ict- 257626 acropolis date: 28/09/2012 - cordis · ict- 257626 acropolis date: 28/09/2012...
TRANSCRIPT
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 1/42
Advanced coexistence technologies for radio optimisation in licensed and unlicensed spectrum
(ACROPOLIS)
Document Number D12.3
First Complete Set of Solution Categorization and Approaches
Contractual date of delivery to the CEC: 30/09/2012
Actual date of delivery to the CEC: 30/09/2012
Project Number and Acronym: 257626 ‐ ACROPOLIS
Editor: UPRC
Authors: Vera Stavroulaki (UPRC), Yiouli Kritikou (UPRC), Aimilia Bantouna (UPRC), Kostas Tsagkaris (UPRC), Panagiotis Demestichas (UPRC), Evangelia Tzifa (UPRC), Nikolaos Koutsouris (UPRC), Asimina Sarli (UPRC), Louisa‐Magdalene Papadopoulou (UPRC), Jad Nasreddine (RWTH), Nikos Dimitriou (IASA), Andreas Zalonis (IASA), Jing Lv (TUD), Eduard Jorswieck (TUD), Ragnar Thobaben (KTH), Ricardo Blasco‐Serrano (KTH), Liljana Gavrilovska (UKIM), Adrian Kliks (PUT), Dionysia Triantafyllopoulou (UoS), Youngwook Ko (UoS)
Participants: UPRC, RWTH, IASA, TUD, KTH, UKIM, PUT, UoS
Workpackage: WP12
Security: Public (PU)
Nature: Report
Version: 1.0
Total Number of Pages: 42
Abstract:
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 2/42
Deliverable D12.1 [1] presented, among others, 10 case studies, i.e. 10 decision making problems that will be addressed by the project. Deliverable D12.3 proposes how these decision making problems can be addressed, i.e. presents a high level description/ the approach that was followed for solving them. In particular, for every proposed solution the adapted theoretical approach is provided while the more detailed description of each proposed mechanism and how the theoretical approach is differentiated in the specific mechanisms will be discussed in Deliverable D12.4 “Description of Decision Making Approaches and Decision Making Engine”. Finally, based on the commonalities and the differences between the proposed decision making solutions, a first categorization of the solutions has been provided.
Keywords: Decision Making Mechanisms, categorization, model‐based, Learning‐based, context‐aware
Document Revision History
Version Date Author Summary of main changes
0.1 02.11.2011 UPRC Table of Contents
0.2 14.02.2012 UPRC Updated ToC
0.3 29.05.2012 UPRC Partners’ intentions to contribute
0.4 11.06.2012 IASA IASA, UoS, PUT, RWTH joint contribution
0.5 13.06.2012 KTH KTH, TUD, PUT joint contribution
0.6 02.07.2012 UPRC UPRC Contribution
0.7 15.08.2012 UoS UNIS Contribution
0.8 11.09.2012 UKIM SOTA section added
0.99 18.09.2012 UPRC Text Rearrangements/ Editing
1.0 19.09.2012 UPRC Finalization of the document
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 3/42
Executive Summary
This document provides the categorization of decision making solutions related to various aspects of cognitive radio systems, some of the related work that has been reported during the last years for each category and the first complete set of the proposed decision making solutions in the context of the scenarios examined by ACROPOLIS project.
To begin with, the document starts (in section 2) with the categorization of the decision making mechanisms used for the cognitive networks. More specifically, the decision making mechanisms are divided to 2 large categories depending on the algorithm on which they are based: i) the model‐based and ii) the learning‐based decision making the mechanisms. The 1st category can be analysed even more to 2 sub‐categories, namely the stochastic approaches and the examination ones. Section 2 elaborates on the each of these (sub‐) categories and it provides an overview/ description of them and some state of the art for each category.
Section 3 focuses on the decision making solutions proposed by ACROPOLIS project in order to address the problems described in the previous WP12 deliverables, namely D12.1 [1]. In particular, the section presents the overview of the scenario and the approach to be followed. However, the more detailed description of the mechanisms including the results that come from it will be included in the next WP12 deliverable, i.e. in D12.4 “Description of Decision Making Approaches and Decision Making Engine”
Finally, section 4 concludes the document highlighting its importance within WP12.
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 4/42
Table of Contents
1. Introduction ................................................................................................... 5 2. The Categorization of the Decision Making Mechanisms and the State-of-the-Art ...................................................................................................................... 6
2.1 Model-based spectrum decision mechanisms ................................................... 7 2.1.1 Stochastic approaches ........................................................................... 7 2.1.2 Examination approaches ........................................................................ 9
2.2 Learning-based spectrum decision mechanisms ............................................. 10 3. Proposed Solutions ...................................................................................... 13
3.1 Model-based spectrum decision mechanisms ................................................. 13 3.1.1 Strategy Selection and Power Allocation for Hierarchical Spectrum Access on Licensed Bands ........................................................................................... 13 3.1.2 Context-aware Interference Management in heterogeneous wireless networks ................................................................................................................ 15 3.1.3 Energy-aware spectrum sharing with femtocells ...................................... 27
3.2 Learning-based spectrum decision mechanisms ............................................. 29 3.2.1 Knowledge-based cognitive Radio Resource Management .......................... 29 3.2.2 Radio Resource Management based on Dynamic Sub-carrier Assignment (DSA), Context, Profiles and Policies .............................................................. 31
4. Conclusions .................................................................................................. 35 5. Glossary and Definitions............................................................................... 36 6. References ................................................................................................... 38
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 5/42
1. Introduction Main aim of WP12 “Metric Identification, Decision Making Algorithms and Solutions” is to define and identify the procedures used in the analysis of the context data to identify potential opportunities and to decide on the spectrum allocation. Moreover, this WP defines and identifies the key elements of spectrum sharing and decision making to allow intelligent and efficient choice of spectrum access, based on spectrum access policies and available or unused spectrum. Towards this direction, D12.1 “Specification of Preliminary Set of Appropriate Metrics, Utility Functions and Layer Identification” [1] described 3 scenarios which consisted of 10 case studies in total, i.e. 10 decision making problems to be addressed by ACROPOLIS project. Those are:
Reference scenario 1 – Cellular networks with cognitive capabilities (e.g. femtocells) o Case Study 1: Utility functions related to the resource taxation o Case Study 2: Policy derivation and enforcing in cellular‐like systems o Case Study 3: Interference management and scheduling for femtocells
Reference scenario 2 – Coordinated/coexisting cognitive radio networks o Case Study 1: Utilization of policy coordinated access in coordinated/
coexisting cognitive radio networks o Case Study 2: Spectrum sharing between operators in licensed and
unlicensed spectrum from a PHY and MAC point of view o Case Study 3: Knowledge based management of reconfigurable B3G
infrastructures o Case Study 4: Self‐optimization of cognitive devices
Reference scenario 3 – Opportunistic cognitive networks (e.g. secondary ad‐hoc networks in presence of primary systems)
o Case study 1: Impact of cross layer utility function in routing for cognitive networks
o Case study 2: Underlay CR with multiple‐antenna secondary users o Case study 3: Opportunistic coverage extension and capacity extension
This deliverables discusses on the approaches that can be followed for resolving these case studies. In particular, the deliverable presents high‐level descriptions of the proposed solutions of the decision making problem that had been described in D12.1 [1] while more detailed description of the decision making solutions will be provided in D12.4 “Description of Decision Making Approaches and Decision Making Engine”. Finally, a categorization, based on the commonalities and the differences, of the proposed mechanisms is also presented. The document is structured as follows: Section 2 presents how decision making mechanisms for spectrum access are categorized and summarizes the respective State‐of‐the‐Art per category. Section 3 presents the followed approaches for solving the decision making problems and thus introduces the mechanisms that will be detailed in D12.4 “Description of Decision Making Approaches and Decision Making Engine”. Finally, Section 5 draws conclusions.
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 6/42
2. The Categorization of the Decision Making Mechanisms and the State‐of‐the‐Art One of the crucial aspects in Cognitive Radio (CR) networks is the decision making for spectrum access (i.e. spectrum decision) process. It is responsible for enabling higher spectrum utilization of the CR system as well as fulfilling the Second User's (SU's) Quality of Service (QoS) requirements. In order to derive an efficient spectrum decision the SU needs to continuously monitor the radio environment and utilize a long‐term analysis so that it can reason and act upon its own and the environmental behaviour, Figure 2‐1.
Secondary User Radio Environment
Act • Actuation • Parameter reconfiguration
Observe • Radio Environmental Information
Decide • Reasoning • Parameter decision
Figure 2-1: High-level cognitive cycle model
As seen on Figure 2‐1 the decision making is one of the pillars in the cognitive cycle and it is tightly related with the observation and action processes. In the observation process the SUs gather the radio environmental information. The term radio environmental information refers to any type of information that could be utilized by the decision process (e.g. interference level, spectrum occupation, channel state, regulators’ rules and policies, SU performance, geolocation, noise level, etc.). The gathered information (raw or pre‐processed) is then fed to the decision process (i.e. decision making mechanism) upon which the SUs reason and decide. Based on the decision made the action process performs parameter reconfiguration and actuation, closing the loop of the cognition cycle. Depending on the level of complexity the decision process can deal with metric analysis, performance optimization, scheduling, etc., and utilizes either model‐based or learning‐based decision tools. In the case of model‐based tools the decision is being made by classical optimization methods which can often be ineffective for solving large scale or non‐convex problems. In the case of learning‐based tools the decision is being made based on techniques like, Bayesian Networks, Reinforcement Learning, etc. Regardless of the decision tool the spectrum decision represents a complex process that spreads across multiple CR spectrum functions, i.e., Spectrum Sensing, Spectrum Access and Spectrum Sharing, Figure 2‐2.
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 7/42
Radio Environment
Spectrum Sharing
Spectrum Access
Spectrum Sensing
Radio environmental
information
Spectrum
hole
detec
tion
PU
detection
Spectru
m
Mobility
Channel
Selection
Signal
Transm
ission
Figure 2-2: Spectrum decision cognitive cycle model
In the case of Spectrum Sensing the spectrum decision process determines which spectrum portions are available for opportunistic access. For this case the decision represents a binary hypothesis testing where the two hypotheses are refereeing to the Primary User (PU) presence and PU absence, respectively [2]. The goal of the decision process in the case of Spectrum Access is to select the best available channel as well as to vacate the given channel when an incumbent user is detected (i.e. spectrum mobility). For the Spectrum Sharing case the decision process is responsible for coordinating the access to the selected channel with the other active SUs.
The reminder of this section will elaborate on the State of the Art (SoA) (model‐based and learning‐based) spectrum decision mechanisms that focus on the Spectrum Access and Spectrum Sharing functions.
2.1 Model‐based spectrum decision mechanisms
The model based decision approaches for cognitive radio networks generally rely on a significant prior knowledge to make the decisions for the spectrum sensing, access and/or sharing. Regarding the amount and type of required knowledge prior the decision process the model based approaches can be further divided into two subclasses: stochastic decision making approaches and examination based decision making approaches.
