A Survey of Self-optimization Approaches for HetNets

Abstract

Network convergence is regarded as the development tendency of the future wireless networks, for which self-organization paradigms provide a promising solution to alleviate the upgrading capital expenditures (CAPEX) and operating expenditures (OPEX). Self-optimization, as a critical functionality of self-organization, employs a decentralized paradigm to dynamically adapt the varying environmental circumstances while without relying on centralized control or human intervention. In this paper, we present comprehensive surveys of heterogeneous networks (HetNets) and investigate the enhanced self-optimization models. Self-optimization approaches such as dynamic mobile access network selection, spectrum resource allocation and power control for HetNets, etc., are surveyed and compared, with possible methodologies to achieve self-optimization summarized. We hope this survey paper can provide the insight and the roadmap for future research efforts in the self-optimization of convergence networks.

1. Introduction

In recent years, with the rapid popularity of smart/mobile devices, users can access network more frequently and conveniently, resulting in an explosion of mobile traffic. Relying only on a single wireless access network or radio access technology (RAT) would hardly satisfy the customers' traffic requirements. A variety of RATs are thus necessarily deployed for facilitating network convergence and simultaneously satisfying the customers' service requirements. Both the channel capacity and the radio coverage can be substantially improved by combining variant RATs and enabling the users to choose the optimal access network in a heterogeneous network (HetNet).

However, it may impose extra complexity on the network operators to maintain a seamless and heterogeneous service. Accordingly, self-organization is inspired in order to provide a promising solution for alleviating the upgrading capital expenditures (CAPEX) and operating expenditures (OPEX) in the large-scale HetNets. The users are thus enabled to dynamic adapt to the varying environmental circumstances without relying on the centralized control or human intervention [1].

The concept of self-organization mainly comprises four aspects, including self-configuration, self-optimization, self-healing, and plug-and-play. In this survey, we mainly focus on self-optimization, which is the critical functionality of self-organization. Mechanisms of self-optimization have been widely emphasized in HetNets for addressing the problems of mobile access network selection, spectrum resource allocation, and power control, etc. Furthermore, with the aid of various self-optimization approaches, the network throughput can be optimized by employing algorithms such as dynamic spectrum sharing and optimal power allocation in HetNets. Mathematical methods ofgame theory, combinatorial optimization, stochastic process and stochastic geometric theory, etc. may play an important role in implementing the self-optimization paradigms. Furthermore, in the existing self-organized methodologies, bio-inspired algorithms are among the most famous techniques and have already shown their advantages in optimizing the performance of HetNets [2, 3].

In this paper, we survey various aspects of HetNets such as network convergence, coexistence of multiple RATs and heterogeneous formats. The fundamentals of self-optimization for HetNets are first investigated, followed by surveying/comparing variant self-optimization algorithms such as dynamic mobile access network selection, spectrum resource allocation and power control, etc. Furthermore, we investigate various techniques that have been employed for implementing self-optimization in HetNets.

The remainder of this paper is organized as follows. In Section 2, we classify the HetNets into three types. The fundamentals of self-optimization are introduced in Section 3, followed by the detail of variant self-optimization algorithms presented in Section 4. In Section 5 we discuss the possible methodologies for achieving self-optimization. Finally the conclusions of this article are stated in Section 6.

 

2. Categorizations of HetNets

HetNet is a quite broad concept, which can be classified from the perspective of network types, radio access technologies and heterogeneous cellular formats.

2.1 Convergence of Different Network Types

In this subsection, we focus on the convergence of infrastructure networks such as Long Term Evolution (LTE) and infrastructure-less networks such as Wireless Sensor Network (WSN), as depicted in Fig. 1, in which each structure exhibits its pros and cons. For instance, WSN is featured by the characters of flexible deployment, whilst exhibits several disadvantages such as a short radio coverage, low throughout, and limited terminals' ability, etc. LTE networks, on the other hand, have larger radio coverage, higher throughout and better mobility robustness, whereas suffering from costly and complicated operation, administration and maintenance (OAM). It is therefore attractive to converge the aobve-mentioned architectures to facilitate a better service for the users. The authors in [4] gave an overview of architecture for the convergence of mobile cellular network and WSN by emphasizing the following key challenges.

