. The ask
price is periodically updated by the obtained rewards.

6.2.2. Buyer agent bidding strategy

Buyer agent also aims at maximizing his profit by submitting the lowest bid price for interesting channel to cloud
auctioneer-agent in the bid round time . Bidding time is decreased from to zero.

( ) ( ) ( )

 ( ) ( )( )  

 ( ) 

Where ( ) is buyer agent bid price for channel at bidding time , which in bidding time range . Each bid price
action is assigned with a probability ( ). It presents how spectrum channel is important for anticipated profit to the buyer agent
at bidding time , and ( ) . is reward function, equals to the number of channels with satisfied QoS requirements. The
bid price is periodically updated by the obtained rewards.

6.2.3. Cloud auctioneer-agent determines auction price

In each channel bidding time unit, the seller agents and buyer agents determine ask and bid price, then send them to the cloud
auctioneer-agent. The cloud auctioneer agent sorts the ask price descending order and bid price in ascending order.
 If ( ( )) ( ( )), the lowest seller agent ask price less than or equal to the highest buyer agent bid
price, then the channel transaction price ( ) is determined by equation (10).

( ) ( ∀ [ ]( ( )) ∀ [ ] ( ( )))
(11)
Or:
( ) ( ( ) ( ) ( ) ( )) (12)

7. Algorithm design
MLCADA as a non-cooperative game is to reach the Nash equilibrium and all players try to maximize their own social utility.
Before discussing the algorithm, we define two definitions.

Definition 5: Social Utility. The utility of the player is cost utility for the channel. Every player tries to get maximal utility.
MLCADA mechanism is essential to maximize the following social utility:

(∑ ∑ )

Where ∑ is the utility of sellers and ∑ is the utility of buyers.

Definition 6: Intelligent Cloud assisted Spectrum Sharing Nash Equilibrium. A sharing strategy profile ( ) is a
pure Nash equilibrium if no player can get better utility by deviation from it. is the utility function of player . Because of the
limited interference range, it is expressed as ( ), where is the sharing strategy of player and is the sharing
strategy profile of the neighboring players of .

( ) ( ) 

With intelligent multi-agent learning cloud computing assistance, we try to find the sharing strategy algorithm with maximum
utility. The algorithm is described in Algorithm 1.
Algorithm 1: Multi-agent learning cloud assisted double auction (MLCADA)
Initialization: Set each seller agent and buyer agent a random sharing strategy ( ) .
1). Updating seller agent and buyer agent set:
Loop
For
Update every seller agent and buyer agent spectrum resource information (sell/buy) in real time during the auction.
End for.
Every seller agent and buyer agent bid or sell needed spectrum with maximum utility probability to the Nash Equilibrium in
double auction game strategy (seller agent or buyer agent use bidding strategy introduced above).
if then
( ( ) ( ) ( )) ( ( ) ( ) ( )) in decreasing
order.
( ( ) ( ) ( )) ( ( ) ( ) ( )) in increasing
order.
while ( ( )) ( ( )) do
calculate ( ) by equation (8).
end while
 With the intelligent cloud assistance find the best strategy action selection for each seller agent or buyer agent :

( )
( ) ( ( ( ( )))) 

( )
if ( ) ( ) then
end loop
Terminate the auction procedure.
2). Bid strategy update:
Each player agent updates its spectrum sharing strategy according to the last time success bid result.
( )
() ( )

3). Go to step 1).

8. Trial
In this section, simulations are conducted to evaluate the performance of the proposed MLCADA mechanism in ICB-RAN
architecture. Simulations are carried out in MATLAB with the help of TUGlab toolbox [24]. In physical radio access layer, we
consider a small cell network deployed in a square region. Where the square region is 300m × 300m, there are varying in number [10,
20, 30, 40, 50, 60, 70] small cells. The coverage distance of each small cell is 50m, and the interference distance is 100m. There are 6
channels available in the network. Each small cell can support 30 users that spread randomly in every area. They are represented by
cognitive agents in the intelligent cloud computing resource pool layer, which use MLCADA mechanism processing techniques to
allocate spectrum resource. All simulation results are averaged over 1000 times independent Monte Carlo simulations.

8.1. Convergence behavior

300
Latency of convergence of MLCADA
Latency of convergence (ms)

250
Latency of convergence of VCG
200

150

100

50

0
10 20 30 40 50 60 70
Number of small cells (N)

Fig. 4. The convergence speed comparison between the MLCADA and conventional double auction.