2.1.1 Stochastic approaches
The stochastic approaches commonly rely on the assumptions of a known PU/SU activity/occupancy model and parameters to derive the optimal decision for the spectrum access and/or sharing. The aim is to apply the most realistic model and through extensive numerical and simulation analyses to assess the performances of different decision schemes in various cognitive radio networks scenarios. The gained conclusions are used to derive the
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 8/42
decision rules which can be applied to the cognitive radio devices to adapt on‐the‐fly based on the different radio environment context. The stochastic decision making approaches can be generally divided based on the type on:
PU/SU system model
Optimization objective In order to produce an efficient decision the stochastic decision making approaches must rely on adequate and accurate PU/SU system models. Most of the SoA work in terms of the model type is mainly based on either Markov transition models or queuing theory models. When considering the former approach the opportunistic access process can be modelled either using a classical Markov transition model [3][4][5], or utilizing Partially Observable Markov Decision Process (POMDP) [6]. Moreover, the POMDP approach is suitable for scenarios where the state of the system is partially observed (e.g. in the case of wideband sensing the state of all candidate spectrum bands may not always be known) [6][7][8]. Queuing theory has proven to be a suitable tool for the stochastic decision making approaches where the PU/SU system can be modelled with different queuing networks models (e.g. M/M/1 queuing model [9] or a Pre‐emptive Resume Priority (PRP) M/G/1 queuing model [7][9]). In most of the works the PU and SU arrivals are modelled as Poisson [3][4][7][10] and the departures as exponential, while the channel occupancy model of the primary system can be modelled as a Bernoulli process [10] or a as a two‐state birth‐death process [11]. Several works have also considered the possibility to model different SU application types with different traffic queues models [11][12].
The stochastic decision making approaches are based on a predefined optimization objective upon which they derive the given decision. They either target a single objective or multiple objectives regarding the optimization problem i.e. decision making process. Most of the single objective works focus on variety of parameters like the maximization of the secondary throughput [5][8][11][12], minimization of the overall SU service time [3][7], minimization of the capacity variance for real‐time applications [11], minimization of the overall SU channel switching delay (SU spectrum handovers time) [6], achieving fairness among the SUs and applications while satisfying quality of service [12] etc. The stochastic decision making approaches target multiple objectives in the optimization problem. They perform a joint optimization on a set of parameters that refer to the cognitive cycle. One example is a group of stochastic decision making algorithms that target a set of the sensed channels and the set of the candidate channels for spectrum access and usage [5][6][7], covering the decision making process in the first two stages of the cognition cycle. Another group of stochastic decision making algorithms attempts to jointly optimize the spectrum access decisions, admission decisions and/or the spectrum/resource allocation decisions [3][8][11][12], whether between the different secondary users and applications or between the primary user and the secondary user. These algorithms simultaneously cover the decision making in the spectrum access and spectrum sharing stages of the cognition cycle, depicted on Figure 2‐2.
The stochastic decision making algorithms are simple and easy to implement since they usually have closed‐form expressions or the target optimization space is narrower and they focus on a limited set of scenarios. However, due to their assumptions and simplifications they are not always capable to cope with the dynamism of the radio environment and unexpected events. Thus, the model based techniques performing examination, as well as
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 9/42
the learning based decision approaches are better suited to the inconsistent and the unpredictable radio environment.
2.1.2 Examination approaches
The examination approaches are a subset of the model based approaches which target more complex optimization problems. They also rely on a prior knowledge, but in this case on the complex relationships between the objectives, the measured metrics and the optimization parameters. However, since they generally target multi‐objective optimizations a closed form expression for the optimal parameters is hard to obtain. Therefore, they use fitness functions measuring and quantifying the satisfactory level of the predefined criteria. In [13][14] the authors have singled out the following objectives relevant for the physical and MAC layer optimizations in cognitive networks. Figure 2‐3 illustrates the dependences and the relations between the objectives in such a multi‐objective optimization. It is clear that the optimization of multiple goals would require a substantial amount of time due to the complex relationships between the objectives.
Figure 2-3: Multi-objective optimization in cognitive networks [13]
There is a vast amount of papers dealing with the multi‐objective decision making in cognitive radio networks. Genetic algorithms (GAs) are proposed as efficient tools for decision making in the field of CR [15] due to their ability to operate in a large search space. GAs have high ability to adapt to dynamically changing environments and add significant amount of power and flexibility in the decision making engine. As cognitive radios are likely to face dynamic environments and situations, genetic algorithms are particularly applicable. The GAs can be successfully used to perform joint optimization on the channel allocation and power control optimization in the quest of achieving an optimal Signal to Interference‐plus‐Noise Ratio (SINR) per SU device [16]. Genetic algorithms can be also used to perform a joint maximization of multiple objectives such as BER minimization, power minimization, and throughput maximization [17]. In [18] the proposed Cognitive Resource Manager (CRM) selects an algorithm from a toolbox of algorithms to solve a particular problem, where the genetic algorithms are preferred for multi‐dimensional problem analysis.
Besides the GAs, the Swarm Algorithms are another biologically inspired and efficient tool for optimizations and decision making in the cognitive radio networks. They are more suited to the distributed scenarios, where multiple secondary users jointly cooperate to make the best time‐frequency allocation of the spectrum resources [19][20]. The iterative evolution of the swarm guaranties an efficient spectrum allocation only after several iterations.
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 10/42
Game Theory [21] represents a possible tool for strategic decision making as well. Mostly used in the mathematical economics in the history, recently the game theory approaches arise as an efficient tool for spectrum access and spectrum sharing in the cognitive radio networks. The game theoretic approaches such as Cournot game, Nash Equilibrium, Nash Bargaining, are seen as an efficient solution for the decision making in spectrum markets for the spectrum trading process [22]. Additionally, the game theoretic approaches have been used to jointly optimize the channel and power allocation to maximize the number of served SUs and maximize the secondary throughput [23].
Besides the previously noted examination based decision making approaches, the simulated annealing [24] and fuzzy logic [25] approaches have also been used for the decision making in the cognitive networks. Since all the examination based approaches usually perform multi‐objective optimizations they mostly target the spectrum sharing stage of the cognition cycle or they perform the joint spectrum access and sharing decisions in the cognitive networks.
2.2 Learning‐based spectrum decision mechanisms
The learning‐based spectrum decision concept extends the common notion of decision making by introducing knowledge‐base and learning functionalities, Figure 2‐4, in the decision making engine.
Figure 2-4: Learning-based decision making engine
The learning engine enables the CR to evaluate the quality of the past actions and enables the decision engine to learn from the past successes and failures, adjust its parameters, and adapt its decision rules to its specific environment. The learning techniques involve the use of an objective function to determine the value of the learned data, where in a CR context these objective functions reflect the overall goal of the application, such as maximizing the throughput or minimizing the interference. The most commonly utilized learning techniques (tools) regarding the spectrum decision process are:
Bayesian Networks
Reinforcement Learning
Artificial Neural Networks Bayesian Network (BN) is a unique tool for modelling the network protocol stack as it not only learns the probabilistic dependencies of the system but also provides an opportunity to fine‐tune the cognitive network parameters in order to achieve the desired performance. A BN model for a CR network incorporates the following steps (analog to Mitolas’ defined cognitive cycle): sampling the network protocol parameters (Observe), learning the structure of the BN and its parameters from the data (Learn), using a BN‐based inference
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 11/42
engine (Decide) to make decisions, and finally effecting the decisions (Act) [26]. The work presented in [27] introduces a novel autonomic decision‐making approach which extends the common Bayesian Network paradigm by introducing the concept of influence diagrams. Utilizing the influence diagrams, in order to decide and execute the action, can maximize and effectively predict the overall CR network performance. The spectrum decision mechanism in [28] exploits BNs to model the probabilistic dependences of spectrum occupancies and influence diagrams for a cross‐layer decision making approach. The mechanism integrates the inference result of the Bayesian network model regarding the spectrum occupancy and the QoS requirements from the upper layers in order to calculate (derive) the most efficient decision solution.
Reinforcement Learning (RL) is a biologically inspired machine learning technique widely used in CR decision making, in which an agent attempts to perform optimal actions to maximize long‐term rewards achieved by interacting with the environment [29]. The dynamic interaction with the environment and the adaptivity of the learning process are two of the main features which make RL techniques suitable for Cognitive Radio Ad Hoc Networks [30], mainly for routing and spectrum decision tasks. As discussed in [30] the most suitable RL techniques for CR Networks are:
Model‐based learning. These algorithms require a model of the environment, i.e. the reward and state‐distribution functions;
Q‐learning. Q‐learning is an on‐line RL algorithm which attempts to estimate the optimal action‐state function without requiring a model of the environment;
Dual RL. This algorithm is very similar to the Q‐learning scheme. The difference is that it updates the action‐state function values based on the previous states, instead of relying only on the subsequent states;
Multi‐Agent Learning. These techniques extend classical RL algorithms, in a system of homogeneous agents which have system‐wide optimization goals.
The authors in [29] propose a value function based approach for evaluating the suitability of different transmission parameters, which can provide more efficient decision making. Q‐learning based spectrum allocation has been extensively investigated in CR networks. The work in [30] proposes a novel decision framework based on RL techniques for joint power and spectrum allocation as an example of Q‐learning. The elaborated method in [31] considers the discussed RL‐formulation assigning rewards to the SU after each data transmission. In [32], the authors consider a spectrum sharing case where a set of transmitting‐receiving pairs of nodes utilize the information of the received Signal‐to‐Noise Ratio (SNR) as an input in the spectrum decision process. If the SNR value is higher than a given threshold, then the transmitter node increases its Q‐value of a fixed weight factor, otherwise it applies a penalty and quits the current spectrum.
Authors in [33][34][35][36] suggest that many problems related to CR’s decision making can be formalized as Multi‐Armed Bandit (MAB) problems and that solving such problems by using Upper Confidence Bound (UCB) algorithms can lead to high‐performance spectrum decision making. The MAB framework models a gambler sequentially pulling one of the several levers on the gambling machine. Every time a lever is pulled, it provides the gambler with a random income usually referred to as reward. Although it is assumed that the gambler has no a priori information on the rewards' stochastic distributions, he aims at maximizing his cumulated income through iterative pulls. In the spectrum decision case, the SU is modelled as the gambler while the frequency bands represent the levers. A possible
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 12/42
way to solve the MAB modelled problem is to use the class of UCB algorithms [35][36][37][38][39]. The main advantage of UCB methods for CR is to offer a balance between the exploration and the exploitation phases without interrupting the communication process and to provide learning solutions even in the cases when the cognitive nodes face new environment conditions [40].
The general approach suggested in [36] aims at selecting the most suitable decision based on the UCB indexes. These indexes are based on the rewards associated with the channels that the SU can potentially exploit. In [34] authors utilize the UCB algorithm to derive the most optimal spectrum decision. They formulate the spectrum decision problem through a general learning model that takes into account the observation limits of the SUs and derive fundamental bounds on the UCB algorithm. Moreover, in order to maximize the overall SU network performance without any a priori knowledge of the system behaviour the authors propose a general spectrum decision mechanism based on the Hungarian algorithm [34].