Fig. 1.Convergence of LTE and WSN

2.2 Coexistence of Multiple RATs

As the rapid development of communication technologies, various wireless access networks (WANs) with different RATs, such as the second generation cellular system (GSM), the third generation cellular system, 3rd Generation Partnership Project (3GPP) LTE, IEEE 802.11 wireless standard, etc., have been deployed to facilitate an overlapped radio coverage, as illustrated in Fig. 2. Mobile devices such as smart mobile phones thus tend to support multiple RATs simultaneous for facilitating a ubiquitous and seamless wireless access. This tendency enables the customers to select a best WAN at any time. However, the operational costs of dynamic spectrum access control and distributed resource sharing make the wireless access much more complicated than ever. It is highly expected to employ a self-optimization approach to substantially reduce the operational cost in HetNets.

Fig. 2.Coexistence of Multiple RAT and Multiple cellular formats.

2.3 Coexistence of Multiple cellular formats

The traditional method for improving network throughput is performed by increasing the density of base stations (i.e., deploying more base stations in a given geographic area). However, this method does not always work well in practice due to the negative effect of serious inter-cell interference. An effective way for increasing network capacity is to overlap small cells (such as pico-cells, femtocells, and relay nodes) upon the traditional macro-cell coverage to enhance the cell-split gains (as shown in Fig. 2).

Currently, two-tier femtocell network has become a hot topic and attracted a lot of attentions by both the academia and industry. As a key challenge to it, cross-layer interference must be suppressed to make dynamic spectrum allocation and power control work well as a viable solution. Furthermore, self-optimization approach is likely to provide a potential way for dynamically allocating resource and effectively sharing the spectrum in a distributed manner.

 

3. Fundamentals of Self-optimization

Self-organization (SON), as exists in many branches of science such as biology, medicine, sociology, communication, etc, is originally regarded as an interdisciplinary and heterogenous research field that provides a promising solution for alleviating the upgrading CAPEX and OPEX in the large-scale HetNets. The most famous and successful SONs are wireless ad hoc and WSN, in which several well-known optimization algorithms have already been employed for improving the systems’ performance. Furthermore, 3GPP LTE/LTE-Advanced systems have already emphasized the self-X capabilities (e.g., self-configuration, self-optimization, and self-healing) as a critical feature for achieving the systems’ goals of robustness, reliability, scalability and power efficiency:

In this paper, we mainly focus on the self-optimization approaches in HetNets. The actual network environment is assumed to be time variant, thus the optimal network settings may not always be maintained. By employing self-optimization approach, the network states can always be maintained in an optimal or near-optimal condition by dynamically adjusting network parameters with minimal human intervention. Particularly, in HetNets, in which the network parameters become more complicated than homogeneous networks, the capability of self-optimization becomes more critical and indispensable. The critical role of self-optimization (which has been regarded as the main functionality of self-organization) in the HetNets will be analyzed in the following sections [3].

 

4. Self-optimization Approaches for HetNets

Self-optimization enables users to dynamically adapt to the varying environment in a decentralized manner in order to facilitate a cost-efficient performance optimization. Although self-optimization approaches has been widely studied in HetNets, the self-optimization approaches are still necessarily treated. In this section, we survey and compare variant optimization approaches that are mainly applied to the functionalities such as mobile access network selection, spectrum resource allocation and power control for HetNets (as shown in Fig. 3). Furthermore, variant approaches are compared in Table 1.

Fig. 3.Classification of the works on self-optimization approaches for HetNets.

Table 1.Comparison among variant self-optimization approaches for WAN selection, Spectrum allocation and power control.

4.1Self-Optimization Approaches for WAN Selection

WAN selection is necessarily performed in HetNets for facilitating a seamless service, whilst achieving traffic load balancing. Typical approaches for WAN selection are enumerated in follows.

1)Prediction Method

Prediction method is employed by mobile users to estimate channel information in order to enable an earlier response. The following mechanisms are emphasized in prediction method.

2) User-Centric Method

User-centric method emphasizes the initiative of users in order to satisfy their requirement preferably.

