We consider the energy-constrained task allocation problem in large-scale Aerial Edge Computing (AEC) systems, which encompasses a series of tightly coupled decision-making processes, includingwhichtasks need to be processed by unmanned aerial vehicles (UAVs),howto allocate these tasks and balance energy across UAVs for delay-sensitive requirements. However, little attention has been devoted to exploring the above coupled decision-making problem in AEC with various resource and energy constraints, which is further complicated by energy dynamics (UAV battery states), task-specific consumption, and allocation-feedback balance. In this paper, we formulate a multi-dimensional joint optimization problem, simultaneously optimizing task allocation and energy rewarding to maximize long-term system rewards while balancing service quality and energy efficiency. To this end, we propose a green aerial edge computing framework where partial UAVs are equipped with energy harvesting modules to collect ambient energy. To circumvent the intractable computational complexity arising from the coupled energy states of massive UAVs, we design a distributed solution method based on the mean field game, which decouples the dense multi-agent interactions into a game between an individual UAV and the aggregate population state, thereby transforming the complex global optimization problem into a set of equivalent scalable subproblems. We develop an optimal energy valuation scheme to guide UAV behavior. Numerical results show that our mechanism can effectively ensure sustainable system operation while maintaining high quality of service for metaverse users, outperforming existing methods in both system sustainability and service responsiveness.

We consider the efficient resource allocation task in the blockchain-based mobile edge computing (MEC) system that requires decentralized transaction management to validate transactions between edge servers (ESs) and mobile devices (MDs). In such task allocation process (where MDs' resources are limited and privacy-sensitive), it is a significant challenge to guarantee individual rationality with satisfactory system stability while enabling flexible task offloading under various locality constraints (e.g., communication distance, bandwidth and delay). In this paper, we formulate the target problem as a blockchain-assisted task-resource matching model, and then propose a near optimal locality-aware resource allocation mechanism over smart contract to enable automatic and efficient transactions in MEC system. More specifically, for the service agents selection, we design the preference-based selection strategy to get highest estimated profit. For the flexible task offloading, we develop the minimum delay task graph partitioning algorithm to determine the optimal task offloading solution for MD under different resource bundles. For the task-resource matching, we propose a task-resource matching game (based on potential game) with the second lowest cost strategy to determine the matching of task-resource and decide the price of resource bundle. For the transaction verification and block allocation, we propose a social welfare-driven consensus mechanism to enable verified transaction and fair block allocation in a reward-free way. Strict theoretical analysis and extensive simulations demonstrate that our mechanism guarantees individual rationality, Nash Equilibrium, and stable near optimal solution.

We consider the collaborative inference acceleration task via cloud-edge-end collaboration, which involves a series of tightly coupled decision-making steps, includingwhichDNN model to be selected,how muchto compress model,howto partition model, andwhereto offload partitioned submodels. In practical deployments, these decisions jointly affect both fine-tuning and inference performance, and must jointly account for such aspects as the model being used, the computational resources and local datasets available at each device, as well as network latencies, which significantly increases the complexity of optimizing the problem. Yet, no existing studies focus on such joint optimization problem for these tightly coupled decisions. In this paper, we model this problem as a multi-dimensional optimization problem, jointly optimizing collaborative inference and fine-tuning by selecting the DNN model, compression level, partition strategy, and computational resource allocation, with the objective of minimizing the overall energy consumption of the learning-inference process, subject to accuracy and latency constraints. To this end, we propose an algorithmic framework called JQODI combining a time-energy tree diagram to represent the learning process, a dynamic programming solution strategy, and a data-driven theoretical approach to predict the expected total number of training epochs that meet the accuracy requirements. We prove that JQODI approximates the optimal solution with polynomial complexity. Numerical results demonstrate that JQODI surpasses state-of-the-art methods in both energy efficiency and latency.