Artificial Neural Networks (ANN) may also be used by the decision making engine to learn and infer decision making rules. Similar to the brain, the ANN is composed of artificial neurons (or processing units) and interconnections [41]. Thus it forms a programming structure that mimics the behaviour and neural processing (organization and learning) of biological neurons. Based on ANN, the method of evaluating and learning best decision for CR decision making is proposed in [42]. In the same work several key architectural issues for cognitive radio engine based on ANN are discussed, including knowledge base information model and learning model Neural Network design. Several works also investigate the performance gain in terms of the decision process for the Multilayer Neural Network (MNN) approach. It has been shown that the MNN can be exploited as an effective tool for real‐time decision making in cognitive radio [43][44].
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 13/42
3. Proposed Solutions This section, and deliverable in general, presents the context/ scenario under which the decision making mechanisms are proposed and the theoretical basis/ approach of the proposed mechanisms but not the mechanisms per se. The detailed description of the mechanism with the inputs/outputs and results will be part of D12.4 “Description of Decision Making Approaches and Decision Making Engine”.
3.1 Model‐based spectrum decision mechanisms
This section summarizes the proposed solutions that belong in the first category of section 2, i.e. solutions that are based on models for addressing the decision making problem. The reasons of why each approach has been selected reside in the specific analysed scenario and can be found in the respective analysis of the mechanism.
3.1.1 Strategy Selection and Power Allocation for Hierarchical Spectrum Access on Licensed Bands
In this section, we consider strategy selection and power allocation in context of hierarchical spectrum access on licensed bands, which is one of the main scenarios of interest of the ACROPOLIS project, as indicated in deliverable D4.1 [45]. The discussion provided in the following is mainly based on joint work between TUD, KTH, and PUT (see, e.g., [46][47] and [48]) and individual work from TUD in [49].
Scenario and Overview
We consider the scenario where a secondary system wishes to establish a link in an occupied licensed band; that is, we explicitly exclude interweave spectrum sharing in this scenario. For simplicity, we assume that the secondary system consists of one transmitter‐receiver pair and that only one primary transmitter‐receiver pair is active in the considered resource blocks. The situation is illustrated in Figure 3‐1, and it is similar to the two‐user interference channel for which utility functions have been discussed in deliverable D12.1 [1]. The theoretical background for this channel model and its extension to the so‐called cognitive radio channel model are furthermore discussed in detail in deliverable D8.2 [50].
Figure 3-1: Hierarchical spectrum access on occupied licensed bands.
In our work in [46][47][48][49], we considered the special case of a Multiple‐Input Single‐Output (MISO) secondary link coexisting with a Single‐Input Single‐Output (SISO) primary system. The scenario was designed to be sufficiently complex to cover the main technical details of the problem while being simple enough in order to obtain tractable results. In this scenario, the secondary system has two choices: it can enter the channel in a non‐ or semi‐
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 14/42
coordinated fashion using an underlay strategy (see, e.g., [49]) or it can coordinate its transmission strategy with the primary system and use an overlay strategy (see, e.g., [46][47] and [48]). If the secondary system now decides for an underlay strategy as described in [49], it can choose to either split the secondary’s message into multiple data streams or transmit a single data stream, and it has to optimize the rate splitting, the corresponding beamforming vectors, and the power allocation in order to maximize the secondary rate while ensuring a given primary rate constraint. We refer to [49] and deliverable D13.1 [51] for the technical details. On the other hand, if the secondary system decides for coordinated spectrum sharing using the schemes investigated in [46][47][48], this requires that the secondary transmitter serves as a relay for the primary system while it transmits its own message to the secondary receiver at the same time. To implement this scheme, the secondary system needs to “learn” the primary’s message first, and it has to optimize the beamforming vectors, power allocation, and time‐sharing parameters in a second step. We refer to [46][47][48] and deliverable D13.1 [51] for the technical details.
The decision, which spectrum sharing strategy will be used, depends on application requirements like throughput, delay, and reliability as well as on the availability of an interface that enables coordination between the primary and the secondary system, the willingness of the primary system to share certain key parameters, and the allowed complexity at the secondary transmitters and receivers. We summarize therefore in the following the context information that is required for the secondary user for assessing the expected performance of the underlay scheme described in [49] and the overlay schemes studied in [46][47] and [48].
For the underlay strategy [49], the following features are essential:
Channel‐state information: The secondary system requires full knowledge of the realizations of the fading processes on all links, as well as the primary transmit power and the noise variance at the receivers in order to be able to reliably predict the expected performance.
Primary‐rate constraint: The optimization of the transmit strategy requires knowledge of the primary rate constraint. If full channel‐state information is available and all transmit powers are known, this constraint can be expressed as an equivalent interference temperature constraint as well.
Complexity: In [49], both a single‐stream decoder and a more advanced multi‐stream decoder based on successive decoding are considered. Which strategy is selected depends therefore on the allowable complexity of the secondary receiver.
For the overlay strategies developed in [46][47] and [48], both the requirements on the channel‐state information as well as the requirements on the primary‐rate constraint remain the same. In addition, the following features need to be provided:
Primary message‐state information: The considered cooperation schemes are based on decode‐and‐forward relaying. This requires that the secondary transmitter has knowledge of the primary codebook and is therefore able to decode (“learn”) the primary message.
Common set of relaying strategies: The secondary transmitter has to be able to support all relaying and cooperation strategies that are implemented in the primary system.
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 15/42
Synchronization: Cooperative transmission from distributed transmitters requires synchronization to ensure that the transmitter signals are perfectly aligned.
Complexity: The transmission scheme in [47] is based on dirty paper coding which requires a quite involved transmitter structure. To reduce complexity at the transmitter, we proposed in [48] an approach based on linear pre‐coding that allows the secondary system to trade performance for reducing complexity. Which strategy is selected depends therefore on the allowable complexity of the secondary transmitters
At this point, it is important to mention that the summary of requirements, given above, lists the idealized assumptions. How the performance of the considered schemes is altered due to practical imperfections is for example studied in [46], where robust beamforming based on imperfect channel state information is considered.
Theoretical Approach
To decide whether an underlay or an overlay strategy is chosen, the decision engine has to evaluate the following aspects for all available strategies:
1. Are all parameters that are required to use the considered scheme available? 2. Does the expected performance satisfy the requirements set by the application (rate,
delay, reliability, etc.) while satisfying constraints set by the primary system (e.g., interference), and how can the expected performance be achieved?
3. Is the solution that leads to the expected performance feasible given the complexity constraints?
4. If multiple solutions are available, which scheme provides the required performance at the lowest price (e.g., complexity or power consumption)?
In order to be able to do this, the decision engine needs a reliable tool to predict the performance of the considered scheme. In [46][47][48][49], we employed achievable rates that are based on information theoretic models for that purpose. We refer to deliverable D8.2 [50] for a large collection of information theoretical models that are suitable for that purpose. Now, given appropriate rate expressions, the selection of the strategies and the optimization of the respective parameters can be stated as a rate optimization problem. The solution of the optimization gives then the expected performance and the set of parameters that achieve it. To illustrate this, equation (1) shows the simplest case of a rate optimization problem for spectrum underlay that was considered in [49]. It describes the optimization of
the beamforming vector and the power allocation for maximizing the secondary rate cR if a
single‐stream underlay strategy is used and the primary rate constraint is lR . As shown in
[49], a closed‐form solution can be obtained for this specific problem.
2 222 11
2 22 2,12 21
max log 1 s.t. log 1 , 1,01 1c
Hc c I
c I c c cHWc p
I c c
h w p h PR R w p P
h P h W p
(1)
3.1.2 Context‐aware Interference Management in heterogeneous wireless networks
In recent years, in order to increase system capacity, network operators enhance their infrastructure with the addition of small cells (e.g. relays, microcells, femtocells). This action transforms classical one‐tier networks to multi‐tier networks, where the proper resource allocation plays an important role in the control of interference. Moreover, network operators try to find alternative ways for coverage and capacity increase in densely
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 16/42
populated countries. One possible option is to offload the part of the traffic to unlicensed spectrum utilizing WiFi technology [52]. However, the interference between WiFi users seems to be a crucial factor that reduces the potential capacity improvement. Since WiFi terminals operate on non‐licensed bands, the Interference Management (IM) problem cannot be treated as it is typically done in the so‐called two‐tier approach, e.g. for femtocells. The degree of freedom in interference management and avoidance is strictly limited in this case, since the operator does not have any influence on the other WiFi users and other wireless systems operating in the 2.4 GHz band. Similar analysis can be performed for cellular networks (LTE/4G/WiMax), where the full frequency‐reuse approach has been applied. In such a situation the whole set of the resources assigned to a certain operator can be used in each separated cell. In all cases the problem of effective yet simple interference management as well as of effective Resource Allocation (RA) among users becomes crucial.
To control interference, one can list the two main approaches: (i) interference cancellation techniques, where the user subtracts the strongest interferers from the received signal, and (ii) interference avoidance techniques, especially through RA, where the users try to avoid, instead of suppressing, the interference. For both cases rich context information about the current situation of the terminal plays significant role. Efficient interference management will be impossible if the information about the vicinity of the particular user will not be available. However, depending on the approach, different amount and types of information are required. Based on [53] context information is the information possessed by the entity that can be used to describe and characterize the current situation of this entity (element of the network). The accuracy of the characterization of the current situation depends not only on the amount of information gathered by the entity but also on the quality of the context information. Moreover, depending on the entity type (e.g. mobile or fixed terminal) the useful and available context information change significantly. For instance some information accessible in cellular networks cannot be obtained in WiFi systems.
In the following interference management approaches for selected scenarios of interests are discussed along with the available context information for each scenario. Algorithmic solutions for those scenarios are proposed along with an assessment for the impact on the system performance.
3.1.2.1 Cellular Networks with Full Frequency Reuse
Scenario and Overview
IM in cellular networks relies on some feedback parameters sent by user equipment to Base Stations (BSs). Usually these parameters are related to measurements of the received signal strength or the SINR, on specific frequencies that may assist in determining the propagation losses and its variations. Apart from assessing the link quality between the BS and the mobile, it is important to assess the interference experienced due to activities in other co‐channel links, and to estimate the expected consequences to other users after allocating a specific resource to the user of interest.
The traditional resource management techniques in legacy systems have the following characteristics: (i) the cellular infrastructure is considered as fixed and the amount of resources allocated in each cell does not change in short term; (ii) users transfer only channel and signal quality metrics and not any location information; (iii) users do not sense explicitly other systems that may be existing in other frequency bands. One should note that the specifications for Long Term Evolution Advanced (LTE‐Advanced) are considering the use
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 17/42
of more rich information especially for the Minimization of Drive‐Tests (MDT) procedure; (iv) the base‐line configuration of cellular systems does not involve the cooperation between base stations for optimized resource usage; (v) there is no provision for opportunistic resource access or for primary/secondary prioritization of users in RA.
