Tactical driving decision making is crucial for autonomous driving systems and has attracted considerable interest in recent years. In this paper, we propose several practical components that can speed up deep reinforcement learning algorithms towards tactical decision making tasks: 1) non-uniform action skipping as a more stable alternative to action-repetition frame skipping, 2) a counter-based penalty for lanes on which ego vehicle has less right-of-road, and 3) heuristic inference-time action masking for apparently undesirable actions. We evaluate the proposed components in a realistic driving simulator and compare them with several baselines. Results show that the proposed scheme provides superior performance in terms of safety, efficiency, and comfort.

Base station (BS) sleeping is an effective way to reduce the energy consumption of mobile networks. Previous efforts to design sleeping control algorithms mainly rely on stochastic traffic models and analytical derivation. However, the tractability of models often conflicts with the complexity of real-world traffic, making it difficult to apply in reality. In this paper, we propose a data-driven algorithm for dynamic sleeping control called DeepNap. This algorithm uses a deep Q-network (DQN) to learn effective sleeping policies from high-dimensional raw observations or un-quantized systems state vectors. We propose to enhance the original DQN algorithm with action-wise experience replay and adaptive reward scaling to deal with the challenges in nonstationary traffic. We also provide a model-assisted variant of DeepNap through the Dyna framework for inferring and simulating system dynamics. Periodical traffic modeling makes it possible to capture the nonstationarity in real-world traffic and the incorporation with DQN allows for feature learning and generalization from model outputs. Experiments show that both the end-to-end and the model-assisted version of DeepNap outperform table-based Q-learning algorithm and the nonstationarity enhancements improve the stability of vanilla DQN.

近年提出的云无线接入网,通过集中化改善了小区为中心无线接入网在成本和灵活性方面的瓶颈.但是,云无线接入网需要通过前传网大范围汇聚带宽极高的前传采样,造成了过高的建设与运营成本.本文针对云无线接入网过度依赖前传通信资源的问题,依据控制与数据分离的思想设计了一种软件定义超蜂窝网络新型架构,并提出了在该架构下通过通信与计算协同的资源部署大幅降低前传汇聚带宽的具体方案.首先基于排队论给出了虚拟基站池中计算资源统计复用增益与池规模的定量关系,并依据统计复用增益边际效应迅速递减的性质得到了部署中等规模基站池更经济的指导方针.然后基于图聚类框架针对基带处理功能分割部署问题给出了一种遗传算法,该算法给出的分割方案可以根据设计偏好在前传带宽成本与计算代价间进行灵活折衷.

Channel state information (CSI) plays an important role in next-generation cellular systems with massive multiple-input multiple-output (MIMO) technology as the indicator of wireless channels.In hypercellular networks (HCNs),traffic base stations (TBSs) improve energy efficiency by dynamical sleeping.However,conventional pilot-based CSI acquisition methods cannot be applied to sleeping cells.We propose a novel CSI scheme based on channel learning to address this problem.Unlike location-aided CSI acquisition schemes,the proposed method utilizes CSI at the control base station (CBS) as input to avoid errors caused by positioning.We validate our scheme in an HCN generated by the geometry-based stochastic channel model (GSCM).The prediction accuracy of the proposed scheme is better than the K-nearest neighbor (KNN) method and close to the location-aided CSI acquisition scheme,which requires the knowledge on user position.

Data from mobile cellular networks has been extremely valuable in helping us understand traffic dynamics, monitor network operations, and conceive better deployment plans. Among these applications, trace-driven emulation is particularly interesting due to its potential in enabling open, reproducible, and cost-effective examination of innovative designs of cellular networks. Nevertheless, the inanimate nature of off-line data restricts it from being directly used for testing interactive network control algorithms, which may change the original data distribution. To address this issue, we present the DragonEye interactive emulation framework. DragonEye combines offline network data with microscopic refining and reacting models to emulate the live interaction between mobile users and the cellular network. For demonstration, we implement DragonEye using session-level user traffic logs captured from a real WLAN. This implementation is then used to test a reinforcement-learning-based base station (BS) sleeping control algorithm. Results show that emulated users can interactively queue and cancel requests in response to dynamic sleeping operations, which demonstrates the effectiveness of DragonEye in emulating live interaction using offline data.

