IEEE Journal on Selected Areas in Communications

Federated learning has emerged recently as a promising solution for distributing machine learning tasks through modern networks of mobile devices. Recent studies have obtained lower bounds on the expected decrease in model loss that is achieved through each round of federated learning. However, convergence generally requires a large number of communication rounds, which induces delay in model training and is costly in terms of network resources. In this paper, we propose a fast-convergent federated learning algorithm, called $\mathsf {FOLB}$ , which performs intelligent sampling of devices in each round of model training to optimize the expected convergence speed. We first theoretically characterize a lower bound on improvement that can be obtained in each round if devices are selected according to the expected improvement their local models will provide to the current global model. Then, we show that $\mathsf {FOLB}$ obtains this bound through uniform sampling by weighting device updates according to their gradient information. $\mathsf {FOLB}$ is able to handle both communication and computation heterogeneity of devices by adapting the aggregations according to estimates of device’s capabilities of contributing to the updates. We evaluate $\mathsf {FOLB}$ in comparison with existing federated learning algorithms and experimentally show its improvement in trained model accuracy, convergence speed, and/or model stability across various machine learning tasks and datasets.

Proxy caching has been used to speed up Web browsing and reduce networking costs. In this paper, we study the extension of proxy caching techniques to streaming video applications. A trivial extension consists of storing complete video sequences in the cache. However, this may not be applicable in situations where the video objects are very large and proxy cache space is limited. We show that the approaches proposed in this paper (referred to as selective caching), where only a few frames are cached, can also contribute to significant improvements in the overall performance. In particular, we discuss two network environments for streaming video, namely, quality-of-service (QoS) networks and best-effort networks (Internet). For QoS networks, the video caching goal is to reduce the network bandwidth costs; for best-effort networks, the goal is to increase the robustness of continuous playback against poor network conditions (such as congestion, delay, and loss). Two different selective caching algorithms (SCQ and SCB) are proposed, one for each network scenario, to increase the relevant overall performance metric in each case, while requiring only a fraction of the video stream to be cached. The main contribution of our work is to provide algorithms that are efficient even when the buffer memory available at the client is limited. These algorithms are also scalable so that when changes in the environment occur it is possible, with low complexity, to modify the allocation of cache space to different video sequences.

We present a wideband land mobile satellite channel model based on a three-state Markov chain model valid for any kind of satellite constellation and for L, S, and Ka bands. Data from measurement campaigns was used to extract the parameters for the statistical distributions involved in the model. We also present a methodology describing how to use the proposed channel model to interface with system-level analysis. We use the time series generated by the model to parameterize power control (PC) imperfection as a function of the environment and user speed by acting upon the time series through a seven-level PC algorithm. We then apply the obtained parameterization to the analysis of a direct sequence-code division multiple-access scenario with users requiring different quality-of-service (QoS). It is shown that request of different QoS produces an unbalance in the users power distribution and less users can be served with the most demanding QoS. PC error worsens this effect and considerably reduces capacity efficiency which is highly dependent on the environment and mobile speed. Bit-error rate statistics are also investigated through the log-normal characterization obtained by using the channel model obtaining results on the effect of required E/sub b//(N/sub o/+I/sub o/) of a particular user.

A significant number of propagation and channel modeling papers have reported channel parameters like RMS delay spread, correlation bandwidth, and the Rician K factor, derived by various new methods from instantaneous, or snapshot measurements. This conflicts with the original definitions of these parameters, which formally should be derived from time averages, under an assumption of ergodicity and applied for the assessment of time-averaged link performance. It appears that the origins of fading channel characterization parameters, and the conditions under which they can be estimated and applied are now often ignored, leaving interpretation of new results subject to skepticism. This paper, therefore, provides a critical review of the estimation of frequency correlation and one of these parameters, correlation bandwidth, using channel impulse response estimates derived from propagation measurements. The practical application for knowledge of frequency correlation is summarized. Then, the derivation of the equation relating a channel's average power delay profile to its frequency correlation function via a Fourier transform is reviewed, with emphasis on conditions needed for validity. An alternate method, free from most of these conditions is also reported. Examples are given and comparisons are made of results from analyses using the two methods to estimate frequency correlation on Rayleigh and Rician mobile radio channels, which are shown to have significantly different frequency correlation characteristics. Finally, a measure of frequency correlation that is free from ambiguity concerning the value (e.g., 0.5, 0.75, or 1/e) at the correlation band edges is recommended.