Context-Aware User-Driven Framework for Network Selection in 5G Multi-RAT environments

To improve inter-operability of future 5G systems with existing technologies, this paper proposes a novel contextaware user-driven framework for network selection in multiRAT environments. It relies on fuzzy logic to cope with the lack of information usually associated with the terminal side and the intrinsic randomness of the radio environment. In particular, a fuzzy logic controller first estimates the out-ofcontext suitability of each RAT to support the QoS requirements of a set of heterogeneous applications. Then, a fuzzy multiple attribute decision making (MADM) methodology is developed to combine these estimates with the various components of the context (e.g., terminal capabilities, user preferences and operator policies) to derive the in-context suitability level of each RAT. Based on this novel metric, two spectrum selection and spectrum mobility functionalities are developed to select the best RAT in a given context. The proposed fuzzy MADM approach is validated in a dense small cell environment to perform a contextaware offloading for a mixture of delay-sensitive and best-effort applications. The results reveal that the fuzzy logic component is able to efficiently track changes in the operating conditions of the different RATs, while the MADM component enables to implement an adjustable context-aware strategy. The proposed fuzzy MADM approach results in a significant improvement in achieving the target strategy, while maintaining an acceptable QoS level compared to a traditional offloading based on signal strength. I. CONTEXT/MOTIVATION The fifth-generation (5G) of wireless networks is being developed to meet the strict requirements of future applications (e.g., two-way gaming and the Tactile Internet [1]). In this respect, an extensive amount of research has been devoted to develop various technologies and key-enablers (e.g., ultradensification, design of new radio interface and use of higher frequencies) to boost the performance of future 5G radio access technologies (RATs) compared to what can be achieved using the current technologies [2]. However, less effort has been made to ensure interworking with existing wireless systems and standards (e.g., WLANs and LTE). From a practical perspective, 5G RATs need to interwork with the existing RATs for various reasons. First, from a commercial perspective, operators often prefer to continue to use their existing network infrastructures to serve the traditional applications (e.g., voice and web-browsing) as much as possible. Second, whenever user equipments (UEs) go out of 5G coverage, traditional RATs need be used to provide a seamless perception to the end-users. In such circumstances, multi-RAT selection could be considered as complementary mode of operation to new 5G RATs. The multi-RAT selection problem has been extensively investigated in the context of 3GPP (e.g., WCDMA and LTE) and non-3GPP (e.g., WLANs) access networks [3–5]. Most of these works considered a common radio resource management (cRRM) based on a tight coupling architecture of the considered RATs. However, such a coordinated approach may not be valid to complement new 5G RATs. As a matter of fact, recent 5G architectures do not integrate most of legacy RATs (e.g., GSM, WCDMA and WLANs) or prefer to interconnect them only at the core network level because a full integration would be too costly in terms of multi-RAT measurements and interworking [6]. This means that a user-driven decisionmaking would be much more suitable particularly when the selection decision has to be made fast for specific applications (e.g., delay-sensitive) or under particular circumstances (e.g., fast degradation of radio conditions). However, the limited amount of information that is typically available to UEs through beacons and pilot channels (e.g., SS and SNR) is not enough to select the best RAT in a given context (e.g., QoS requirements, terminal capabilities and network constraints). This calls for a form of network assistance to inform UEs about all the relevant pieces of information that would be needed to perform an efficient decision. To cope with the lack of information and uncertainty associated with UEs, most proposals have relied on fuzzylogic to infer the best RAT out of the available pieces of information. In this respect, some works developed a single fuzzy logic controller (FLC) that is fed with all the relevant attributes of the available RATs [4, 7]. The main limitation of these works is the lack of scalability. As a mater of fact, when more RATs and attributes are considered, the number of inference rules exponentially increase. Other works proposed a single FLC per RAT that combines the set of available radio parameters with the various components of the context [8, 9]. The main drawback of this approach is its lack of flexibility. First, it assumes that all data is fuzzy, while some attributes can be obtained precisely (e.g., QoS requirements). Second, using fuzzy logic rules to combine both QoSand context-related attributes does not offer flexibility in adjusting the importance (i.e., weight) of each of attribute depending on the operating conditions. To offer a higher degree of flexibility to fuzzy logic, few proposals have combined it with multiple attribute decision making (MADM) that is known for its ability to efficiently combine various heterogeneous attributes [10–12]. All these proposals have made no distinction between the radio parameters that are directly related to the achievable QoS (e.g., SS, bandwidth and SNR) and the various components of the context (e.g., speed, battery consumption and price) whose importance vary from case to case. The suitability level of each RAT to meet the QoS requirements is not explicitly assessed before considering the various contextual information, which means that, in practice, the selected RAT may not meet the QoS requirements. Another key limitation of these schemes is that they all assumed that the decision-maker has full access to all required infirmation (i.e., MADM attributes and weights). Therefore, they are not valid to tackle the unique issues and constraints associated with a user-driven mode of operation, e.g., which attributes can be in practice obtained by UEs or how the network