Self-organizing networks (SONs) are expected to minimize operational and capital expenditure of the operators while improving the end users' quality of experience. To achieve these goals, the SON solutions are expected to learn from the environment and to be able to dynamically adapt to it. In this work, we propose a learning-based approach for self-optimization in SON deployments. In the proposed approach, the learning capability has the central role to perform the estimation of key performance indicators (KPIs) which are then exploited for the selection of the optimal network configuration. We apply this approach to the use case of dynamic frequency and bandwidth assignments (DFBA) in long-term evolution (LTE) residential small cell network deployments. For the implementation of the learning capability and the estimation of KPIs, we select and investigate various machine learning and statistical regression techniques. We provide a comprehensive analysis and comparison of these techniques evaluating the different factors that can influence the accuracy of the KPI predictions and consequently the performance of the network. Finally, we evaluate the performance of learning-based DFBA solution and compare it with the legacy approach and against an optimal exhaustive search for best configuration. The results show that the learning-based DFBA achieves on average a performance improvement of 33 % over approaches that are based on analytical models, reaching 95 % of the optimal network performance while leveraging just a small number of network measurements.
In this paper we describe an experimental testbed to empirically study the construction of Radio Environmental Maps (REMs) in indoor environments. The testbed allows investigating the characteristics and modeling of the radio environment for indoor scenarios. The deployed system is a network of over 80 heterogeneous wireless spectrum sensors with significantly different measurement capabilities in an office building consisting of multiple rooms. As application examples we consider two scenarios, one illustrating the indoor propagation conditions and another showing temporal aspects of primary node activities as observed by sensing devices. The observed phenomena strongly indicate that development of general radio environment map solutions for indoor use are extremely challenging, unless heterogeneity of spectrum sensors and non-linearity of propagation conditions is considered. Our measurement results advocate dynamic construction of REMs instead of static solutions. We strongly believe that the deployed testbed and obtained experimental data can further facilitate research in the area of REMs.
Abstract-In this paper we discuss the design of optimization algorithms for cognitive wireless networks (CWNs). Maximizing the perceived network performance towards applications by selecting appropriate protocols and carrying out cross-layer optimization on the resulting stack is a key functionality of any CWN. We take a "black box" approach to the problem and study the use of simulated annealing for solving it. To improve the convergence rate of the basic algorithm we apply machine learning techniques to construct graphical models on the perceived relations between network stack parameters and application-specific network utilities. We test our optimizer design both in a simulation environment as well as on a network testbed with low-power radios. Our results show that even basic simulated annealing works well, but simple graphical models can further increase the convergence rate. However, use of sophisticated models such as Bayesian networks does not always lead to substantially better performance. The results indicate that enhanced simulated annealing indeed appears to be a promising foundation for future cognitive engine designs.
We present the implementation architecture and performance evaluation of the nanoIP protocol stack. The stack consists of miniaturized versions of UDP, TCP, SLP and HTTP protocols with reduced header sizes and complexity to make the protocols usable in wireless sensor networks. Similarity to the TCP/IP stack facilitates the development of gateways towards IP-based networks and makes the use of the stack easier for developers accustomed to network programming. Our implementation work and experiments show that the footprint of the stack is acceptable even for the most resource constrained sensor nodes.
I. INTRODUCTIONMany mission critical applications for Wireless Sensor Networks (WSNs) are not useful when operated in isolation. For example, applications for detecting weapons of mass destruction in homeland security scenarios [1] and WSNs for surveillance missions used by police and military [2] obviously need to be connected to outside networks. Besides simple data transfer, such scenarios also impose the requirement of being able to remotely access and manage the sensor network. Therefore embedded protocols for WSNs should be chosen so that connection of them directly or through gateways toward internet is as simple as possible.As the TCP/IP protocols are standard solution for connectivity in the Internet using those either unmodified or with minimal modifications is the most straightforward approach for WSNs [3]. By applying these IP-based protocols programmers and network administrators do not need to learn new skills in order to make them functional. However, common constraints related to the nature of the WSN have to be tackled. For example, the wireless sensor nodes typically have very limited memory, power resource and computational abilities. These constraints have made the adaptation of any internet protocol for WSN use very challenging and more complicated than a simple porting procedure.The objective of the nanoIP stack is to offer the familiar TCP/IP transport and application layer services to application and system developers for WSNs with low overhead. Highlevel protocol design for the nanoIP stack protocols, nanoUDP, nanoTCP, nanoHTTP and nanoSLP was given in [4] and we now focus on reporting the implementation and performance
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.