This paper investigates the energy-efficient power scheduling (EEPS) problem in extended reality (XR) transmission with hard-latency constraints. Highly dynamic wireless channels and low packet dropout requirements render this a challenging non-convex stochastic constrained sequential decision problem, further complicated by XR’s multi-timeslot large-packet transmissions and non-stationary traffic. Traditional resource scheduling techniques are limited to simple and known traffic/channel models, while existing constrained reinforcement learning (CRL) algorithms lack theoretical guarantees for satisfying non-convex stochastic constraints and struggle to adapt to rapidly changing traffic dynamics. To address this, we propose a Context-aware Constrained Reinforcement Learning (CACRL) algorithm, consisting of a CRL module and a context inference (CI) module. The CRL module uses a policy network for EEPS decision-making and optimizes it through a novel constrained stochastic successive convex approximation (CSSCA) method, which effectively handles the original problem by solving a sequence of convex surrogate problems. The CI module integrates context-aware meta-learning and reward-reshaping mechanisms to infer varying traffic dynamics and transform sparse packet dropout signals caused by multi-timeslot transmissions into dense ones, guiding the CRL module to quickly converge under XR traffic. Theoretical analyses provide insights into the CACRL, while simulations demonstrate it outperforms advanced baselines in both power conservation and meeting packet dropout constraints.

In the emerging extended reality (XR) applications, downlink transmission struggles with tricky packet dropout issues caused by large-sized data packets and hard-latency constraints, usually requiring significant reliance on transmission power resources for support. This paper proposes an accelerated constrained reinforcement learning (ACRL)-based energy-efficient power scheduling (EEPS) algorithm for XR transmission to ensure a satisfactory packet dropout rate for each user with minimal power consumption. Based on a actorcritic framework, the proposed method employs a policy network to make energy-efficient power scheduling decisions online, with the critic module evaluating it through interactions with the environment, and the actor module optimizing it using the constrained stochastic successive convex approximation (CSSCA) method, which is particularly suitable for addressing non-convex stochastic constraints related to packet dropout rates. Moreover, we integrate a transfer learning method, policy reuse, into the actor module and apply a signal-reshaping mechanism to transform sparse delayed packet dropout signals into dense signals. Both techniques considerably accelerate the convergence speed of traditional constrained reinforcement learning (CRL) algorithms in XR scenarios. Simulation results demonstrate that the proposed ACRL-EEPS algorithm outperforms advanced baselines in both power conservation and meeting packet dropout constraints.

Salient Object Detection (SOD) aims to identify the most attractive objects in an image. To solve the problem that existing RGB-D SOD methods cannot fully utilize multimodal information to localize objects accurately, we propose a novel Refined Fusion Three-Branch network(RFNet). Firstly, the Comprehensive Attentive Fusion module is designed to encode the fusion of features from different modalities and suppress the background noise. Secondly, the Intermediate Refinement Connection module is designed to remove redundant multimodal information and refine the features. Finally, experiments on public benchmark datasets demonstrate the good performance of our method for both quantitative and qualitative evaluation. The source code is publicly available as https://github.com/Corgislam/RFNet-code

This paper presents an innovative downlink transmission power scheduling (TPS) scheme for the emerging delay-sensitive applications in the 6G era, focusing on enhancing the transmission efficiency while ensuring a high quality of service (QoS) for each user. Specifically, we first adopt the hard-delay constrained effective throughput and users’ packet dropout rates as performance metrics and formulate the TPS problem as a constrained Markov Decision Process (CMDP), which appropriately characterizes the requirements of delay-sensitive applications. Then, we propose a novel constrained deep Actor-Critic-based TPS (CDAC-TPS) algorithm, which can dynamically make TPS decisions by a policy network in its Actor module and evaluate the current policy by Q-networks in its Critic module without any prior information of the environment. In particular, the CDAC-TPS adopts a constrained stochastic successive convex approximation (CSSCA) method to optimize the policy network, which can better handle the stochastic non-convex objective and constraints in the TPS problem, while most of the existing policy optimization methods for CMDP are only suitable for simple convex constraints. Finally, simulation results demonstrate that the proposed TPS scheme outperforms baselines in both transmission efficiency and QoS guarantee.

Abstract Facial attractiveness prediction is an important research topic in the computer vision community. It not only contributes to the development of interdisciplinary research in psychology and sociology, but also provides fundamental technical support for applications like aesthetic medicine and social media. With the advances in 3D data acquisition and feature representation, this paper aims to investigate the facial attractiveness from deep learning and three‐dimensional perspectives. The 3D faces are first processed to unwrap the texture images and refine the raw meshes. The feature extraction networks for texture, point cloud, and mesh are then delicately designed, considering the characteristics of different types of data. A more discriminative face representation is derived by feature fusion for the final attractiveness prediction. During network training, the cyclical learning rate with an improved range test is introduced, so as to alleviate the difficulty in hyperparameter setting. Extensive experiments are conducted on a 3D FAP benchmark, where the results demonstrate the significance of deep feature fusion and enhanced learning rate in cooperatively facilitating the performance. Specifically, the fusion of texture image and point cloud achieves the best overall prediction, with PC, MAE, and RMSE of 0.7908, 0.4153, and 0.5231, respectively.

Radar sensing will be integrated into the 6G communication system to support various applications. In this integrated sensing and communication system, a radar target may also be a communication channel scatterer. In this case, the radar and communication channels exhibit certain joint burst sparsity. We propose a two-stage joint pilot optimization, target detection and channel estimation scheme to exploit such joint burst sparsity and pilot beamforming gain to enhance detection/estimation performance. In Stage 1, the base station (BS) sends downlink pilots (DP) for initial target search, and the user sends uplink pilots (UP) for channel estimation. Then the BS performs joint target detection and channel estimation. In Stage 2, the BS exploits the prior information obtained in Stage 1 to optimize the DP signal to further refine the performance. A Turbo Sparse Bayesian inference algorithm is proposed for joint target detection and channel estimation in both stages. The pilot optimization problem in Stage 2 is a semi-definite programming with rank-1 constraints. By replacing the rank-1 constraint with a tight and smooth approximation, we propose an efficient pilot optimization algorithm based on the majorization-minimization (MM) method. Simulations verify the advantages of the proposed scheme.

The demand bound function (DBF) is a powerful abstraction to analyze the feasibility/schedulability of real-time tasks. Computing the DBF for expressive system models, such as graph-based tasks, is typically very expensive. In this article, we develop new techniques to drastically improve the DBF computation efficiency for a representative graph-based task model, digraph real-time tasks (DRT). First, we apply the well-known quick processor-demand analysis (QPA) technique, which was originally designed for simple sporadic tasks, to the analysis of DRT. The challenge is that existing analysis techniques of DRT have to compute the demand for each possible interval size, which is contradictory to the idea of QPA that aims to aggressively skip the computation for most interval sizes. To solve this problem, we develop a novel integer linear programming (ILP)-based analysis technique for DRT, to which we can apply QPA to significantly improve the analysis efficiency. Second, we improve the task utilization computation (a major step in DBF computation for DRT) efficiency from pseudo-polynomial complexity to polynomial complexity. Experiments show that our approach can improve the analysis efficiency by dozens of times.