3) Cost-based VHO Method

This method can be designed based on the cost of possible target networks considering user/application's preferences for performing an optimal handover decision and network selection [9]. A cost function with weights of bandwidth’s and monetary’s costs can be defined in order to decide whether to perform a VHO or not. Furthermore, the cost function can be optimized by considering both service types and user preferences in HetNets. According to the cost function, the required delay and processing complexity can be reduced. In addition, the throughput for mobile terminals with multiple active sessions can be improved by employing the adaptive and intelligent VHO protocol.

4) Competition-based VHO Method

Although the optimal network usually allows limited users to access simultaneously, users may still try to select the optimal WAN to maximize their own reward, causing severe network-access competition among the users. Existing methodologies such as game theory can be employed as a ideal guidance to manage the competition and guide users to dynamically adjust their behavior according to the time-varying network conditions without relying on external control [10].

5) Intelligence-based VHO Method

In order to deal with the complicated conditions in practical network, it is necessary to make the handover decision more intelligent. A fuzzy Q-learning vertical handoff control strategy has been proposed to support the mobility management in vehicular HetNets [11], enabling an optimal handoff decision relying on the always-best-connected concept based real time learning. However, only throughput and delay aspects are taken into consideration in this strategy.

4.2 Self-Optimization Approaches for Spectrum Resource Allocation

A challenge arises in HetNets due to the limited spectrum resource and the ever increasing requirements of user services, inspiring effective spectrum allocation mechanisms that attracts extensive studies in recent literatures.

1) Graph-based Mechanism

In this mechanism, the channel assignment can be modelled as a problem of multi-coloring graph, in which each node represents a femtocell and each edge between two nodes represents the interference between two femtocells. A distributed and adaptive scheme comprising three steps for graphic radio resource allocation can also be developed for improving the fairness among users [12]: i) a graph-based model can be developed according to the received signal to interference and noise ratio (SINR) in femtocells, ii) an initial random-subchannel-allocation algorithm is performed, and iii) a tradeoff between system’s throughput and user’s fairness is achieved by executing an iteratively distributed adaptive-subchannels-allocation algorithm voluntarily in each femtocell.

2) Competition-based Mechanism

In a distributed resource-allocation mechanism, each mobile station tries to select the most unoccupied subchannels. With the help of a non-cooperative game, a Nash equilibrium subchannel allocation can be proposed for maximizing every mobile station's payoff as well as each networks' utility in a distributed fashion by considering the bandwidth allocation for various applications. Furthermore, a self-optimized bandwidth allocation for the new connections can be performed by using a bankruptcy cooperative game. In addition, based on the non-cooperative game, the bandwidth-allocation scheme can also be divided into network-level and connection-level self-optimized functionalities, where the former stands for a long-term viewpoint for different access network, but the latter is a short-term viewpoint for each new connection [14].

3) Cognitive Radio-based Spectrum Allocation Mechanism

Cognitive radio technologies have been employed in the femtocells by utilizing a self-organization approach to allocate the spectrum resource [15]. Methods of mixed integer-nonlinear-programming and decomposition can be utilized for processing the spectrum-sharing problem among macro/femtocells and achieving a lower user-blocking rate than using the coloring method. A four-step method for cognitive radio resource management can be performed: i) periodically sense the available spectrum hole, ii) select unoccupied resource block (RB) based on the received interference power, iii) allocate the unoccupied RB, and iv) extract other important parameters.

In order to address the tradeoff between the spectrum utilization and interference avoidance in HetNets, a partial spectrum reuse (PSR) scheme can be performed to leverage the spectrum reuse and interference mitigation between macrocells and small cells [16]. In PSR, small cells only partially reuse the system spectrum to reduce the intensity of interfering small cells, and consequently the inter-cell interference can be reduced. Furthermore, the energy efficiency of HetNets can also be improved by choosing the optimal spectrum reuse factor.

4) Intelligence-based Resource Allocation

Swarm intelligence inspired approaches can also be employed in a dynamic resource allocation mechanism for improving the spectral efficiency of an OFDMA system. An evolutionary algorithm called Particle Swarm Optimization (PSO), which is originally used to exchange information between biotic population, has exhibited an ideal benefit in distributing OFDMA subcarriers due to its capabilities of learning and forgetting in performing parameter control [17].