In the case of a macro‐cellular architecture with nearly full frequency reuse in adjacent cells, there is a need for optimum RA and IM. The resulting co‐channel interference that may be generated needs to be carefully controlled to ensure seamless operation between the BSs. In next generation cellular systems where various small cells (femtocells, picocells, metrocells, microcells) will coexist and share the common resources, the coordination between those “small cell” APs and macro‐BSs is of great importance.. This co‐ordination can be assisted by using rich context information: (i) BS and AP locations – this requires a low update rate and will serve for making long‐term network planning decisions; (ii) information about network infrastructure, such as BS power control capabilities; (iii) real‐time or almost real‐time measurements conducted by users, along with their locations. Each user may collect information regarding its surrounding radio environment, containing even more dynamically updated measurement information; and (v) user requirements on QoS.
In the considered scenario the important issue is to determine the possibility of re‐using the same frequency channels for different users and the corresponding consequences related to the generated co‐channel interference. The system model is presented in Figure 3‐2, where three omnidirectional transmitting APs, named A, B, and C, and three receiving Mobile Stations (MSs), named 1,2,3, communicate within the same area.
A
C 3
1
B
2
Figure 3-2: Co-channel allocation scenario
The objective is to examine how the available context information can help in increasing the efficiency of IM techniques. The analysis is valid for any scenario involving co‐channel frequency allocation in neighbouring cells; the assumption of 3 co‐existing co‐channel users within the same area is valid and covers most practical system layout scenarios for macro‐cells, small cells (e.g. femtocells) and opportunistic access systems. One such example is
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 18/42
cellular OFDMA full frequency reuse scenarios, involving the wide re‐use of the same sub‐channels. Another example is co‐channel LTE femtocells deployment in a specific area, or WiFi APs deployment with a provision of re‐using of the same sub‐channels among them.
Theoretical Approach and assessment
The calculation of the interference in each possible interference scenario of Figure 3‐2 is based on the parameters in Table 3‐1. The total channel gain gm,n = L(α,d,S,H) can be segregated into path‐loss component at distance d with attenuation factor α, shadowing S, and fast fading term h. Depending on the scenario of interest and the available context information the exact values or probabilistic models can be used to represent the above parameters. The exact values, in the generic case, are dependent on the positions of the transmitter and the receiver. In each scenario the available channel knowledge can be represented accordingly as a special case of the complete total channel gain gm,n. Below an exemplary sum‐utility optimization problem is presented as an example, with the intention to be as generic as possible to cover as much as possible sub‐problems in the area. Clearly, other classes of generic optimization problems can be also identified. In the ideal case, where we have the exact channel knowledge of all the interconnections in the network and a centralized controller that will collect in real‐time this information, a generic optimization function that targets to the maximization of a specific function X(.) can be formulated as:
. . .,
ˆˆmax m n m n m nm n
X r P (2)
where the sum is over the tuples (m,n). Power and rate allocated for a MSm are represented as pm,n and rm,n, respectively. Function X(.) can be the capacity, the sum good put, or something else. Various optimization functions could be selected depending on the scenario of interest. This generic optimization example can be used as a reference for various practical sub‐problems in the area of IM and RA in femtocell networks.
Parameter Description
, 1, 2,3mMS m m‐th MS user (total number of MS is equal to M)
, , ,nAP n A B C n‐th AP (total number of AP is equal to N)
,m ng Total channel gain between the nth AP and m‐th MS
Table 3-1: Exemplary set of context parameters
Conditions for feasible co‐channel allocation in the scenario of Figure 3‐2: Each user is assumed to have a specific QoS requirement in terms of SINR value, i.e. the m‐th mobile station MSm has an SINR requirement equal to γm. By considering the conditions that have to be fulfilled by user MSi to be connected to APX, it can be shown that there can be feasible power allocation strategy for all three users if the expression concerning the Feasible Power Allocation Criterion (FPAC) is valid:
1 2 31 0sFPAC L (3)
Here Ls represents the interference coupling factor that depends on the path losses between the APs and the MSs as has been introduced in [54] for CDMA systems and
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 19/42
adopted herein for any full‐frequency reuse system. This criterion may be integrated with the maximum power conditions that may be imposed by the AP transceiver characteristics.
We assume the knowledge of the location and the QoS requirements of all users, along with the propagation models for all possible user links. It is obvious that, if this information is available in each AP, it can be directly inferred whether there is a feasible co‐channel allocation for all users. Thus, the previously described mathematical model can be used to determine feasible power allocation options for each user according to its QoS requirements and location. In case there are more than one connection options using the same frequency (e.g. connection to a macrocell BS or a femtocell AP), this model can be used to identify better connection type.
A practical algorithm, which uses the above‐mentioned information for performing power allocation for each user of the scenario of Figure 3‐2, is the following:
1. The user in each cell sends to the AP information regarding the received SINR measurements.
2. The AP calculates the user location (this can be done either by assuming that the user sends directly his/her location information or by assuming that the BS uses some triangulation or other localization determination techniques, in collaboration with the neighbouring BSs that also can locate the user of interest).
3. The AP updated a local database the new user data. 4. The AP sends the updated information for all users to a cluster controller (a cluster
Radio Environmental Map ‐ REM). 5. The updated cluster controller is used to calculate the FPAC from (3) and to allocate
in an optimum fashion the same frequency to users in the adjacent cells. 6. If no feasible solution exists, the users with the higher propagation loss values are
allocated different channels to avoid excessive interference. It was assumed that if the FPAC was positive, any feasible transmitted power vector corresponding to the specific user QoS requirements and user location would be within the power limit bounds of each AP.
The simulation results show the benefit of exploiting context information for the power allocation of co‐channel users in adjacent cells in order to control the generated interference. An area of 3 neighbouring omnidirectional cells (each having radius equal to 1 km) was simulated. User locations were generated randomly in each cell. This procedure is compared to the case of radio distance‐based fractional frequency reuse‐based schemes [55] according to which the AP allows full frequency reuse within a radius Rc smaller than the cell range (radio distance in this simulation coincides with the user distance from the base station due to the chosen path loss model). The tests were performed assuming a Full Frequency Reuse (FFR) area of radius 0.5Rc, 0.75Rc, and Rc (denoted as FFR‐0.5, FFR‐0.75, FFR‐1, respectively). The FFR‐1 scheme coincided with the case, according to which the AP allocated resources to all users assuming full frequency re‐use, and was considered as a lower bound case. In each case, following the co‐channel allocation, the SINR of each user was measured. If the SINR si was below a specified threshold (set to 12 dB for the tests), the user was considered to be in outage.
Figure 3‐3 depicts the frequency reuse utilization (defined as the outage‐free channel usage per cell) measured in the 3 neighbouring cells due to the generated co‐channel interference. It is apparent that in the case of the uncoordinated resource allocation (FFR‐1), users close
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 20/42
to the cell borders suffered from excessive inter‐cell interference and thus a considerable outage rate was observed (around 20%). In the case FFR‐0.75, the outage rate is considerably decreased (down to 1%) since the distance between co‐channel users in different cells is increased and hence the co‐channel interference dropped. The same trend is observed for the FFR‐0.5 scheme that has zero outage rate, which also coincides with the outage rate of the proposed REM‐based allocation. However, limiting the full frequency reuse users within a fixed part of each cell may lead to underutilization of resources. The FFR‐1 scheme allocates the same channel to all users in each cell, whereas the FFR‐0.75 and FFR‐0.5 schemes exclude around 25% and 50% of the users in each cell, respectively, from the common access channel. The proposed scheme, based on calculating the FPAC by using REM parameters related to the channel gains between the 3 APs and the 3 co‐channel users, outperforms both abovementioned schemes and leads to better utilization of resources, excluding a much lower percentage of users (around 13%) from the common channel allocation and leading to an optimum interference‐free frequency channel re‐use per user and cell area (see Figure 3‐3). Its drawback is its dependency on the stored context information that might have inaccuracies, which will have a negative impact on power allocation calculations and on frequency channel exploitation.
Figure 3-3: Frequency reuse utilization results
3.1.2.2 Two‐tier Cellular Networks with Full Frequency Reuse
Scenario and Overview
The most common deployment of Heterogeneous Networks are the 2‐tier network deployments, according to which the small cells (picocells, femtocells) will be deployed inside buildings in an area already covered by a macrocell. IM is the most critical challenge for a successful such deployment. The worst interference scenarios are when we have close access femtocells that co‐exist in the same frequency bands as the macro‐cell network. In Figure 3‐4 the common types of interference are depicted. This requires the design of sophisticated interference management techniques, exploiting relevant environment and context information, such as propagation characteristics in both macro and femto environments.
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 21/42
M2F: Interference from macro to femtocellF2F: Interference from femto to femtocellF2M: Interference from femto to macrocell
Interference
Transmission
F2F
M2F
M2F
F2M
Macro user with poor reception
Macro user with high
transmission power
1
Close femtocells to the macro BS
Figure 3-4: Type of interference in two-tier networks
One of the most challenging interference scenarios in femtocell networks is the downlink interference from the Femtocell AP (FAP) to the Mobile User Equipment (MUE) in the case of shared spectrum and closed access femtocells (Figure 3‐5). One approach to mitigate this type of interference is to apply on the FAP a power control mechanism based on the detection of ‘victim’ UEs in the area.
Figure 3-5: Two-tier cellular network downlink interference to MUE
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 22/42
Theoretical Approach and assessment
In [56] a FAP power control algorithm was proposed in order to reduce this type of interference. The objective is to provide protection to a “victim” MUE against FAP downlink interference while maintaining acceptable FAP performance. The algorithm was proposed for LTE femtocells and is summarized below:
The FAP (Home e‐NodeB (HeNB) in [56]) measures the Reference Signal Received Power (RSRP) and the SINR from the MBS and neighbouring FAPs.
Upon the detection of a neighbouring victim MUE, the FAP reduces its transmission power aiming to maintain a predefined SINR target for the MUE. The algorithm in the FAP assumes that the FAP and the MUE measure the same RSRP from the MBS and that the path loss distance between them has always a specific value (80dB).
In case the target cannot be achieved, the FAP transmits at a predefined minimum power level.
In order to highlight the potential gain of using rich context information in a two‐tier network deployment, this algorithm was modified in [57] by taking into consideration rich context information for the deployment area. The context information was used for the accurate calculation of the FAP transmit power and it was compared with the baseline algorithm of [56]. The accuracy and the level of the available context information depend on the deployment scenario. The fast fading term is highly dynamic and cannot be used from a pragmatic algorithm since it will be outdated. In [57] complete knowledge of the network component locations, the path loss and shadowing terms was assumed. While the location of MBS is known in advance, the FAP location needs to be estimated since the user may place the FAP randomly inside the house. The victim MUE position relative to the FAP and the building also has to be estimated. The path loss term for the calculation of the channel gains between the network elements is also assumed to be stored for a specific area based on long term operator measurements. A sensor network can be used to estimate the shadowing term, but this is not an easy task and it may result in inaccurate estimates.
