Multi-task reinforcement learning (MTRL) seeks to learn a unified policy for diverse tasks, but often suffers from gradient conflicts across tasks. Existing masking-based methods attempt to mitigate such conflicts by assigning task-specific parameter masks. However, our empirical study shows that coarse-grained binary masks have the problem of over-suppressing key conflicting parameters, hindering knowledge sharing across tasks. Moreover, different tasks exhibit varying conflict levels, yet existing methods use a one-size-fits-all fixed sparsity strategy to keep training stability and performance, which proves inadequate. These limitations hinder the model’s generalization and learning efficiency. To address these issues, we propose SoCo-DT, a Soft Conflict-resolution method based by parameter importance. By leveraging Fisher information, mask values are dynamically adjusted to retain important parameters while suppressing conflicting ones. In addition, we introduce a dynamic sparsity adjustment strategy based on the Interquartile Range (IQR), which constructs task-specific thresholding schemes using the distribution of conflict and harmony scores during training. To enable adaptive sparsity evolution throughout training, we further incorporate an asymmetric cosine annealing schedule to continuously update the threshold. Experimental results on the Meta-World benchmark show that SoCo-DT outperforms the state-of-the-art method by 7.6% on MT50 and by 10.5% on the suboptimal dataset, demonstrating its effectiveness in mitigating gradient conflicts and improving overall multi-task performance.

The real-time processing of large-scale, heterogeneous tasks—including CPU-only, general-purpose GPU, and specialized GPU tasks—poses significant challenges in Internet of Things (IoT) systems, driven by severe GPU resource fragmentation, inefficient CPU and memory resource utilization on edge servers. These issues often compromise system processing performance and server stability. To address these issues, we formulate a multi-stage mixed-integer nonlinear programming (MINLP) model, to jointly optimize GPU fragmentation rate and system processing capability. We then introduce a novel deviation-based Lyapunov optimization framework that explicitly maintains memory utilization around a predefined optimal threshold, effectively balancing resource usage and system stability. Finally, to achieve real-time decision-making for massive tasks in dynamic systems with randomly arriving tasks, we propose the MA-LHTO algorithm, a multi-agent deep reinforcement learning approach that incorporates a multi-head architecture, entropy-based exploration, and a parameter reset mechanism. Experimental results confirm that our algorithm significantly improves resource utilization, and exhibits good performance under various working conditions.

As wireless communication technology advances, the convergence of the Internet of Vehicles (IoV) and edge computing is increasingly underpinning the progression of intelligent driving systems. Despite these advances, challenges persist in the stability and efficiency of vehicle-edge collaborative computing frameworks. This study introduces a service availability and task complexity-aware scheduling framework for intelligent driving tasks (SATCAS-IDTs), ensuring the reliability of task scheduling between user vehicles and in-vehicle computing platforms (IVCPs) during high-speed mobility. Initially, a task complexity quantification model for intelligent driving is constructed, incorporating metrics such as information entropy, image contrast, energy, and color space. Subsequently, a vehicle mobility model tailored for high-speed multilane environments is designed, upon which the mobility and service availability aware IVCP initial screening algorithm (MSAIF-IVCP) is developed, preselecting IVCPs that meet mobility and service availability requirements. Lastly, the double deep recurrent reinforcement learning task scheduling algorithm with multiobservation hybrid encoder (DDRRL-MOHE) is proposed. By intelligently balancing task delay and processing accuracy, this algorithm dynamically adapts edge computing resources, effectively reducing response times and enhancing overall system performance. Experimental validation shows that our proposed method improves the task processing accuracy by about 1.70% on average and reduces the average latency by about 1.97% over the baseline method.

Achieving real-time processing of tasks has become a crucial objective in the Internet of Vehicles (IoV) field. During the online generation of tasks in IoV systems, many dependency tasks arrive randomly within continuous time frames, and it is impossible to predict the number of arriving tasks and the dependencies between sub-tasks. Offloading dependent tasks, which are quantity-intensive and have complex dependencies, to appropriate vehicle edge servers (VESs) for online processing of large-scale tasks remains a challenge. Firstly, we innovatively propose a VES task parallel processing framework incorporating a multi-level feedback queue to enhance the cross-slot parallel processing capabilities of the IoV system. Secondly, to reduce the complexity of problem-solving, we employ the Lyapunov optimization method to decouple the online task offloading control problem into single-stage mixed-integer nonlinear programming problem. Finally, we design an online task decision-making algorithm based on multi-agent reinforcement learning to achieve real-time task offloading decisions in complex dynamic IoV environments. To validate our algorithm’s superiority in dynamic IoV systems, we compare it with other online task offloading decision-making algorithms. Simulation results show that ours significantly reduces the all-task processing latency of IoV system by 15% compared to the comparison algorithms, and the task average latency time is reduced by 14%.

As smart cities evolve, rising computational demands strain infrastructures. Offloading tasks to edge cloud data centers offers potential but faces challenges like high latency, energy use, and data leakage, especially in dense urban areas. This paper presents a low-latency, energy-efficient digital twin (DT) architecture tailored for smart cities, integrating edge computing (EC) and multiple s (IRS) to enhance communication. Dynamic voltage and frequency scaling (DVFS) technology is considered for user devices to reduce energy consumption. To mitigate the risk of user privacy leakage during task offloading, we address sensitive user location data that may be exposed by proposing a perturbed sliding task queue (PSTQ) algorithm based on differential privacy (DP), and demonstrate the effectiveness of the algorithm. To optimize task processing time and energy efficiency, we decompose the complex problem using block coordinate descent and propose an intelligent scheduling for energy sustainability (ISES) algorithm based on Karush-Kuhn-Tucker conditions and deep reinforcement learning (DRL). Experimental results demonstrate that our proposed architecture and algorithms achieve over 90% improvement in key optimization objectives, alleviating the computational pressure on existing devices while significantly enhancing task processing efficiency and energy sustainability.