4.3 Self-Optimization Approaches for Power Control

Although low-power nodes can be deployed in an overlay fashion in HetNets, the appendant interference may substantially erode the systems’ capacity. Power control, which is regarded as an effective approach for mitigating the interference and maximizing system performance simultaneously, has been widely studied.

1) Competition-based Power Control Mechanism

A non-cooperative game can be employed by a distributed utility-based SINR adaption for implementing an effective power control [21], in which the utility function consists of an SINR-dependent reward and the interference can be manipulated. A scheme of link-quality protection can also be proposed for protecting the cellular user from the strongest femtocell interference. However, a lot of femtocell users cannot achieve their SINR target at Nash Equilibrium point due to the interference imposed by macrocell users. In order to alleviate the aforementioned condition, femtocell users can be divided into two classes, i.e., the qualified users, whose channel gain is lower than that of the interference link, and interference users. The number of users not achieving their SINR target can be reduced due to the removing of users in the severe radio conditions [22]. Furthermore, power control based on Stackelberg game can also be performed, where each user prices its power and sells it to subchannels under the total power constraint. The optimization problem can be solved by using a distributed power-bargaining algorithm and emphasizing self-organization paradigm in it. In addition, a distributed power-optimization scheme in femtocell network can also be proposed based on non-cooperative game in order to mitigate inter-tier interference and reduce the strongest femtocell power when cellular users cannot achieve their SINR target.

However, most of the existed power control games perform active user selection in a random manner, resulting in a lose sight of user-diversity. In the following, we first compare different user–pairing methods and find an optimal way in considering user-diversity gain to maximize the macrocell capacity, and then perform the power control game. The user-pair-scheduling-based power control game shows a superior performance compared to the traditional stochastic schedule strategy based power control game, as illustrated in Fig. 4 [24].

Fig. 4.Average macrocell capacity under schemes of no power control, traditional stochastic schedule strategy based power control and user-pair-scheduling-based power control.

2) Power Control in Cognitive Radio Environment

In cognitive radio networks, users in both macrocells and femtocells are assumed to have the cognitive capability and can dynamically sense/utilize the available licensed spectrum resource. Power control and spectrum allocation can thus be jointly performed in cognitive radio by using a decentralized radio resource allocation mechanism, which is part of the self-organization features. A three-step joint-resource-allocation scheme can be performed:

The above-mentioned approach can be regarded as a Stackelberg game, in which a gradient based iteration algorithm can be used for achieving the Stackelberg equilibrium. However, a joint resource-allocation scheme with imperfect channel information and under some constraints (e.g., minimum-rate and proportional-fairness constraint [19]) is necessarily performed, because it is practically hard to obtain a perfect channel state information. In order to avoid the local sensing degradation owing to the local shadowing phenomena, cooperative sensing mechanism can be performed by employing a consensus algorithm for aggregating the sense result in a decentralized manner.

In light of the fact that the secondary users in cognitive radio do not know the interference power imposed by primary users, the interference constraint can be formulated in a probabilistic framework. In order to make the overall non-convex problem solvable, the solution can be decomposed into two parts, i.e., i) find the optimal power allocation scheme for each pair of false alarm rate and sensing time, and ii) maximize the cost function with respect to false alarm rate and sensing time [20]. A tradeoff between the collaborative/sensing power and transmission power can also be achieved in this solution.

3) Base Station Sleep (BSS) Scheme

In contrast to the traditional power control scheme, in which base station power can be set to arbitrary value between zero and its maximum value, Base Station Sleep (BSS) scheme just supports two power levels corresponding to the active and sleep states. BSS scheme is regarded as an effective method for network energy saving by turning the BSs with small activities to sleep. In [21], the authors derived both the success transmission probability and energy efficiency under BSS strategies by formulating the optimization problems in the form of power-consumption minimization and energy efficiency maximization in HetNet. In order to further reduce the energy consumption, the benefits of both PSR and BSS scheme can be combined to improve energy efficiency in terms of two dimensions: spectrum reuse factor and BS active probability. The simulation results in Fig.5 show that the proposed joint PSR and BSS scheme outperforms the method relying on any individual scheme in terms of energy efficiency.