In [57] it was shown that context‐aware power control offers better protection to the MUE than the baseline power control of [56]. The performance metric was the outage experienced by the “victim” MUE when it cannot meet its SINR target because of the FAP interference. As the femtocell is deployed closer to the cell edge, the gain in the MUE outage was reduced. This is expected since the provision for a minimum FAP Transmitter (Tx) power level has a bigger effect in the SINR calculation when the MBS signal is weaker. Also as expected, the outage increases as the MUE SINR target increases.
3.1.2.3 Channel allocation in two‐tier networks
Scenario and Overview
The scenario that is considered here is the same with the one described in Section 4.1.1.3 of D12.1 [1] "Coordinated and opportunistic channel access in LTE systems" (Reference scenario 1 – Cellular networks with cognitive capabilities (e.g. femtocells), Case Study 3: Interference management and scheduling for femtocells). It consists of a Heterogeneous Network (HetNet) with macrocells and femtocells that share a pool of N sub‐channels in an OFDMA‐based LTE system. The sub‐channels are split between the macrocells using a frequency reuse of 1/3. The femtocells are assumed to use the sub‐channels of the non‐covering cells as in [58] to increase spectrum efficiency (See Figure 3‐6).
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 23/42
Figure 3-6: Sub-channel allocation to femtocells and macrocells [58].
Since femtocell traffic is very dynamic, in general, dynamic sub‐channel allocation between macrocells and femtocells will not be able to cope with the traffic changes due to the slow connection between femtocells and the backbone, which is usually through land connection. Therefore, we proposed in D12.1 [1] an opportunistic sub‐channel access technique that allows the femtocell to access to the unused sub‐channels of the covering cell when more sub‐channels are needed by that femtocell to server its users. In addition, the technique allows the femtocells of the same building to share their allocation in order to avoid interference between femtocells that are very close. Therefore, we assumed the existence of a building‐REM (building Radio Environment Map) that contain the information about the location of the femtocells and their instantaneous sub‐channel allocation. Based on the REM information, the femtocells that are very close will use different sub‐channels.
Theoretical Approach and assessment
In D12.1 [1] we studied the performance of this approach using a Monte‐Carlo simulator. In this section, we show how such approach can be implemented in real system and evaluate its performance using a dynamic simulator. We assume an LTE downlink system using Frequency Duplex Division (FDD) with a sub‐frame of 1 ms where N = 14 OFDMA symbols can be transmitted.
In order to implement the algorithm in real system, the sub‐channel allocation of the covering macrocell should be determined by the femtocells. This can be done using two approaches:
By sensing the first n OFDMA symbols of the first slot (i.e. 0.5 ms) of a sub‐frame. In this approach we consider that n = 3.
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 24/42
By decoding the control channel Physical Downlink Control CHannel (PDCCH) that contains the Downlink Control Information (DCI). By decoding the latter, a node can assess the allocation of the sub‐channels to the users. In order to decode it, the femtocell should know the Radio Network Temporary Identifier (RNTI) of the macrocell users. This is possible since the dynamics of such information is slow, and therefore can be stored and updated in the building REM.
Since the two slots of a sub‐frame have the same sub‐channel allocation, knowing the allocation during the first n OFDMA symbols will allow the femtocell to know the allocation during the following N ‐ n OFDMA symbols. In order to enable the femto base station, or the HeNB in the 3GPP terminology, to determine the Macro Base Station (MBS) sub‐channel allocation, the approach requires that the HeNB frame is shifted by the duration n OFDMA symbols. In addition the HeNB will be only able to transmit during the first slot of the sub‐frame in the sub‐channels accessed opportunistically. By considering the OFDMA symbols used for control channels in the two time slots, the average ratio between the number of OFDMA symbols that can be used to transmit data in the first slot of an opportunistic channel to the one transmitted during a slot of a normal sub‐channel is 0.97.
We have evaluated the performance of the proposed approach using the same environment in D12.1 [1] for different configurations. The system consists of 7 tri‐sectorized macrocells. In each macrocell, one block of two strips of building is considered. We assume a fixed femtocell deployment ratio of 0.8 inside the building. The macro users are considered to be voice users, white the femtocell may have voice and data users. The characteristics of voice and data users are summarized in Table 3‐2 and Table 3‐3 [59]. The macro users are uniformly distributed between the 21 macro sectors. The average load in a sector is L and is different for different configurations. The femto users are also uniformly distributed between the femtocells that have an activation ratio a that depends on the configuration. The number of voice users in an active femtocell is in average 1. The number of data users d depends on the configuration. The mapping between throughput and the SINR is taken from Annex A of [60]. The transmit power of the MBS and the HeNB are 20 Watts and 100 mWatts, respectively. The number of Physical Resource Blocks (PRBs) allocated to each femtocell is 6 PRBs, in addition to a maximum of 3 opportunistic PRBs if possible. The load of a cell is the ratio between the number of users and the allocated PRBs.
Parameter Encoder frame length
Voice activity factor
Transition probability
Packet payload
Assigned PRBs
Satisfaction condition
Value 20 ms 50 % 0.01 40 bytes 1 98% of packets experience a delay less than 50 ms
Table 3-2: Voice traffic parameters
Parameter File size Reading time
Value Truncated log‐normal distribution
(mean, variance) = (2, 0.52) Mbytes
Maximum size = 5 Mbytes
Exponential distribution
Mean = 180 seconds
Table 3-3: Data traffic parameters
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 25/42
Each configuration is simulated for 4 minutes. The average data rate provided for a data user in the central cell when the macro cell load is 0.8 are depicted in Figure 3‐7, Figure 3‐8, Figure 3‐9 and Figure 3‐10. This is the worst case for the proposed algorithm since most of the macro sub‐channels will be allocated. Although it is a worst case, these figures show that the data rate provided to a user when the proposed approach is used is much higher than the one provided by the approach without opportunistic access, denoted here traditional access. The increase of the average data rate ranges between 22.2 % and 24 % depending on the number of data users and the activation ratio. It should be noted that for lower macrocell loads, the increase in data rate is more significant. In addition, for this case the number of unsatisfied voice users is very close for the two approaches.
0.5 1 1.5 2 2.5 3 3.50
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Data rate (Mbps)
Pro
ba
bili
ty
Tradional AccessOpportunistic Access
Figure 3-7: Data rate for l = 0.8, a = 0.3, and d = 1.
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 26/42
0.5 1 1.5 2 2.5 3 3.50
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
Data rate (Mbps)
Pro
ba
bili
ty
Tradional AccessOpportunistic Access
Figure 3-8: Data rate for l = 0.8, a = 0.6, and d = 1.
0.5 1 1.5 2 2.5 3 3.50
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Data rate (Mbps)
Pro
ba
bili
ty
Tradional AccessOpportunistic Access
Figure 3-9: Data rate for l = 0.8, a = 0.3, and d = 3.
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 27/42
0.5 1 1.5 2 2.5 3 3.50
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
Data rate (Mbps)
Pro
ba
bili
ty
Tradional AccessOpportunistic Access
Figure 3-10: Data rate for l = 0.8, a = 0.6, and d = 3
3.1.3 Energy‐aware spectrum sharing with femtocells
Scenario and Overview
In this work, we consider Orthogonal Frequency Division Multiplexing Access (OFDMA) based femtocells, distributed in indoor deployment and overlaid to the existing macrocell by spectrum sharing. For given sub–channels available at the femtocells, we propose a method for joint energy and spectrum resource (i.e., sub–channels) utilization in the indoor dense femtocell networks. This aims not only for the co–existence with the conventional macrocell, but also for an energy–aware implementation of multi–femtocells deployment. To this end, aggregate energy usage among femtocells is designed. Particularly, we provide methods of (i) mathematically formulating the aggregate amount of the energy usage by taking into account the cost of both the sub–channels at the femtocells and individual femtocell energy usage for the channel feedbacks and the data transmissions, (ii) finding the maximum amount of the energy usage per femtocell based on the aggregate interference (by the nearby femtocells) at the incumbent macrocell receiver, and (iii) balancing the cost of two energy usages by varying the number of active sub–channels per femtocell and their power level in a distributed manner. This is to improve the system performance while guaranteeing the interference rise at the incumbent receiver.
Theoretical Approach
Consider that there are L femtocells, each of which serves lN Femtocell User Equipments
(FUE) in a radius r and, for the downlink, L femtocells cognitively access the radio spectrum, licensed to the underlaid macrocell network. L femtocells are randomly distributed in a
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 28/42
small area (e.g., enterprise environment) on the coverage of the macrocell network and co–channel deployment of the femtocells causes interference rise at macrocell User Equipments (UEs) located near at the femtocells. Let each femtocell coordinate to each other and orthogonally operate over the available incumbent channels in the frequency domain while avoiding the co–tier interference among neighbouring femtocells.
For the downlink of each femtocell, we employ OFDM (e.g., 3GPP’s LTE). Particularly, let each FUE in femtocell l be allocated to n sub–channels (or, equivalently, sub–carriers) for given m incumbent radio channels, where it is assumed that m channels are originally licensed to a macrocell UE near the femtocells and for the simple analysis and without loss
of the generality, the value of m is given such that l lNmn / is an integer hereafter.
For given m inactive channels as well as the threshold of aggregate interference rise at the incumbent receiver from L femtocells, we propose a method to jointly allocate the power and the spectrum resources to coordinated femtocells. For corresponding decision‐making and management of the spectrum, we consider following decision problems:
How to share the m inactive primary channels among femtocells: In this problem, we equivalently focus on the issue that how much spectrum per FUE would be assigned out of m inactive primary channels. In this problem, the worst‐case interference will be considered such that the minimum impact toward the incumbent is guaranteed.
In the meantime, how to increase the system rate of femtocells. On every data transmission, we are aware of the energy usage at two subsequent planes inherent in the networked femtocells: (i) the channel feedbacks, and (ii) the data transmission. In the channel feedbacks plane, firstly, exchange of channel state information between FAPs and FUEs happens. In particular, for given n sub–channels per FUE, the FAP l randomly activates only a subset of n (≤ n ) sub–channels per FUE according to the Uniform distribution at every time slot. Then, among
lN FUEs at femtocell l, lN n sub–channels result in active for scheduling in a given
time slot. Only on such lN n sub–channels randomly activated, therefore, the
exchange of channel feedback information occurs for the purpose of radio resource
scheduling. Secondly, for given such lN n active sub–channels, each femtocell
employs the opportunistic radio resource scheduling scheme. That is, in the data plane, only the best among active sub–channels is motivated to allocate for the data transmission at femtocell l for all I at each time slot.
The main idea is that when there are less m primary channels inactive, we decide to allocate less number of the Resource Blocks (RBs) to femtocells, followed by the proper selection of the maximum power level per RB. Therefore, the resulting cross‐interference at the incumbent receiver (i.e., MBS) is reduced. Here, one RB is a set of subcarriers in the
OFDM‐based femtocell. Notice that, as a function of both such power allocation level per RB and the number of RBs, the sum rate of the coordinated femtocells is determined as well.