The latency of convergence of the MLCADA and the conventional VCG double auction [25] in the different cell number network
that the time needed for converging to stabilization are shown in Fig. 4.
• Convergence behavior, the time needed for converging to stabilization, which is the speed of convergence.
As been mentioned above, the results are averaged by 1000 independent trials of each network scale. The figure shows that the time
needed for converging to stabilization of the MLCADA are significantly less than the conventional VCG double auction at same size
network. Moreover, when the network size scales up from N = 10 to N = 70, the latency of convergence of the MLCADA is
significantly slowly increased than the conventional VCG double auction. The VCG mechanism is a generic mechanism to select an
outcome that for achieving a socially optimal solution. With intelligent cloud assistance, MLCADA finds the best strategy action
selection more efficiently. The results imply that the proposed MLCADA is especially suitable for large-scale networks.
8.2. Throughput performance
In Fig.5, the aggregate interference level [26] among the proposed MLCADA algorithm, best optimum NE and random selection at
different small cell numbers are compared.
Random selection, in which the auctioneer randomly distributes the buyers among the sellers and then determines the final winning
buyers as the same as VGC double auction.
The aggregate interference experienced at the primary receiver is the sum of interference from all the active interferer nodes
transmitting in the same frequency spectrum block.

-20 Bset optimum NE

Proposed MLCADA
Random selection
Aggregate interference level

-40

-60

-80

-100

-120

-140

-160
10 20 30 40 50 60 70
Number of small cells (N)

Fig. 5. The aggregate interference level comparison among the MLCADA, best optimum NE and random selection.

The aggregate interference experienced at the primary receiver is the sum of interference from all the active interferer nodes
transmitting in the same frequency spectrum block.
The figure notes that the aggregate level increases quickly with the network scale increases as we anticipate. The results imply that
the proposed intelligent approach performance is close to the optimum solution and significantly better than the random selection
strategy.

8.3. Spectral efficiency

We compare the spectral efficiency of the proposed ICB-RAN and the C-RAN (depicted in Fig. 1) with VCG double auction
mechanism at different small cell numbers.
• Spectral efficiency, the ratio of net data in bits per second (bps) divided by the bandwidth in hertz, which is expressed in the
format “bits per second per hertz”, abbreviated as b/s/Hz [27].
8
Spectral efficiency of conventional C-RAN
Spectral efficiency of ICB-RAN

6
Spectral efficiency (b/s/Hz)

0
10 20 30 40 50 60 70

Number of small cells (N)

Fig. 6. The spectral efficiency comparison between the ICB-RAN and conventional C-RAN.

( )
 Where is the data rate and B is the bandwidth required.
The compare results are shown in Fig. 6. It is noted that the spectral efficiency increases with the network scale increases as
expected. Furthermore, the companion results note that the spectral efficiency performance of the proposed ICB-RAN is better than
C-RAN at same network size.

8.4. Energy efficiency

3.0
Energy efficiency of ICB-RAN
Energy efficiency of conventional C-RAN
Energy efficiency (bits per Joule) 2.8

2.6

2.4

2.2

2.0

1.8

1.6

1.4

1.2
0 6 12 18 24

Hours of the day

Fig. 7. Energy utilization of the day.

• Energy efficiency, the ratio of unit power in bits per second, which is means how many the system capacity can achieve by one-unit
power consumed. It is expressed in the format “bits per joule” [28].

( )( ) ( )

Where is an independent parameter controlled by the network designer, N is the number of nodes randomly deployed onto the
surface of a sphere and q is the path loss exponent, cq is the number of cell with path loss exponent q.
In a day, energy efficiency is different from the hour, because the network traffic load is different. The results imply that the
proposed intelligent approach always is less energy consuming than conventional. This simulation results note that the performance of
the proposed ICB-RAN architecture with ICADA mechanism has significantly better performs than the conventional resources
management strategy. Figure 7 depicts that ICB-RAN offers 30% improvement in the overall energy efficiency a day than C-RAN
with VCG mechanism.

9. Conclusion and future directions

The wireless network has several changes such as multimedia-oriented and heterogeneous. The next generation wireless network
needs a revolutionary architecture while supporting the smooth evolution. This article proposes an intelligent cloud-based RAN
(ICB-RAN) architecture with a novel intelligent multi-agent learning cloud assisted double auction (MLCADA) mechanism for future
5G network. It achieves intelligent dynamic spectrum sharing, at the same time ensuring the spectral SE and low energy consumption.
The proposed mechanism provides an efficient method to manage limited spectrum for massive services demand in 5G network.
There are still some new challenges to be studied in the future. First, the mechanism is needed to prevent the users from reporting their
demands untunefully. Secondly, learning algorithms with imperfect information caused by the delay should be developed to achieve
robust spectrum access in the cloud. In the future, new intelligent spectrum sharing ideas will be trialed in both laboratory and live
network environments.

