Virtual reality (VR) applications require wireless systems to provide a high transmission rate to support 360-degree video streaming to multiple users simultaneously. In this paper, we propose an intelligent reflecting surface (IRS)-aided rate-splitting (RS) VR streaming system. In the proposed system, RS exploits the shared interests of the users in VR streaming, and the IRS creates reflected channels to facilitate a high transmission rate. The IRS also mitigates the performance bottleneck caused by the requirement that all RS users have to be able to decode the common message. We formulate an optimization problem for maximization of the achievable bitrate of the streamed 360-degree video subject to the quality-of-service (QoS) constraints of the users. We propose a deep reinforcement learning (DRL)-based algorithm, in which we leverage imitation learning and the hidden convexity of the formulated problem to optimize the IRS phase shifts, RS parameters, beamforming vectors, and bitrate selection of the 360-degree video tiles. Simulations based on a real-world dataset show that the proposed IRS-aided RS VR streaming system outperforms two baseline schemes in terms of system sum-rate and average runtime.

The growing demand for virtual reality (VR) applications requires wireless systems to provide a high transmission rate to support 360-degree video streaming to multiple users simultaneously. In this paper, we propose an intelligent reflecting surface (IRS)-aided rate-splitting (RS) VR streaming system. In the proposed system, RS facilitates the exploitation of the shared interests of the users in VR streaming, and IRS creates additional propagation channels to support the transmission of high-resolution 360-degree videos. IRS also enhances the capability to mitigate the performance bottleneck caused by the requirement that all RS users have to be able to decode the common message. We formulate an optimization problem for maximization of the achievable bitrate of the 360-degree video subject to the quality-of-service (QoS) constraints of the users. We propose a deep deterministic policy gradient with imitation learning (Deep-GRAIL) algorithm, in which we leverage deep reinforcement learning (DRL) and the hidden convexity of the formulated problem to optimize the IRS phase shifts, RS parameters, beamforming vectors, and bitrate selection of the 360-degree video tiles. We also propose RavNet, which is a deep neural network customized for the policy learning in our Deep-GRAIL algorithm. Performance evaluation based on a real-world VR streaming dataset shows that the proposed IRS-aided RS VR streaming system outperforms several baseline schemes in terms of system sum-rate, achievable bitrate of the 360-degree videos, and online execution runtime. Our results also reveal the respective performance gains obtained from RS and IRS for improving the QoS in multiuser VR streaming systems.

In this paper, we formulate a joint uplink scheduling, phase shift control, and beamforming optimization problem in intelligent reflecting surface (IRS)-aided systems. We consider maximizing the aggregate throughput and achieving the proportional fairness as objectives. We propose a deep reinforcement learning-based user scheduling, phase shift control, beamforming optimization (DUPB) algorithm to solve the joint problem. The proposed DUPB algorithm applies the neural combinatorial optimization (NCO) technique to solve the user scheduling subproblem, in which a stochastic user scheduling policy is learned by deep neural networks with attention mechanism. Curriculum learning with deep deterministic policy gradient (CL-DDPG) is used in the proposed DUPB algorithm to jointly optimize the phase shift control and beamforming vectors. The knowledge on the hidden convexity of the joint problem is exploited to facilitate the policy learning in CL-DDPG. Simulation results show that, with the maximum aggregate throughput as the objective, the proposed DUPB algorithm achieves an aggregate throughput that is higher than the alternating optimization (AO)-based algorithms. Moreover, the throughput fairness among the users is improved when proportional fairness is used as the objective. The proposed DUPB algorithm outperforms the AO-based algorithms in terms of runtime when the number of reflecting elements is large.