Fig. 5.The total heterogeneous cellular network energy cost of four conditions.

4) Other Mechanisms for Power Allocation

Authors in [25] proposed two power control schemes for mitigating cross-tier interference. The first scheme uses open-loop control to adjust the maximum transmit power of femtocell users in order to reduce the cross-tier interference to the level of lower than a fixed interference threshold. The second one, on the other hand, uses closed-loop control to make the cross-tier interference meet an adaptive interference threshold based on the SINR level at macrocell base station. It is shown that the closed-loop control can significantly improve the femtocells throughput compared to the open-loop control, whilst suffering only a minimal degradation of macrocell throughput.

 

5. Methodologies

In this section, we summarize the possible methodologies (including modelling, learning and optimization) for realizing self-optimization functionalities of HetNet, as shown in Fig. 4. The comparison among variant methodologies is given by Table 2.

Fig. 4.Classification of the possible methodologies.

Table 2.Comparison among variant methodologies.

5.1 Modelling

A general model of automatic management, which comprises five steps, i.e., knowledge-based monitor, analysis, plan, evaluate and execution (MAPE2-K), was proposed in [26]. In order to give a quantitative description of the self-optimization functionality, several mathematical tools associated with modelling are employed:

5.2 Learning

One of the major features of self-optimization is shown in that the system could improve its performance via self-learning from previous actions. As one of the most well-known learning approaches, game theory has been widely studied for scheduling resource in mobile communication systems, in which non-cooperative games play an important role in performing distributed self-optimization schemes. A game is usually expressed as G = [N,{Si},{Pi}], where N denotes the set of game players, {Si} represents the set of all possible strategies, and {Pi} is the set of utility function corresponding to all the strategies in {Si} (i.e., payoff). Players in a game insist to maximize their own selfish objectives, which usually damage other nodes’ performance. Furthermore, Nash equilibrium, in which each node cannot unilaterally take an action to improve its state, can be reached under the given state of other nodes in the system, i.e., pi(si|ni)≤P∗(S∗),∀si ∈ S .

5.3 Optimization

Compared with the learning approach, classic optimization approaches such as convex optimization and intelligence optimization (e.g., that using bionic theory) can also be employed for designing self-optimization functionalities in HetNet.

 

6. Conclusion and Remaining Challenges

HetNets have been developed rapidly due to the coexistent of various RATs and cell infrastructures. Self-optimization plays an important role in alleviating the upgrading CAPEX and OPEX in the large-scale HetNets. In this paper, we summarize the architecture of different HetNets and survey variant self-optimization approaches. Furthermore, some well-known methodologies have also been surveyed and compared, followed by discussing a number of potential research directions.

1) Ultra-Dense Network

User density may become extremely high in some specific time and/or place, such as assembly, airports, and subway stations. In order to meet the traffic demands of users, ultra-dense network deployment has been regarded as one of the future research directions. In particular, ultra-dense HetNets have already become the major characteristic of future wireless network (5G) [27]. Furthermore, interference management and resource virtualization have been regarded as two critical technologies in densely deployed networks. In dense HetNet, self-optimization approaches have exhibited enormous advantages in improving the network robustness, adaptation and load balancing capabilities and simultaneously decreasing the CAPEX/OPEX compared with the traditional approaches.

2) Energy Efficiency Issues

Green communication has attracted more and more attentions to both academia and industry. The development of HetNets has been motivated by the fact that deploying a high density network with low-power BSs can substantially improve the energy efficiency compared to the conventional macrocell deployment. Apart from it, the energy-cost problem is still a challenge in the dense HetNets due to the existence of “data traffic tidal effect” in practical systems.

3) Millimeter Wave and Massive MIMO

Cell size is becoming more and more small in future HetNets for the purpose of achieving more spectrum reuse gain. The shrinking cell sizes are attractive for the mmWave spectral band where RF path loss (PL) increases with frequency. The radio coverage can be effectively extended by employing massive MIMO technology, in which the the large beamforming gains may help overcome the high mmWave path loss [28]. However, a high complexity in terms of signal processing is observed in the above-mentioned technologies, highly requiring some self-optimization approaches to be performed to relieve the complexity burden of 5G systems.