Based on these, we investigate a method that, for given m primary channels and L coordinated femtocells, the system performance of femtocells is enhanced. In particular, consider the case of using the outage–sensitive and the energy–limited femtocells application. In this case, we focus on maximizing the outage capacity per femtocell. The outage capacity used in this work is referred to as the maximum achievable sum rate R such that the probability that the sum rate per femtocell is less than or equal to the rate (R) is
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 29/42
less than or equal to a given threshold ϵ for ϵ > 0. To this end, we account for optimal selection cost (i.e., optimal number) of active sub–channels per FUE and their power allocation levels.
More details and the respective results will be provided in D12.4 “Description of Decision Making Approaches and Decision Making Engine”.
3.2 Learning‐based spectrum decision mechanisms
Accordingly to section 3.1, this section summarizes the proposed solutions that belong in the second category of section 2, i.e. solutions that are based on knowledge for addressing the decision making problem. The reasons of why each approach has been selected reside in the specific analysed scenario and can be found in the respective analysis of the mechanism.
3.2.1 Knowledge‐based cognitive Radio Resource Management
Scenario and Overview
The solution proposed for the knowledge‐based cognitive Radio Resource Management (RRM) is actually a management platform attributed with cognitive RRM capabilities, targeted to LTE compliant network segments. The cognitive features in the management process are introduced and the context acquisition mechanism architecture is presented.
Theoretical Approach
Introduction and Scope
Figure 3‐11 depicts the management functionality developed for enhancing RRM within LTE network segments [61].
Context. It reflects the status of the elements of the network segment, and the status of their environment. Each element monitoring procedures provide the traffic requirements, the mobility conditions, the configuration used, and the offered QoS levels.
Figure 3-11: Introduction and scope of cognitive features in the management process.
Profiles. It provides information on the capabilities of the elements and terminals of the segment, as well as the behaviour, preferences, requirements and constraints of users and applications. For users this part designates the applications required, the preferred QoS levels and the constraints regarding costs.
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 30/42
Policies. They designate rules and functionality that should be followed in context handling. Sample rules can specify allowed QoS levels per application, allocations of applications to RATs and assignments of configurations to transceivers.
Decision. The output is consisted of four main reconfiguration decisions. The first one is the carrier bandwidth assignment for each cell. The second is the number of sub‐carriers assignment per user. The third is the power allocation in the used sub‐carries used and finally the fourth is the modulation scheme that should be used for each subcarrier.
Optimization. It can be based on several Dynamic Sub‐carrier Assignment (DSA), Adaptive Power Allocation (APA) and Adaptive Modulation (AM) algorithms [62], [63] and machine learning techniques. The target of each one of these utility based algorithms is to maximize a utility function which reflects the most efficient configurations, considering the network environment parameters provided by Context.
Infrastructure Abstraction. Furthermore, the management infrastructure interfaces with the network segment through infrastructure abstraction which provides technology independent information on network parameters. This information is used for perceiving the context encountered in the network segment.
Learning. Learning functionalities embedded in the management infrastructure enhance the means for addressing complexity. The management entities are able to learn from the past interactions of the system with the environment, and identify situations addressed in the past. Thus, known solutions will be provided faster since complex optimization procedures can be skipped. Essentially, management components can find and gradually learn the best spectrum carriers that can be used for addressing certain situations.
Enhanced Context Acquisition for Cognitive Intra‐cell RRM. Figure 3‐12 presents the architecture of the enhanced context acquisition mechanism [61].
Figure 3-12: Enhanced context acquisition mechanism architecture for cognitive intra-cell RRM.
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 31/42
Reference Context Repository. It contains the information for each context addressed in the past. A registry table is used, in order to “remember” the solution for each context after the optimization procedure. In order to be clear on the difference between an arbitrary context derived from the network segment, and a context stored in the repository, the term “reference context” is used for the information stored in the repository.
Context Matching. The target of this module is to find the closest pattern for the new context. The algorithm that is used is based on the k‐Nearest Neighbour(s) (k‐NN) algorithm [64], [65], [66], and it is going to be presented in the next paragraphs.
Figure 3‐12 also shows the potential interactions between the modules. Interaction one is the starting phase in which the Context Acquisition module retrieves all the relevant information from the network segment. Through interaction two, the Context Matching and the Optimization modules are triggered. Context matching will use the repository data for finding if there is a reference context that is close to the current context. In parallel, the Optimization module can be triggered to start processing the context, as a “new” situation.
Through interaction three, the Context Matching module will pass the control to the Reconfiguration Enforcement or to the Optimization modules. The first is selected if a match is found. The second is done if no reference context is close to the new context. The Reconfiguration Enforcement may also pass the control to the Optimization module, through interaction four, in case the solution proposed by the Context Matching module cannot be applied. Through interaction five, the Optimization module will ask the Reconfiguration Enforcement module to apply the derived configuration to the network segment. Moreover, through interaction six, the context and the solution are sent to the repository, in order to ensure that if the same context derives again, the solution may be retrieved directly. In this way the management infrastructure has the ability to “learn” and apply “known” solutions reducing the time needed for context handling.
3.2.2 Radio Resource Management based on Dynamic Sub‐carrier Assignment (DSA), Context, Profiles and Policies
Scenario and Overview
In an Orthogonal Frequency Division Multiplexing Access based (OFDMA‐based) system, the available spectrum is divided into multiple narrowband, interference‐free sub‐carriers. The target of DSA algorithms is then to find the most appropriate assignment of those sub‐carriers to the users by taking into account their channel state information and resulting in an efficient multiple access schemes that increases the system performance as much as possible. As it is depicted in Figure 3‐13, multiple sub‐carriers are allowed to be assigned to a single user. The number of sub‐carriers needed to be assigned to a user may depend on several parameters like user location, requested service, user profile and Network Operator's (NO’s) policies. Moreover, information on the channel state, as seen by a specific user on each sub‐carrier is also considered in order to select the most appropriate modulation scheme in an adaptive manner.
Theoretical Approach
Future wireless communication systems cannot guarantee lower cost and better QoS levels to the users without proper network management functionalities targeted to that scope. Furthermore, management functionalities except for technological constraints and capabilities should also take into account NO’s policies and strategies in order to provide
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 32/42
optimized network decisions, while respecting operator’s business visions. Considering the above, it is quite easy to understand why this is a complex task that introduces time‐consuming procedures. However, when a problem is addressed to the service area, it should be solved the fastest possible in order to keep network Key Performance Indicators (KPIs), such as QoS, QoE and delay (more details will be provided in D12.4 “Description of Decision Making Approaches and Decision Making Engine”), in acceptable levels and keep the users satisfied for the services they experience. Thus, fast network adaptation to the environment changes should be characterized from faster and less complex procedures. The introduction of an efficient RRM scheme based on DSA algorithms tries to achieve this target.
Figure 3‐13 depicts an overview of the RRM scheme for wireless OFDMA network segments, which is similar to the scheme described in Section 3.2.1 and Figure 3‐12. As it can be shown, it is fed with input from modules which are classified as context acquisition, profiles management and policies derivation, respectively, and produces output by applying an optimization process [67]. The RRM scheme is interfacing with the wireless network segments. The latter provide information, expressed in a high level manner and are used by the management framework for perceiving the context encountered in the network segment. Another interface is needed to enforce the implementation of the actions dictated by the RRM scheme. The input as well as the optimization process and corresponding decision will be described in the following paragraphs.
Figure 3-13: Radio resource management scheme overview
Context acquisition. This component collects the status of the elements of the network segment, the status of their environment as well as user terminals [68], [69]. Essentially,
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 33/42
each element may use monitoring and discovery (sensing) procedures [70], which can be sensing based [71][72], and/or pilot channel based [73]. Monitoring procedures provide, for each network element of the segment and for a specific time period, the traffic requirements, the mobility conditions, the current configuration in terms operating parameters (e.g., spectrum assignment, power level, etc.) and the QoS levels offered. Furthermore, it includes procedures for capturing the channel state information which is reflected by the estimation of the mean SNR value for each sub‐carrier. Context information will be used from the system not only for the estimation of KPIs but also to address possible problems and trigger the optimization procedure whenever it is necessary.
Profiles management. This component provides information on the capabilities of the elements and terminals of the segment, as well as the behaviour, preferences, requirements, and constraints of users and applications. Essentially, this part designates the configurations on the operating parameters that will be checked for network elements and terminals. For users, this part designates the applications required, the preferred QoS levels, and the constraints regarding costs. This information is necessary during the optimization procedure in order to decide the most appropriate sub‐carrier assignment considering also current context information.
Policies derivation. The optimized decisions of the management functionalities should not only be feasible from technological perspective but also have to be aligned with NO’s policies and strategies. Policies designate rules that should be followed during context handling. Sample rules can specify allowed (or suggested) QoS levels per service and assignments of spectrum to transceivers.
Optimization. The target of the optimization process is to exploit all the available radio and network resources in order to achieve high bit rates providing users with the maximum possible QoS level. Part of the optimization procedure of the RRM scheme is the DSA technique which is expressed by the corresponding DSA algorithms. DSA algorithms try to find the best possible sub‐carriers assignment in order to serve users with the highest possible QoS level. However, an efficient RRM scheme will try to exploit also all the available management information in order to guarantee acceptable QoS levels to the users as well as to introduce fairness. In order for the aforementioned characteristics to be reflected from the sub‐carriers assignment decisions, the additional input of Context, Profiles, and Policies (CPP) from the RRM scheme should be considered.
Figure 3-14: Efficient RRM scheme incorporating DSA algorithms
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 34/42
Figure 3‐14 depicts how the DSA algorithm is encompassed inside the RRM scheme. Assuming that we have several users in our service area, where each one of them is allowed to request multiple services in parallel called sessions, context acquisition will be responsible to provide the traffic and network conditions in terms of the number of active sessions, the requested service for each session, the available sub‐carriers as well as their channel information in terms of their mean SNR values. The profiles management entity will be responsible to provide the user preferences in terms of the QoS level that must be achieved according to the requested service, based on what the users are willing to pay. The impact of this information to the DSA algorithms decision will be twofold. First, the RRM scheme will be capable to guarantee that the users will be served with the QoS level that they prefer. Second, fairness will be introduced, since in cases where unassigned subcarriers are available, they will be used only for sessions that are still below their preferred QoS level and will not be equally distributed among the sessions regardless of whether their preferred QoS level is achieved or not. Finally, the input from policies derivation entity will be considered as being complementary to the above. The NO’s policies can specify, for example, the thresholds for the minimum and maximum allowed QoS levels for each offered service. Thus, it will be guaranteed that all sessions will be served with an acceptable QoS level for each service and also that the system will be capable of serving new potential sessions.