Offloading a plethora of end workflows to edge servers in mobile edge computing (MEC) systems involves a series of coupled decision-making steps, includinghow muchedge resources will be allocated for each workflow,whichsubtasks will be offloaded, andhowto determine edge-end transaction prices to ensure system stability. Particularly, these decisions must jointly account for workflow characteristics, the resources available at each edge server, as well as local constraints (e.g., communication distance, task latency, and bandwidth conditions), which again increases the difficulty of optimizing the problem. However, no existing study addresses such joint optimization problems for these tightly coupled decisions. To fill this gap, a minimum-delay workflow partitioning algorithm is first designed to determine the optimal task offloading solution under various resource conditions. Based on this algorithm, two locality-based social welfare maximization models (basic and dynamic) are constructed. Specifically, for basic model, a multi-stage task matching game with the second lowest cost strategy is developed to determine the resource selection and pricing. For the dynamic model with uncertain requests, an online learning algorithm is introduced to track the dynamic valuations of mobile devices and to ensure that the resulting task allocation solution achieves an upper-bounded regret. Strict theoretical analysis demonstrates that our mechanism guarantees individual rationality, Nash Equilibrium, and stable approximation ratio. Simulation results verify the effectiveness and efficiency of our mechanism, and show that the proposed mechanisms obtain at most 18% higher social welfare than existing studies.

Mixed Precision Quantization (MPQ) has become an essential technique for optimizing neural network by determining the optimal bitwidth per layer. Existing MPQ methods, however, face a major hurdle: they require a computationally expensive search for quantization strategies on large-scale datasets. To resolve this issue, we introduce a novel approach that first searches for quantization strategies on small datasets and then generalizes them to large-scale datasets. This approach simplifies the process, eliminating the need for large-scale quantization fine-tuning and only necessitating model weight adjustment. Our method is characterized by three key techniques: sharpness-aware minimization for enhanced quantized model generalization, implicit gradient direction alignment to handle gradient conflicts among different optimization objectives, and an adaptive perturbation radius to accelerate optimization. It offers advantages such as no intricate computation of feature maps and high search efficiency. Both theoretical analysis and experimental results validate our approach. Using the CIFAR10 dataset (just 0.5\% the size of ImageNet training data) for MPQ policy search, we achieved equivalent accuracy on ImageNet with a significantly lower computational cost, while improving efficiency by up to 150\% over the baselines.

One-shot neural architecture search (NAS) trains an over-parameterized network (termed as supernet) that assembles all the architectures as its subnets by using weight sharing for computational budget reduction. However, there is an issue of multi-model forgetting during supernet training that some weights of the previously well-trained architecture will be overwritten by that of the newly sampled architecture which has overlapped structures with the old one. To overcome the issue, we propose an orthogonal gradient learning (OGL) guided supernet training paradigm, where the novelty lies in the fact that the weights of the overlapped structures of current architecture are updated in the orthogonal direction to the gradient space of these overlapped structures of all previously trained architectures. Moreover, a new approach of calculating the projection is designed to effectively find the base vectors of the gradient space to acquire the orthogonal direction. We have theoretically and experimentally proved the effectiveness of the proposed paradigm in overcoming the multi-model forgetting. Besides, we apply the proposed paradigm to two one-shot NAS baselines, and experimental results demonstrate that our approach is able to mitigate the multi-model forgetting and enhance the predictive ability of the supernet with remarkable efficiency on popular test datasets.

Weight quantization is an effective technique to compress deep neural networks for their deployment on edge devices with limited resources. Traditional loss-aware quantization methods commonly use the quantized gradient to replace the full-precision gradient. However, we discover that the gradient error will lead to an unexpected zig-zagging-like issue in the gradient descent learning procedures, where the gradient directions rapidly oscillate or zig-zag, and such issue seriously slows down the model convergence. Accordingly, this paper proposes a one-step forward and backtrack way for loss-aware quantization to get more accurate and stable gradient direction to defy this issue. During the gradient descent learning, a one-step forward search is designed to find the trial gradient of the next-step, which is adopted to adjust the gradient of current step towards the direction of fast convergence. After that, we backtrack the current step to update the full-precision and quantized weights through the current-step gradient and the trial gradient. A series of theoretical analysis and experiments on benchmark deep models have demonstrated the effectiveness and competitiveness of the proposed method, and our method especially outperforms others on the convergence performance.