Conflicts of Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to
influence the work reported in this paper.
 Author Statement
Jin Xu : Conceptualization, Methodology, Software, Writing, Original draft preparation.
Zbigniew Dziong : Data curation.
Yan Luxin : Visualization, Investigation.
Zhe Huang : Supervision.
Ping Xu : Software, Validation.:
Adnane Cabani : Writing- Reviewing and Editing.

Acknowledgment
This work was supported in part by the projects of National Natural Science Foundation of China under Grant 61571207, and Hubei
Provincial Natural Science Foundation of China under Grant 2018CFA089.

References
[1] P. Demestichas, A. Georgakopoulos, D. Karvounas, K. Tsagkaris, V. Stavroulaki, J. Lu, C. Xiong, and J. Yao, “5G on the horizon: key challenges for the
radio-access network," IEEE Vehicular Technology Magazine, vol. 8, no. 3, Sept. 2013, pp. 47-53.
[2] M. Tehrani, M. Uysal, and H. Yanikomeroglu, “Device-to-device communication in 5G cellular networks: challenges, solutions, and future directions," IEEE
Communications Magazine, vol. 52, no. 5, May. 2014,pp. 86-92.
[3] White paper, “C-RAN: The Road Towards Green RAN,” C. M. R. Institute, http://labs.chinamobile.com/cran/wpcontent/uploads /CRAN white paper v2 5
EN.pdf, Oct. 2011.
[4] N. Himayat, S. ping Yeh, A. Panah, S. Talwar, M. Gerasimenko, S. Andreev, and Y. Koucheryavy, “Multi-radio heterogeneous networks: architectures and
performance," in International Conference on Computing, Networking and Communications (ICNC), Feb. 2014, pp. 252-258.
[5] I. Chih-Lin et al., “Toward Green and Soft: A 5G Perspective,” IEEE Commun Mag., vol. 52, no. 2, Feb. 2014, pp. 66–73.
[6] Agarwal, Satyam, Francesco Malandrino, Carla Fabiana Chiasserini, and Swades De. "VNF placement and resource allocation for the support of vertical
services in 5G networks." IEEE/ACM Transactions on Networking, vol. 27, no. 1, Feb. 2019, pp. 433-446.
[7] C.-X. Wang, F. Haider, X. Gao, X.-H. You, Y. Yang, D. Yuan, H. Aggoune, H. Haas, S. Fletcher, and E. Hepsaydir, “Cellular architecture and key
technologies for 5G wireless communication networks," IEEE Communications Magazine, vol. 52, no. 2, Feb. 2014,pp. 122-130.
[8] N. Bhushan, J. Li, D. Malladi, R. Gilmore, D. Brenner, A. Damnjanovic, R. Sukhavasi, C. Patel, and S. Geirhofer, “Network densication: the dominant theme
for wireless evolution into 5G," IEEE Communications Magazine, vol. 52, no. 2, Feb. 2014, pp. 82-89.
[9] Malandrino, Francesco, Carla Fabiana Chiasserini, Claudio Casetti, Giada Landi, and Marco Capitani. "An Optimization-Enhanced MANO for
Energy-Efficient 5G Networks." IEEE/ACM Transactions on Networking, vol. 27, no. 4, Aug. 2019, pp. 1756-1769.
[10] T. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang, G. Wong, J. Schulz, M. Samimi, and F. Gutierrez, “Millimeter wave mobile communications
for 5G cellular: it will work!" IEEE Access, vol. 1, 2013, pp. 335-349.
[11] S. Talwar, D. Choudhury, K. Dimou, E. Aryafar, B. Bangerter, and K. Stewart, “Enabling technologies and architectures for 5G wireless," in IEEE MTT-S
International Microwave Symposium (IMS), June 2014, pp. 1-4.
[12] Clarke, Richard N. "Expanding mobile wireless capacity: The challenges presented by technology and economics." Telecommunications Policy, vol. 38, no.
8-9, Sep. 2014, pp. 693-708.
[13] E. Hossain, L. B. Le, and D. Niyato, Radio resource management in multi-tier cellular wireless networks. John Wiley & Sons, 2013.
[14] C. M. R. Institute. C-ran: The road towards green ran white paper. Oct. 2011.
[15] B. Haberland, F. Derakhshan, H. Grob-Lipski, R. Klotsche, W. Rehm, P. Schefczik, and M. Soellner, “Radio Base Stations in the Cloud,” Bell Labs Technical
Journal, Vol. 18, No. 1, Apr. 2013, pp. 129–152.
[16] L. Xiang, G. Sun, J. Liu, X. Wang, and L. Li, "A discriminatory pricing double auction for spectrum allocation," in IEEE Wireless Communications and
Networking Conference (WCNC), Shanghai, China, 2012, pp. 1473-1477.
[17] M. Dong, G. Sun, X. Wang, and Q. Zhang, "Combinatorial auction with time-frequency flexibility in cognitive radio networks," in IEEE INFOCOM,
Orlando, FL, USA, 2012, pp. 2282-2290.
[18] Z. Chen, H. Huang, Y.-e. Sun, and L. Huang, "True-MCSA: A Framework for Truthful Double Multi-Channel Spectrum Auctions," in IEEE Transactions on
Wireless Communications, 2013, pp. 3838-3850.
[19] Chouayakh, Ayman, Aurélien Bechler, Isabel Amigo, Loutfi Nuaymi, and Patrick Maillé. "A truthful auction mechanism for dynamic allocation of LSA
spectrum blocks for 5G." In Network Games, Control, and Optimization, pp. 219-232. Birkhäuser, Cham, 2019.
[20] Devi, Monisha, Nityananda Sarma, and Sanjib K. Deka. "DAMW: Double Auction Multi-winner Framework for Spectrum Allocation in Cognitive Radio
Networks." In International Conference on Innovations for Community Services, pp. 266-281. Springer, Cham, 2020.
[21] R. B. Myerson and M. A. Satterthwaite, “Efficient mechanisms for bilateral trading,” J. of Economic Theory, vol. 29, no. 2, pp. 265–281, Apr. 1983.
[22] E. H. Clarke, “Multipart pricing of public goods,” Public Choice, vol. 11, pp. 17–33, 1971.
[23] X. Feng, J. Zhang, and Q. Zhang, “Database-assisted multi-AP network on TV white spaces: system architecture, spectrum allocation and AP discovery,” in
2011 IEEE DySPAN.
[24] M. Mirás, E. Sánchez, Juegos Cooperativos con utilidad transferible usando MATLAB: TUGlab, Servicio de Publicaciones da Universidade de Vigo, 2008.
[25] W. Vickrey, "Counterspeculation, Auctions, and Competitive Sealed Tenders," The Journal of Finance, vol. 16, pp. 8-37, 1961.
[26] A. Babaei P. Agrawal B. Jabbari "Statistics of aggregate interference in cognitive wireless ad hoc networks" Proc. ICNC pp. 397-401 2012.
[27] Hong, J. Wang, C. Wang, and J. Shi, “Cognitive radio in 5G: A perspective on energy-spectral efficiency trade-off,” IEEE Commun.Mag., vol. 52, no. 7, Jul.
2014, pp. 46–53.
[28] V. Rodoplu T. Meng "Bits-per-joule capacity of energy-limited wireless networks" IEEE Trans. Wireless Commun. vol. 6 no. 3 pp. 857-865 Mar. 2007.
Jin Xu received his Ph.D degree in Institut national des sciences appliquées de Rouen France in 2008. He is currently an Associate
Researcher in the Department of Electrical Engineering, École de Technologie Supérieure, University of Quebec, Montreal,
Canada. His research interests include wireless network and distributed system, multi-agent system, cloud computing.