Behaviour configuration. In general, decisions are targeted at producing feasible network configurations in terms of (1) sub‐carriers assignment to users’ sessions and (2) modulation rate assignment for each sub‐carrier based on its mean SNR value. Thus, the decision basically affects the application layer since its target is to guarantee fair QoS levels assignment as well as lower medium access control/ physical layers due to changes to the number of assigned subcarriers and their modulation type. Furthermore, in cases of high traffic demand, the decision may impose efficient traffic distribution to nearby cells by exploiting potentially free sub‐carriers. Framed within the above, the role of this entity is twofold. First, it is responsible to translate accordingly the optimization decision to the necessary configuration actions, considering the network elements reconfiguration capabilities. Thus, each network element’s capabilities will be fully exploited so as the corresponding configuration actions to be applied with the minimum possible overhead, delay, and cost. Second, all the necessary information targeted for learning purposes in terms of currently addressed network and radio environment conditions as well as the corresponding optimum decisions will be provided to Learning functional block in order to be further analyzed and processed.
Learning. Learning attributes will yield knowledge and experience. Such attributes may regard context acquisition (capability to identify previously tackled situations and their suitable solutions) [74], [75], [76], profiles management (certain user classes may require a specific number of sub‐carriers to be served), as well as policy derivation (storage of data on NO policies and exploitation of them in future situations). Knowledge and experience may help the RRM scheme to predict problems and act proactively to solve them. The functionalities of this entity will not be further analyzed since it is out of the scope of this paper and will be subject of our future work.
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 35/42
4. Conclusions This deliverable is the third deliverable of WP12 "Metric Identification, Decision Making Algorithms and Solutions", the main aim of which is to define and identify the procedures used in the analysis of the context data, to identify potential opportunities and to decide on the spectrum allocation. Moreover, this WP defines and identifies the key elements of spectrum sharing and decision making to allow intelligent and efficient choice of spectrum access, based on spectrum access policies and available or unused spectrum. WP12 includes a) definition of a framework for negotiation and decision making based on utility functions, utility function policies and constraints and pre‐classified context information, b) investigation of required profile definition and evaluation techniques, c) enhanced service negotiation and policy enforcement mechanisms and d) efficient spectrum allocation solutions.
Towards this direction, activity 12.5 "Solutions Categorisation" targeted at categorizing and providing the proposed mechanisms that will be incorporated in the decision making framework. This deliverable summarizes the outcome of this activity by presenting the categorization of the proposed solutions and an overview of them. In these terms, D12.3 has presented a) the state of the art of decision making mechanisms, b) the existing categorization of them and c) the approach of the proposed mechanisms that will address decision making issues posed by the scenarios that had been described in D12.1 [1].
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 36/42
5. Glossary and Definitions Acronym Meaning
3GPP 3rd Generation Partnership Project
AM Adaptive Modulation
ANN Artificial Neural Networks
AP Access Point
APA Adaptive Power Allocation
BN Bayesian Network
BS Base Station
CPP Context, Profiles and Policies
CR Cognitive Radio
CRM Cognitive Resource Manager
DCI Downlink Control Information
DSA Dynamic Sub‐carrier Assignment
FAP Femtocell Access Point
FDD Frequency Duplex Division
FFR Full Frequency Reuse
FPAC Feasible Power Allocation Criterion
FUE Femtocell User Equipments
GA Genetic algorithm
HeNB Home e‐NodeB
HetNet Heterogeneous Network
IM Interference Management
k‐NN k‐Nearest Neighbour
KPI Key Performance Indicator
LTE Long Term Evolution
MAB Multi‐Armed Bandit
MBS Macro Base Station
MDT Minimization of Drive‐Tests
MISO Multiple‐Input Single‐Output
MNN Multilayer Neural Network
MS Mobile Station
MUE Mobile User Equipment
NO Network Operator
OFDMA Orthogonal Frequency Division Multiplexing Access
PDCCH Physical Downlink Control CHannel
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 37/42
POMDP Partially Observable Markov Decision Process
PRB Physical Resource Block
PRP Pre‐emptive Resume Priority
PU Primary User
QoS Quality of Service
RA Resource Allocation
RB Resource Block
REM Radio Environmental Map
RL Reinforcement Learning
RNTI Radio Network Temporary Identifier
RRM Radio Resource Management
RSRP Reference Signal Received Power
SINR Signal to Interference‐plus‐Noise Ratio
SISO Single‐Input Single‐Output
SNR Signal‐to‐Noise Ratio
SoA State of the Art
SU Secondary User
Tx Transmitter
UCB Upper Confidence Bound
UE User Equipment
WiMax Worldwide Interoperability for Microwave Access
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 38/42
6. References [1] ICT‐ 257626 ACROPOLIS Project, Deliverable D12.1, “Specification of Preliminary Set of Appropriate
Metrics, Utility Functions and Layer Identification”, Sept. 2011, available at http://www.ict‐acropolis.eu/images/Deliverables/acropolis_d12_1_final.pdf
[2] ICT‐ 257626 ACROPOLIS Project, Deliverable D9.1, “Spectrum Sensing”, Sept. 2011, available at http://www.ict‐acropolis.eu/images/Deliverables/acropolis_d9.1_compiled_v1.0.pdf.
[3] Chung‐Wang Wang, Li‐Chung Wang, “Modeling and Analysis for Proactive‐Decision Spectrum Handoff in Cognitive Radio Networks,” ICC '09. IEEE International Conference on Communications 2009, Dresden, Germany, 2009.
[4] Chung‐Wang Wang, Li‐Chung Wang, Fumiyuki Adachi “Modeling and Analysis for Reactive‐Decision Spectrum Handoff in Cognitive Radio Networks,” GLOBECOM '10. IEEE International Conference on Communications 2010, Miami, Florida, USA, 2010
[5] D. Hamza, S. Aissa, “Secondary Access Based on Sensing and Primary ARQ Feedback in Spectrum Sharing Systems,” IEEE Wireless Communication and Networking Conference 2012 (WCNC 2012), Paris, France, 2012.
[6] Lingcen Wu, Wei Wang, Zhaoyang Zhang, “A POMDP‐based Optimal Spectrum Sensing and Access Scheme for Cognitive Radio Networks with Hardware Limitation, ” IEEE Wireless Communication and Networking Conference 2012 (WCNC 2012), Paris, France, 2012.
[7] Li‐Chun Wang et al.,” Load‐Balancing Spectrum Decision for Cognitive Radio Networks,” IEEE Journal on Selected Areas in Communications, vol. 29, no.4, pp. 757 – 769, 2011.
[8] Chao Yang et al., “Optimal Wideband Mixed Access Strategy Algorithm in Cognitive Radio Networks,” IEEE Wireless Communication and Networking Conference 2012 (WCNC 2012), Paris, France, 2012.
[9] I. Malanchini, M. Cesana, and N. Gatti, “On Spectrum Selection Games in Cognitive Radio Networks,” GLOBECOM '09. IEEE International Conference on Communications 2009, Honolulu, Hawaii, December 2009.
[10] A. Banaei and C. N. Georghiades, “Throughput Analysis of a Randomized Sensing Scheme in Cell‐based Ad‐hoc Cognitive Networks,” ICC '09. IEEE International Conference on Communications 2009, Dresden, Germany, 2009.
[11] Won‐Yeol Lee, I.F. Akyldiz, “A Spectrum Decision Framework for Cognitive Radio Networks,” IEEE Transactions on Mobile Computing, vol. 10, no. 2, pp. 161‐174, 2011.
[12] B. Canberk, I.F. Akyldiz, S. Oktug, “A QoS‐aware framework for available spectrum characterization and decision in Cognitive Radio networks,” IEEE 21st International Symposium on Personal Indoor and Mobile Radio Communications (PIMRC 2010), Istambul, Turkey, 2010.
[13] TW Rondeau, D Maldonado, D Scaperoth, CW Bostian, “Cognitive radio formulation and implementation”, in 1st International Conference on Cognitive Radio Oriented Wireless Networks and Communications (CROWNCOM 2006), Mykonos, Greece, 2006.
[14] W. Jouini, C. Moy, J. Palicot, “Decision making for cognitive radio equipment: analysis of the first 10 years of exploration", EURASIP Journal on Wireless Communications and Networking 2012:26, 2012.
[15] TW Rondeau, Application of Artiffcial Intelligence to Wireless Communications (PhD thesis, Virginia Tech, 2006).
[16] M Y. El Nainay, D.H. Friend, A.B. MacKenzie, “Channel Allocation & Power Control for Dynamic Spectrum Cognitive Networks using a Localized Island Genetic Algorithm,” 3rd IEEE Symposium on New Frontiers in Dynamic Spectrum Access Networks, 2008. DySPAN 2008. Chicago, USA, 2008.
[17] T. R. Newman, R. Rajbanshi, A. M. Wyglinski, J. B. Evans, and G. J. Minden, “Population adaptation for genetic algorithm‐based cognitive radios," International Conference on Cognitive Radio Oriented Wireless Networks and Communications (CROWNCOM 2006), Orlando, Florida, Aug. 2007.
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 39/42
[18] P. Mahonen, M. Petrova, J. Riihijarvi, and M. Wellens, “Cognitive wireless networks: Your network just became a teenager (poster)," in IEEE INFOCOM, 2006.
[19] P. Di Lorenzo, S. Barbarossa, “ Bio‐inspired swarming models for decentralized radio access incorporating random links and quantized communications,” 2011 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2011), Prague, Czech Republic, 2011.
[20] P. Di Lorenzo, S. Barbarossa, “ Distributed resource allocation in cognitive radio systems based on social foraging swarms,” 2010 IEEE Eleventh International Workshop on Signal Processing Advances in Wireless Communications (SPAWC 2010), Marrakech, Morocco, 2010.
[21] B. Wang, Y. Wu and K. J. R. Liu, “Game Theory for Cognitive Radio Networks: An Overview,” Elsevier Computer Networks, Vol. 54, No. 14, pp. 2537‐2561, 2010.
[22] D. Niyato, E. Hossain, “A Game‐Theoretic Approach to Competitive Spectrum Sharing in Cognitive Radio Networks,” IEEE Wireless Communication and Networking Conference 2007 (WCNC 2007), Hong Kong, 2007.
[23] B. Jankuloska, V. Atanasovski, L. Gavrilovska, “Novel spectrum sharing algorithm for maximizing supported WiFi‐like secondary users in TV white spaces” 18th European Wireless Conference 2012 European Wireless 2012, Poznan, Poland, 2012.
[24] E. Meshkova, J. Riihijärvi, A. Achtzehn, P. Mähönen, “Exploring Simulated Annealing and Graphical Models for Optimization in Cognitive Wireless Networks” GLOBECOM '09. IEEE International Conference on Communications 2009, Honolulu, Hawaii, December 2009.
[25] N. Baldo, M. Zorzi, “Fuzzy Logic for Cross‐layer Optimization in Cognitive Radio Networks”, IEEE Communications Magazine, Vol. 46, No. 4, pp. 64‐71, 2008.
[26] G. Quer, H. Meenakshisundaram, B. R. Tamma, B. S. Manoj, R. Rao and M. Zorzi, “Cognitive Network Inference through Bayesian Network Analysis,” in IEEE Globecom, Miami, FL, US, Dec. 2010.
[27] J. Qi, S. Zhang, L. Cao, Y. Sun and L. Tan, “Cognitive Networks Autonomic Decision‐making Approach Based on Influence Diagram,” Journal of Networks, Vol.7, No. 5: 876‐883 (2012)
[28] H. Li, Robert C. Qiu, “A Graphical Framework for Spectrum Modeling and Decision Making in Cognitive Radio Networks,” in GLOBECOM 2010: 1‐6.