In this paper, we propose a learning-based low-overhead channel estimation method for coordinated beamforming in ultra-dense networks. We first show through simulation that the channel state information (CSI) of geographically separated base stations (BSs) exhibits strong non-linear correlations in terms of mutual information. This finding enables us to adopt a novel learning-based approach to remotely infer the quality of different beamforming patterns at a dense-layer BS based on the CSI of an umbrella control-layer BS. The proposed scheme can reduce channel acquisition overhead by replacing pilot-aided channel estimation with the online inference from an artificial neural network, which is fitted offline. Moreover, we propose to exploit joint learning of multiple CBSs and involve more candidate beam patterns to obtain better performance. Simulation results based on stochastic ray-tracing channel models show that the proposed scheme can reach an accuracy of 99.74% in settings with 20 beamforming patterns.

Recently, the cloud radio access network (C-RAN) architecture has been proposed to enhance the cost effectiveness and flexibility of traditional cell-centric radio access networks. However, the massive fronthaul bandwidth required to centralize baseband computations in C-RAN results in extremely high costs. This paper summarizes our previous efforts toward solving this problem. We proposed the software-defined hyper-cellular network (SDHCN) based on the control/data separation principle. Under the proposed SDHCN framework, we studied two mechanisms that can greatly reduce fronthaul costs through the joint deployment of communicational and computational resources. First, we quantitatively characterized the relationship between the size of virtual base station (VBS) pools and the gains from computational statistical multiplexing by using queueing theory. We then showed that the marginal gain diminishes quickly with a growing pool size. Therefore, it is most economical to deploy mid-sized VBS pools. Finally, we proposed a genetic algorithm for baseband function splitting within a graph-clustering framework. This algorithm provides splitting schemes that can flexibly achieve different tradeoffs between fronthaul and computational costs based on different design preferences.

Cloud radio access network (C-RAN) was proposed recently to reduce network cost, enable cooperative communications, and increase system flexibility through centralized baseband processing. By pooling multiple virtual base stations (VBSs) and consolidating their stochastic computational tasks, the overall computational resource can be reduced, achieving the so-called statistical multiplexing gain. In this paper, we evaluate the statistical multiplexing gain of VBS pools using a multi-dimensional Markov model, which captures the session-level dynamics and the constraints imposed by both radio and computational resources. Based on this model, we derive a recursive formula for the blocking probability and also a closed-form approximation for it in large pools. These formulas are then used to derive the session-level statistical multiplexing gain of both real-time and delay-tolerant traffic. Numerical results show that VBS pools can achieve more than 75% of the maximum pooling gain with 50 VBSs, but further convergence to the upper bound (large-pool limit) is slow because of the quickly diminishing marginal pooling gain, which is inversely proportional to a factor between the one-half and three-fourth power of the pool size. We also find that the pooling gain is more evident under light traffic load and stringent quality of service requirement.

The fronthaul is an indispensable enabler for 5G networks. However, the classical fronthauling method demands large bandwidth, low latency, and tight synchronization from the transport network, and only allows for point-to-point logical topology. This greatly limits the usage of fronthaul in many 5G scenarios. In this article, we introduce a new perspective to understand and design fronthaul for next-generation wireless access. We allow the renovated fronthaul to transport information other than time-domain I/Q samples and to support logical topologies beyond point-to-point links. In this way, different function splitting schemes can be incorporated into the radio access network to satisfy the bandwidth and latency requirements of ultradense networks, control/data decoupling architectures, and delay-sensitive communications. At the same time, massive cooperation and devicecentric networking could be effectively enabled with point-to-multipoint fronthaul transportation. We analyze three unique design requirements for the renovated fronthaul, including the ability to handle various payload traffic, support different logical topology, and provide differentiated latency guarantee. Following this analysis, we propose a reference architecture for designing the renovated fronthaul. The required functionalities are categorized into four logical layers and realized using novel technologies such as decoupled synchronization layer, packet switching, and session-based control. We also discuss some important future research issues.

User behaviour analysis based on traffic log in wireless networks can be beneficial to many fields in real life: not only for commercial purposes, but also for improving network service quality and social management. We cluster users into groups marked by the most frequently visited websites to find their preferences. In this paper, we propose a user behaviour model based on Topic Model from document classification problems. We use the logarithmic TF-IDF (term frequency - inverse document frequency) weighing to form a high-dimensional sparse feature matrix. Then we apply LSA (Latent semantic analysis) to deduce the latent topic distribution and generate a low-dimensional dense feature matrix. K-means++, which is a classic clustering algorithm, is then applied to the dense feature matrix and several interpretable user clusters are found. Moreover, by combining the clustering results with additional demographical information, including age, gender, and financial information, we are able to uncover more realistic implications from the clustering results.

