Federated learning (FL) enables cross-institution model training under privacy constraints but is prone to catastrophic forgetting in continual learning scenarios. Many federated continual learning (FCL) methods rely on rehearsal (storing or generating past data), which conflicts with medical privacy requirements. We propose FedOCP-LRAT, a rehearsal-free FCL framework. The server constructs a global important subspace from past tasks and preserves prior tasks knowledge by constraining new parameter updates to its orthogonal complement. To facilitate the learning on new tasks, we introduce a trainable scaling matrix to reuse knowledge from old tasks. To reduce the communication overhead introduced by the scaling matrix, we propose a low-rank alternating training strategy to optimize the scaling matrix. Experiments on two medical datasets and CIFAR100 show consistent improvements over strong baselines, enabling privacy-compliant federated continual learning without retaining historical patient data.

Understanding protein function is essential for decoding cellular mechanisms and addressing a wide range of biological challenges. However, current protein function prediction methods face significant limitations due to data silos, heterogeneity, and privacy constraints, which hinder the integration and utilization of large-scale, multimodal datasets across institutions. Moreover, existing approaches lack personalization, making them less effective in adapting to institution-specific data characteristics and prediction needs. To address these challenges, we introduce a novel framework based on personalized federated learning that enables collaborative model training without compromising data privacy. The proposed framework, named PFLGO, incorporates a multi-center aggregation strategy and leverages largescale pre-trained models with knowledge transfer mechanisms to support efficient and accurate protein function prediction. By integrating heterogeneous multimodal data-including sequence, structure, and contextual embeddings-from multiple institutions, our method significantly enhances generalization performance. Extensive experiments conducted on two benchmark datasets demonstrate that PFLGO consistently outperforms both traditional centralized and existing multimodal approaches in terms of accuracy, convergence speed, and communication efficiency. This work highlights the potential of federated learning in advancing collaborative, privacy-preserving protein function prediction.

With the rapid development of intelligent transportation systems, autonomous driving (AD) is gradually becoming the primary mode of transportation for the future. However, safety still remains the critical challenge for the widespread adoption of automated vehicles. The ego vehicle is subject to significant safety risks, primarily due to its limited sensing capabilities and insufficient global situational awareness. To enhance the safety of AD, researchers have proposed two types of frameworks: Vehicle-to-Vehicle (V2V) and Vehicle-to-Infrastructure (V2I). Although these frameworks can improve vehicle safety, they still face issues such as multi-node data transmission delays and a lack of infrastructure flexibility, which can impair the vehicle’s ability to effectively perceive risks in complex environments. In recent years, unmanned aerial vehicle (UAV) has attracted considerable attention from researchers in the AD field, primarily due to its mobility and expansive field of view. The advantages of UAV bring hope for solving the drawbacks of traditional frames. In this context, this paper introduces UAV into the AD environment and proposes an innovative unmanned aerial vehicle assisted autonomous driving (U2AD) framework to address these limitations and enhance vehicles’ risk perception and decision-making capabilities. Experimental results show that U2AD, compared to V2V and V2I, increases the average perception range of vehicles by 27.5% and reduces the average collision rate by 14.1%, fully demonstrating its tremendous potential in the field of AD.

Due to the scarcity of high-quality data, the employment of Federated Learning (FL) for the fine-tuning task of Large Language Models (LLMs) has turned into a research hotspot. However, the low resource capacity of clients and the high memory requirements for LLM fine-tuning pose challenges for the federated fine-tuning of LLMs. The most advanced existing method, FedMeZO, utilizes MeZO’s forward pass optimizer to fine-tune LLMs on clients, using only inference-level GPU memory. However, the randomness of the perturbation vectors in the MeZO optimizer may lead to training oscillations, which hinders the convergence speed and makes overfitting more likely. Inspired by the momentum in backpropagation, this study proposes FedAdamZO, which replaces traditional gradient momentum with perturbation vectors, combining adaptive momentum and perturbation sampling uncertainty to enhance the smoothness of gradient estimation and model parameter changes. Experiments on multiple LLMs and datasets demonstrate that this method improves convergence speed and maintains good model performance.

Federated learning (FL) in edge computing scenarios faces critical challenges such as data heterogeneity and limited communication resources. The traditional FL global model has limited generalization capabilities, resulting in poor performance on individual edge nodes. To address this, we propose a semi-asynchronous personalized federated learning framework, named MHFL. MHFL utilizes deep reinforcement learning (DRL) to dynamically adapt to the FL environment (e.g., network conditions, data distribution, and workload distribution), generating optimized migration strategies to enhance personalized performance, alleviate limited communication resources, and mitigate the impact of non-independent and identically distributed (Non-IID) data. Furthermore, MHFL adopts a semi-asynchronous aggregation strategy, allowing edge nodes to transmit their models to the server without the need to synchronize all devices, thereby avoiding straggler issues, balancing training loss and latency, and further optimizing personalization while reducing communication costs. The experimental results demonstrate that, compared to existing methods, MHFL achieves an improvement in accuracy of up to 15.8% and reduces communication costs by up to 40.3%.