Zbigniew Dziong received his M.Sc. and Ph.D. degrees from Warsaw University of Technology, Poland, both in electrical
engineering. After graduation, he worked at Warsaw University of Technology as an Assistant Professor. During this period, he
was on sabbatical at the Centre National d’Etudes des Telecommunications, Paris, France, and at the Department of
Communication Systems, Lund Institute of Technology, Sweden. From 1987 to 1997, he was with INRS-Telecommunications,
Montreal, Canada, as a professor. From 1997 to 2003, he worked for the Performance Analysis Department at Bell Labs, Lucent
Technologies, Holmdel, New Jersey, USA. Since 2003, he has been with the École de technologie supérieure (University of
Quebec), Montreal, Canada, where he teaches on both the undergraduate and graduate levels as a professor.

Yan Luxin received the B.S. degree in electronic communication engineering and the Ph.D. degree in pattern recognition and
intelligence system from Huazhong University of Science and Technology (HUST), Wuhan, China, in 2001 and
2007,respectively.He is currently a Professor with the School of Automation, HUST. His research interests include multispectral
image processing, pattern recognition, and real-time embedded systems.

1
 Zhe Huang received her master’s degree in University of Quebec in 2017. She is currently a doctoral student in the École de
Technologie Supérieure, University of Quebec, Montreal, Canada.

Ping Xu is senior engineer at Wuhan Junhui Technology Development Co., Ltd. His research interests include DSP, wireless
network, IoT, cloud computing.

Adnane Cabani is Associate Professor at Normandie Univ, UNIROUEN, ESIGELEC, IRSEEM. He earned a Ph.D. degree in
Computer Science in 2007 from the National Institute of Applied Sciences (INSA), Rouen, France. Since 2008, he has been with
the Normandie Univ, UNIROUEN, ESIGELEC, IRSEEM.