[29] C. Wu, K. R. Chowdhury, M. Di Felice and W. Meleis, “ Spectrum management of cognitive radio using multi‐agent reinforcement learning,” in AAMAS 2010: 1705‐1712
[30] M.Di Felice, K. R. Chowdhury, C. Wu, L. Bononi and W. Meleis, “ Learning‐Based Spectrum Selection in Cognitive Radio Ad Hoc Networks,” in WWIC 2010: 133‐145.
[31] Yau, K.‐L.A., Komisarczuk, P., Teal, P.D.: A Context‐aware and Intelligent Dynamic Channel Selection Scheme for Cognitive Radio Networks. In: Proc. Of CROWNCOM’09, Hannover (2009)
[32] Jiang, T., Grace, D., Liu, Y.: Performance of Cognitive Radio Reinforcement Spectrum Sharing Using Different Weighting Factors. In: Proc. of CHINACOM’08, Hangzhou, pp. 1195–1199 (2008)
[33] Jouini W, Bollenbach R, Guillet M, Moy C, Nafkha A, “Reinforcement learning application scenario for Opportunistic Spectrum Access”, 54th IEEE International Midwest Symposium on Circuits & Systems (MWSCAS 2011), August 7‐10 2011, Seoul, South Korea
[34] W. Jouini, M. Di Felice, L. Bononi and C. Moy, “Collaboration and Coordination in Secondary Networks for Opportunistic Spectrum Access,” in CoRR abs/1204.3005: (2012)
[35] W. Jouini, D. Ernst, C. Moy, J. Palicot, "Multi‐Armed Bandit Based Policies for Cognitive Radio’s Decision Making Issues," in 3rd conference on Signal Circuits and Systems, Jerba, Tunisia, 6‐8 November 2009
[36] W. Jouini, D. Ernst, C. Moy, and J. Palicot. Upper confidence bound based decision making strategies and dynamic spectrum access. Proceedings of the 2010 IEEE International Conference on Communications (ICC), May 2010
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 40/42
[37] D. Jouini, C. Moy, and J. Palicot. Upper confidence bound algorithm for opportunistic spectrum access with sensing errors. Proceedings of the 6th International ICST Conference on Cognitive Radio Oriented Wireless Networks, (CrownCom 2011), June 2011.
[38] K Liu, Q Zhao, B Krishnamachari, Distributed learning under imperfect sensing in cognitive radio networks. ASILOMAR (Asilomar Conference Grounds Pacific Grove, CA, USA, 2010), pp. 671–675
[39] K Liu, Q Zhao, Distributed learning in cognitive radio networks: Multi‐armed bandit with distributed multiple players. Proc of ICASSP (San Diego, 2010), pp. 1–10
[40] W. Jouini, C. Moy and J.Palicot, “Decision making for cognitive radio equipment: analysis of the first 10 years of exploration,” in EURASIP Journal on Wireless Communications and Networking 2012, 2012:26
[41] R. Ruslan and T. C. Wan “Learning Ability in Cognitive Radio,” in NEtApps 2008, 2008
[42] Z. Zhang, G. Univ and G. X. Xie, “Intelligent cognitive radio: Research on learning and evaluation of CR based on Neural Network,” in 5th International Conference on Information and Communications Technology (ICICT 2007), 2007
[43] Z. Zhang and X. Xie, “Application research of evolution in cognitive radio based on GA,” 3rd IEEE Conference on Industrial Electronics and Applications (ICIEA 2008), 2008.
[44] N.Baldo and M. Zorzi, “Learning and Adaptation in Cognitive Radios Using Neural Networks,”.in 5th IEEE Consumer Communications and Networking Conference (CCNC 2008), 2008
[45] ICT‐ 257626 ACROPOLIS Project, Deliverable D4.1 “Report on Global Trends and Major Developments Roadmap”, March 2012, available at http://www.ict‐acropolis.eu/images/Deliverables/acropolis_d4_1.pdf
[46] J. Lv, R. Blasco‐Serrano, E. Jorswieck, R. Thobaben, and A. Kliks, “Optimal Beamforming in MISO Cognitive Channels with Degraded Message Sets,” in Proc. IEEE Wireless Communications and Networking Conference 2012 (WCNC 2012), Paris, France, April 2012.
[47] R. Blasco‐Serrano, J. Lv, R. Thobaben, E. Jorswieck, A. Kliks, and M. Skoglund, “Comparison of Underlay and Overlay Spectrum Sharing Strategies in MISO Cognitive Channels,” in Proc. 7th International Conference on Cognitive Radio Oriented Wireless Networks 2012 (CROWNCOM 2012), Stockholm, Sweden, June 2012.
[48] J. Lv, R. Blasco‐Serrano, E. Jorswieck, and R. Thobaben, “Linear Precoding in MISO Cognitive Channels with Causal Primary Message,” In Proc. Ninth International Symposium on Wireless Communication Systems 2012 (ISWCS 2012), Paris, France, Aug. 2012.
[49] J. Lv and E. Jorswieck, “Spatial shaping in cognitive system with coded legacy transmission,” in Proc. of International ITG Workshop on Smart Antennas (WSA), Feb. 2011.
[50] ICT‐ 257626 ACROPOLIS Project, Deliverable D8.2, “Machine Learning and Information Theoretic Techniques for Cognitive Radio and Networks”, Sept. 2011
[51] ICT‐ 257626 ACROPOLIS Project, Deliverable D13.1, “Adaptive and robust signal processing algorithms for directive mapping in interference networks”, March 2012, available at http://www.ict‐acropolis.eu/images/Deliverables/acropolis_d13_1.pdf
[52] M. Andrews, V. Capdevielle, A. Feki, and P. Gupta. “Autonomous spectrum sharing for mixed LTE femto and macro cells deployments”, in Proc. of INFOCOM WiP, 2010.
[53] A.K. Dey. Architectural Support for Building Context‐Aware Applications. PhD thesis, Georgia Institute of Technology, December 2000.
[54] N. Dimitriou, “Network Planning & Resource Management Issues for Mobile Multimedia CDMA Systems”, VTC 2004‐Spring, 2004.
[55] J. Jackson, J.Roy, V. Vaidehi, “Analysis of Frequency Reuse and Throughput Enhancement in WiMAX Systems”, Wireless Personal Communications, Springer Netherlands, April 2010.
[56] www.femtofotum.org: “Interference Management in OFDMA Femtocells”, white paper, March 2010
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 41/42
[57] A. Zalonis, N. Dimitriou, A. Polydoros, J, Nasreddine, P. Mähönen, “Femtocell Downlink Power Control based on Radio Environment Maps”, accepted in IEEE WCNC 2012, Paris, France, April 2012.
[58] T. Kim, T. Lee, “Throughput enhancement of macro and femto networks by frequency reuse and pilot sensing,” in IEEE International Performance, Computering and Communications Conference, Dec. 2008.
[59] F. Khan, LTE for 4G Mobile Broadband: Air Interface Technologies and Performance. Cambridge, U.K.: Cambridge University Press, 2009.
[60] 3GPP TR 36.942 V10.2.0, " Evolved Universal Terrestrial Radio Access (E‐UTRA); Radio Frequency (RF) system scenarios," Dec. 2010.
[61] Aggelos Saatsakis, George Dimitrakopoulos, Panagiotis Demestichas, “Enhanced Context Acquisition Mechanisms for Achieving Self‐Managed Congnitive Wireless Network Segments”, Proc. ICT‐Mobile Summit Conference, 2008.
[62] Ian C. Wong, Zukang Shen, Brian L. Evans, Jeffrey G. Andrews, “A Low Complexity Algorithm for Proportional Resource Allocation in OFDMA Systems”, IEEE Workshop on Signal Processing Systems, 2004.
[63] G. Song and Y. (G.) Li, “Cross‐layer optimization for OFDM wireless networks—Part II: Algorithm development,” IEEE Trans. Wireless Commun., vol. 4, no. 2, pp. 625‐634, Mar. 2005.
[64] Y. Zhan, H. Chen, G.‐C. Zhang,“An optimization algorithm of K‐NN classification”, in Proc. International Conference on Machine Learning and Cybernetics, 2006, pp. 2246‐2251.
[65] H. Samet, “K‐Nearest Neighbor Finding Using MaxNearestDist”, IEEE Transactions on Pattern analysis and Machine Intelligence, pp. 243‐252, 2008.
[66] X.‐G. YU, X.‐P. YU, “The research on an adaptive K‐NN classifier”, in Proc. Fifth International Conference on Machine Learning and Cybernetics, 2006.
[67] Dimitrakopoulos G, Demestichas P, Saatsakis A, Tsagkaris K, Galani A, Gebert J, Nolte K (2009) Functional architecture for the management and control of reconfigurable radio systems. Vehicular Technology Magazine, IEEE 4(no.1):42–48
[68] Van Sinderen MJ, Van Halteren AT, Wegdam M, Meeuwissen HB, Henk Eertink E (2006) Supporting context‐aware mobile applications. IEEE Communications Mag 44(No. 9):96–104
[69] Bellavista P, Corradi A, Montanari R, Tononelli A (2006) Context–aware semantic discovery for next generation mobile systems. IEEE Communications Mag 44(No. 9):62–71
[70] Tsagkaris K, Katidiotis A, Demestichas P (2008) Neural network based learning schemes for cognitive radio systems. Comput Commun 31(14):3394–3404
[71] Liu X, Shankar NS (2006) Sensing‐based opportunistic channel access. Mobile Networks and Applications Journal 11(No. 4):577–591
[72] Kim H, Shin KG (2008) Efficient discovery of spectrum opportunities with MAC‐layer sensing in cognitive radio networks. IEEE Trans on Mobile Computing 7(No. 5):533–545
[73] Perez‐Romero J, Sallent O, Agusti R, Giupponi L (2007) A novel on‐demand cognitive pilot channel enabling dynamic spectrum allocation. In Proc. 2nd International Symposium on New Frontiers in Dynamic Spectrum Access Networks 2007 (DySPAN 2007). Dublin, Ireland
[74] Saatsakis A, Dimitrakopoulos G, Demestichas P (2008) Enhanced context acquisition mechanisms for achieving self‐managed congnitive wireless network segments. Proc. ICT‐Mobile Summit Conference, Stockholm, Sweden
[75] Saatsakis A, Demestichas P (2009) Context matching for realizing cognitive wireless network segments. Wireless Personal Communications Journal, Springer. doi:10.1007/s11277‐009‐9807‐z
ICT- 257626 ACROPOLIS Date: 28/09/2012
ICT-ACROPOLIS Deliverable D12.3 42/42
[76] Saatsakis A, Tsagkaris K, von‐Hugo D, Siebert M, Rosenberger M, Demestichas P (2008) Cognitive radio resource management for improving the efficiency of LTE network segments in the wireless B3G world. DySPAN 2008 Conference, Chicago