Video frame interpolation (VFI) aims to generate intermediate frames between existing ones, enabling applications such as frame rate upscaling and video editing. Most existing methods focus on improving interpolation quality, but they neglect interpolation efficiency, making them less suitable for real-time applications and other performance-constrained scenarios. Although some research has focused on improving efficiency, their interpolation quality remains suboptimal. Thus, achieving a better balance between interpolation quality and efficiency continues to be a significant challenge. In this paper, we propose a novel hierarchical framework that combines horizontal and vertical channel mixing with token mixing to capture global contextual features. Horizontal and vertical channel mixing effectively extracts local features, while token mixing captures global dependencies. Experimental results on several benchmark datasets, including Vimeo90K, UCF101, and Xiph, demonstrate that our method significantly reduces model parameters and computational complexity with minimal sacrifice in interpolation quality. Our framework provides a promising solution for real-time VFI applications and hardware-constrained scenarios.

Federated Continual Learning (FCL) aims to address the issue of Global Catastrophic Forgetting (GCF) in Federated Learning (FL) under dynamic data scenarios. Although various approaches have been proposed in FCL research to mitigate forgetting, most of them belong to the category of rehearsal-based approaches, relying on replaying historical data during learning the new task. However, this contravenes the principle of data forgettability. To address this issue, this paper proposes a rehearsal-free federated continual learning framework, namely Federated Orthogonal Projection with Projection Compensation, FOP-PC. The framework extracts the Global Input Subspace (GIS, the important representation space of old tasks) and constrains the parameters update of the new task within a space orthogonal to GIS, thereby minimizing interference from the new task on old tasks and avoiding global catastrophic forgetting. Meanwhile, the framework constructs a Similar Input Subspace (SIS) for the new task, its each layer is composed of the corresponding layer from the GIS of the most similar old task, and compute projection compensation to facilitate the learning of the new task. Experiment shows that, on three different datasets, FOP-PC outperforms the state-of-the-art FCL method, achieving accuracy improvements of 1.23%, 7.97%, and 5.44%, and reducing the forgetting by 0.1%, 1.12%, and 0.22%.

Mobile edge computing (MEC) has been widely applied in various Internet of Things (IoT) and mobile applications. In the event of network failures that prevent endusers (EUs) from establishing connections with MEC servers (MECSs), unmanned aerial vehicles (UAVs) can be deployed to provide computing services and data transmission. In UAV- assisted MEC systems, task offloading and resource allocation are two critical issues that are closely related and mutually influential, jointly determining the effectiveness of the system. Moreover, due to the complexity of UAV-assisted MEC systems, traditional methods cannot effectively address these two problems, and most existed research fails to fully consider them. In this paper, we consider a scenario of UAV-assisted MEC and propose an innovative deep reinforcement learning (DRL) approach that specifically targets the distinct characteristics of task offloading and resource allocation problems. We utilize the Deep Deterministic Policy Gradient (DDPG) method to solve the task offloading problem and introduce Long Short-Term Memory (LSTM) into DDPG to propose an LSTM-based DDPG (LDPG) method for resource allocation. By allowing partial parameter sharing between these two modules and alternating training, we effectively address the task offloading and resource allocation problems, maximizing system stability and minimizing energy consumption and service latency. Experimental results demonstrate that the performance of the DDPG-LDPG hybrid algorithm surpasses existing task offloading and resource allocation schemes, while exhibiting good scalability.

Currently, the typical framework for cloud gaming involves the use of cloud servers or a combination of cloud servers and edge servers to provide services. At the same time, there is an increasing number of intelligent user devices with computing resources, and the computational capabilities of user devices are also becoming stronger. However, existing frameworks often treat user devices as terminals capable of network connectivity and display, which leads to a waste of computational resources. Therefore, this paper proposes a cloud-edge-end collaborative cloud gaming distributed rendering framework that incorporates user devices into the cloud gaming service network. The framework not only inherits the advantages of traditional frameworks, but also takes advantage of the computing power of user devices, enhances the usability of the framework, makes the utilization of computing resources higher, and makes the framework's work more flexible. Additionally, this paper designs a server operational cost optimization algorithm based on machine learning and metaheuristic algorithms. We compare and evaluate the proposed framework and algorithm with the mainstream framework and algorithm, and the experimental results prove its effectiveness