The baseband-up centralization architecture of radio access networks (C-RAN) has recently been proposed to support efficient cooperative communications and reduce deployment and operational costs. However, the massive fronthaul bandwidth required to aggregate baseband samples from remote radio heads (RRHs) to the central office incurs huge fronthauling cost, and existing baseband compression algorithms can hardly solve this issue. In this paper, we propose a graph-based framework to effectively reduce fronthauling cost through properly splitting and placing baseband processing functions in the network. Baseband transceiver structures are represented with directed graphs, in which nodes correspond to baseband functions, and edges to the information flows between functions. By mapping graph weighs to computational and fronthauling costs, we transform the problem of finding the optimum location to place some baseband functions into the problem of finding the optimum clustering scheme for graph nodes. We then solve this problem using a genetic algorithm with customized fitness function and mutation module. Simulation results show that proper splitting and placement schemes can significantly reduce fronthauling cost at the expense of increased computational cost. We also find that cooperative processing structures and stringent delay requirements will increase the possibility of centralized placement.

Wireless communication networks rely heavily on channel state information (CSI) to make informed decision for signal processing and network operations. However, the traditional CSI acquisition methods is facing many difficulties: pilot-aided channel training consumes a great deal of channel resources and reduces the opportunities for energy saving, while location-aided channel estimation suffers from inaccurate and insufficient location information. In this paper, we propose a novel channel learning framework, which can tackle these difficulties by inferring unobservable CSI from the observable one. We formulate this framework theoretically and illustrate a special case in which the learnability of the unobservable CSI can be guaranteed. Possible applications of channel learning are then described, including cell selection in multi- tier networks, device discovery for device-to-device (D2D) communications, as well as end-to-end user association for load balancing. We also propose a neuron-network-based algorithm for the cell selection problem in multi-tier networks. The performance of this algorithm is evaluated using geometry-based stochastic channel model (GSCM). In settings with 5 small cells, the average cell-selection accuracy is 73% - only an 3.9% loss compared with a location-aided algorithm which requires genuine location information.

Accurate mobile traffic forecast is important for efficient network planning and operations. However, existing traffic forecasting models have high complexity, making the forecasting process slow and costly. In this paper, we analyze some characteristics of mobile traffic such as periodicity, spatial similarity and short term relativity. Based on these characteristics, we propose a Block Regression (BR) model for mobile traffic forecasting. This model employs seasonal differentiation so as to take into account of the temporally repetitive nature of mobile traffic. One of the key features of our BR model lies in its low complexity since it constructs a single model for all base stations. We evaluate the accuracy of BR model based on real traffic data and compare it with the existing models. Results show that our BR model offers equal accuracy to the existing models but has much less complexity.

Cellular networks are one of the cornerstones of our information-driven society. However, existing cellular systems have been seriously challenged by the explosion of mobile data traffic, the emergence of machine-type communications, and the flourishing of mobile Internet services. In this article, we propose CONCERT, a converged edge infrastructure for future cellular communications and mobile computing services. The proposed architecture is constructed based on the concept of control/data (C/D) plane decoupling. The data plane includes heterogeneous physical resources such as radio interface equipment, computational resources, and software-defined switches. The control plane jointly coordinates physical resources to present them as virtual resources, over which software-defined services including communications, computing, and management can be deployed in a flexible manner. Moreover, we introduce new designs for physical resources placement and task scheduling so that CONCERT can overcome the drawbacks of the existing baseband-up centralization approach and better facilitate innovations in next-generation cellular networks. These advantages are demonstrated with application examples on radio access networks with C/D decoupled air interface, delay-sensitive machine-type communications, and real-time mobile cloud gaming. We also discuss some fundamental research issues arising with the proposed architecture to illuminate future research directions.

Facing the explosion of mobile data traffic, cloud radio access network (C-RAN) is proposed recently to overcome the efficiency and flexibility problems with the traditional RAN architecture by centralizing baseband processing. However, there lacks a mathematical model to analyze the statistical multiplexing gain from the pooling of virtual base stations (VBSs) so that the expenditure on fronthaul networks can be justified. In this paper, we address this problem by capturing the session-level dynamics of VBS pools with a multi-dimensional Markov model. This model reflects the constraints imposed by both radio resources and computational resources. To evaluate the pooling gain, we derive a product-form solution for the stationary distribution and give a recursive method to calculate the blocking probabilities. For comparison, we also derive the limit of resource utilization ratio as the pool size approaches infinity. Numerical results show that VBS pools can obtain considerable pooling gain readily at medium size, but the convergence to large pool limit is slow because of the quickly diminishing marginal pooling gain. We also find that parameters such as traffic load and desired Quality of Service (QoS) have significant influence on the performance of VBS pools.