may adjust the controlling parameters to achieve a target strategy. Clearly, all the architectural constraints should be taken into account early in the design of any feasible user-driven decision-making. Therefore, the first main contribution of this paper is to construct a novel functional architecture to enable contextaware user-driven operation in multi-RAT environments. In particular, the proposed architecture relies on a UE connection manager (CM) that selects the best RAT according to a policy that is remotely adjusted by a network policy designer. Correspondingly, the second contribution is to develop a feasible fuzzy MADM implementation of the CM to select the best RAT for a set of heterogeneous applications. Fuzzy logic is used to first estimate the out-of-context suitability level of each RAT to support the QoS requirements of various applications. Second, an MADM component combines these estimates with the various components of the context (e.g., user preferences and operator policies) to derive the in-context suitability levels of the various RATs. Finally, the third contribution is to validate the proposed approach to perform a context-aware offloading in a dense small-cell environment to support a mixture of delay-sensitive and best-effort applications. The remainder of this paper is organized as follows. A context-aware user-driven functional architecture is constructed in Section. II to support spectrum management in multi-RAT environments. Then, the CM is implemented in Section. III to select the best RAT based on novel metric that assesses the in-context suitability levels of the various RATs. The proposed approach is instantiated in Section. IV to perform an intelligent offloading in dense small cell environments for a mixture of VoIP and FTP file transfer applications. The initial results are presented in Section. V, comparing several variants of the proposed approach to a traditional offloading approach. The conclusions and future directions are provided in Section. VI. II. THE PROPOSED CONTEXT-AWARE USER DRIVEN FRAMEWORK A set of K available RATs ({RATk}1≤k≤K) are considered by UEs to establish a set of L applications ({Al}1≤l≤L) that are characterised in term of various sets of QoS requirements ({Reql}1≤l≤L). At the time of establishing each of these applications, the various contextual information that may be available about the UE (e.g., velocity, remaining power and remaining balance) or the network (e.g., operator strategy and regulation rules) should be also taken into account as they may have a strong impact on the suitability of each of these RATs. Therefore, the problem considered here is whenever an application Al needs to be established, how to: • make UE select the best RAT, • to meet the set of QoS requirements, • in the considered context? To enable a context-aware user-driven mode of operation, a functional split between the UE and network domains should be clearly made to identify the logical entities and scope of each side. In this respect, the functional architecture described in Fig. 1 is proposed. Specifically, a connection manager (CM) is introduced at the UE to collect the relevant components of the context from both the terminal and network sides to implement a given decision-making policy (e.g., the proposed fuzzy MADM). The collected contextual information is combined with a radio characterisation of each available RAT in terms of a set of short-term attributes (e.g., SS, SNR and load) obtained e.g., through beacons and some mediumand long-term attributes (e.g., cost and regulation rules) stored in a policy repository together with all the policy-related parameters (e.g., algorithm, fuzzy logic membership functions and MADM weights). The content of the policy repository may be retrieved in practice from a local instance following a pull or push mode using e.g., the Open Mobile Alliance Device Management (OMA-DM) protocol [13]. To offer higher flexibility, a policy designer entity builds and updates the policy repository content based on collected measurements from the various UEs and the set of network-level strategies and constraints (e.g., operator strategy and regulation rules). For instance, the policy designer may dynamically adjust the selection policy (e.g., MADM weights) or the radio resource characterization (e.g., cost) to optimise some network-level metrics (e.g., spectrum efficiency and energy efficiency) or implement a form of traffic steering (i.e., push UEs to use specific radio resources during some periods of time). Finally, the CMs of different UEs may collaborate to further improve their individual performances. III. FUZZY MADM MULTI-RAT SELECTION This section proposes to select, for a given application Al, the RATk that would best meet each of the associated QoS requirements in the considered context. To this end, a threesteps approach is considered: 1) Design a fuzzy logic calculator to estimate the ”outof-context” QoS suitability level (s k,l) based on the available radio parameters. 2) Develop a fuzzy MADM methodology to combine s k,l with the various components of the context to derive the so-named ”in-context” suitability level (s k,l). 3) Select the RAT that maximizes the ”in-context” suitability level (s k,l). A. Out-of-context suitability levels Given the uncertainty and lack of information associated with UEs, this section develops a fuzzy-logic calculator to estimate the suitability levels of each RAT to meet the QoS requirements of the various applications. The key building block of fuzzy-logic reasoning is the fuzzy logic controler (FLC) whose architecture is described in Fig. 2 [14]. It is composed of three main stages, namely fuzzifier, inference engine and defuzzifier. During fuzzification, crisp (i.e., real) input data are assigned a value between 0 and 1 corresponding to the degree of membership in a given fuzzy set. Then, the inference engine executes a set of ifthen fuzzy rules on the input fuzzy sets. These rules, referred to as inference rules, are maintained in a rule base that is typically built based on previous expert knowledge. Finally, the aggregated output fuzzy sets are converted into crisp outputs using a given defuzzification method. In our case, the main challenge is how to design a FLC to reliably estimate the QoS suitability level (s k,l) in the most