Intelligent reflecting surface (IRS) and cell-free multiple-input multiple-output (CF-MIMO) systems are two promising multi-antenna technologies for the fifth generation and beyond (B5G) wireless communication systems. In this paper, we formulate a joint phase shift control and beamforming opti-mization problem to maximize the aggregate throughput subject to the reliability constraint of the users in an IRS-aided CF-MIMO system. We propose an alternating optimization (AO)-based algorithm, in which the joint problem is decomposed into a phase shift control subproblem and a beamforming subproblem. For the phase shift control subproblem, we propose a complex gradient descent (CGD)-based algorithm, which tackles the unit-modulus constraint and guarantees the aggregate throughput to be monotonic increasing in each iteration. We then propose a difference of convex programming (DCP)-based algorithm for beamforming optimization. Simulation results show that the proposed AO-based algorithm achieves an aggregate throughput that is 53.8% and 25.1% higher than the cellular MIMO system with zero-forcing beamformer and the IRS-aided CF-MIMO system with random phase shift control, respectively. Moreover, the reliability requirements of the users are satisfied with the proposed AO-based algorithm. Our results also demonstrate that the proposed algorithm improves the minimum throughput of the users and reduces the standard deviation of the throughput distribution.

Grant-free multiple access (GFMA) is a promising paradigm to efficiently support uplink access of Internet of Things (IoT) devices. In this paper, we propose a deep reinforcement learning (DRL)-based pilot sequence selection scheme for GFMA systems to mitigate potential pilot sequence collisions. We formulate a pilot sequence selection problem for aggregate throughput maximization in GFMA systems with specific throughput constraints as a Markov decision process (MDP). By exploiting multiagent DRL, we train deep neural networks (DNNs) to learn near-optimal pilot sequence selection policies from the transition history of the underlying MDP without requiring information exchange between the users. While the training process takes advantage of global information, we leverage the technique of factorization to ensure that the policies learned by the DNNs can be executed in a distributed manner. Simulation results show that the proposed scheme can achieve an average aggregate throughput that is within 85% of the optimum, and is 31%, 128%, and 162% higher than that of acknowledgement-based GFMA, dynamic access class barring, and random selection GFMA, respectively. Our results also demonstrate the capability of the proposed scheme to support IoT devices with specific throughput requirements.

Intelligent reflecting surface (IRS) is a promising paradigm for enhancing the spectrum efficiency of wireless communication systems. In this paper, we study the joint uplink scheduling and phase shift control in IRS-aided systems. We formulate the throughput maximization problem as a combinatorial optimization problem. We decompose the problem into two subproblems for user scheduling and phase shift control, respectively. We propose a neural combinatorial optimization (NCO)-based algorithm, in which a near-optimal stochastic policy for user scheduling is learned by deep neural networks (DNNs) with attention mechanism, while the phase shifts of the IRS are optimized using fractional programming. Unlike alternating optimization-based approaches which obtain a suboptimal solution by iteratively solving two subproblems, the proposed NCO-based algorithm is capable of obtaining a near-optimal solution while each subproblem is required to be solved only once. Simulation results show that the proposed NCO-based algorithm achieves an aggregate throughput which is within 98% of the exhaustive search algorithm, and outperforms both greedy scheduling and random scheduling algorithms.

Grant-free non-orthogonal multiple access (GF- NOMA) is a promising paradigm for reducing the access delay and improving the spectrum efficiency. As the signals of multiple users are superimposed in GF-NOMA systems, each user is required to select a user-specific pilot sequence to distinguish its own signal from the signals of other users. Packet collisions in the uplink occur when multiple users select the same pilot sequence. In this paper, we first formulate a pilot sequence selection problem for aggregate throughput maximization in GF-NOMA systems. We then design a deep reinforcement learning (DRL)- based distributed algorithm for each user to select its pilot sequence via learning from the past pilot sequence selections. The proposed algorithm does not rely on information exchange between the users and does not require centralized scheduling by the base station. Packet-level simulations show that, for the considered system parameters, the proposed DRL distributed algorithm can achieve an average aggregate throughput which is within 90% of the optimal value, and has a better performance than both acknowledgement-based and random selection GF-NOMA schemes.