Vehicle Edge Computing (VEC) is a critical technology that can achieve low latency and energy consumption for Telematics. However, with the high-speed mobility of new energy-electric vehicles and their cross-regional nature, performing high-quality service of vehicle tasks on the VEC model is still challenging. Considering the electric vehicle range problem, this paper plans to jointly optimize task offloading, task result forwarding and computational resource allocation (OOFR) within the maximum tolerable delay of vehicle tasks to minimize vehicle tasks' delay and energy consumption. The non-orthogonal multiple access (NOMA) technology is used with roadside units (RSUs) to achieve multiplexing of limited spectrum resources. This allows multiple vehicle users to perform task transmission simultaneously, thus reducing vehicle task transmission delay. In addition, we propose a cooperative game approach based on NOMA for task grouping to reduce the signal interference of vehicle task transmission. Finally, a deep reinforcement learning method is proposed for task offloading decision selection. A simulation platform is built to compare with MEC, COMO and MADDPG methods, combined with simulation results, showing that the superiority of our proposed scheme is verified.

In multi-party interactive video live streaming, multiple streams are simultaneously transmitted to the audience. However, in this scenario, the QoE problem caused by factors such as device heterogeneity has received little attention. Firstly, we propose a QoE model for multi-party live streaming applications. In order to improve transmission quality, we propose a live streaming framework, and the rate control of the entire framework is divided into two modules: encoding control and transmission control, which are respectively solved. At the coding level, we have achieved better bit rate allocation by clustering user information. The transmission control layer achieves full utilization of network throughput through buffer feedback adjustment. The results of extensive simulation experiments show that our algorithm is superior to fixed rate algorithms, with an average QoE improvement of 2-4 times. In addition, the playback speed control strategy reduces the delay difference between different streams to an imperceptible level.

With the rapid development and popularization of electric vehicles (EVs), the increased demand for charging stations (CSs) has made configuring charging facilities and optimizing charging station recommendations crucial. Consequently, personalized recommendation services for charging stations have emerged as a crucial strategy to mitigate the "mileage anxiety" experienced by drivers. Previous work has typically focused on optimizing CSs’ resource allocation or drivers’ costs, with little consideration of the complexity of the charging behavior of individuals, such as the distribution of Point Of Interests (POIs) surrounding the CS, the similarity of driver selection in different temporal scales, and the tradeoff between different users’ inherent preferences and external cost (e.g., travel times and electric prices), which may result in lower satisfaction, decreased accuracy of recommendations, and sub-optimal decision-making. To address this problem, we propose a charging station recommendation framework based on Actor-Critic Deep Reinforcement Learning, called DST, to assist electric vehicle drivers in finding the proper spots for charging. In the DST, both the actor and critic networks are implemented by Deep Neural Networks (DNNs). The Actor networks, utilizing a Semantic-based Bidirectional Long Short-Term Memory Network and a Stacked Convolutional Neural Network with Multiple Channels, extract the geographical semantic features of CSs and specific preference patterns from different temporal scales as hidden state representations. The critic networks employ a Twin-Critic architecture for joint optimization of inherent preference rewards and external environmental rewards. Extensive experiments on two real-world datasets demonstrate that the DST achieves the best comprehensive performance compared with seven baseline approaches.

In recent years, the advancement of 4G/5G network technologies and smart devices has led to an increasing demand for smooth, massively multiplayer online games on mobile terminals. These games necessitate high performance and heavy workloads, often consuming substantial amounts of computing and storage resources while imposing strict latency requirements. However, due to the limited resources of end devices, such tasks cannot be efficiently and independently executed. The traditional solution typically involves processing gaming tasks at centralized cloud servers. However, this approach introduces issues such as bandwidth pressure, high latency, load imbalance, and elevated costs. Recently, mobile edge computing (MEC) has gained popularity, and its low-latency capabilities can be integrated with cloud gaming to enhance the gaming performance experience. In this paper, we explore the offloading and placement of rendering services in a scenario that combines MEC with cloud gaming. We propose a model-free algorithm based on deep reinforcement learning to learn the optimal task offloading and placement policy, which optimizes a combination of four metrics: latency, cost, bandwidth, and load balancing. Additionally, the algorithm predicts future bandwidth using LSTM, significantly improving the player's gaming experience and fairness. Simulation results demonstrate that our proposed task placement strategy outperforms state-of-the-art methods applied to similar problems.

With the advancement of device software and hardware performance, and the evolution of game engines, an increasing number of emerging high-quality games are captivating game players from all around the world who speak different languages. However, due to the vast fragmentation of the device and platform market, a well-tested game may still experience text glitches when installed on a new device with an unseen screen resolution and system version, which can significantly impact the user experience. In our testing pipeline, current testing techniques for identifying multilingual text glitches are laborious and inefficient. In this paper, we present AG3, which offers intelligent game traversal, precise visual text glitch detection, and integrated quality report generation capabilities. Our empirical evaluation and internal industrial deployment demonstrate that AG3 can detect various real-world multilingual text glitches with minimal human involvement.

MOTIVATION: Gut dysbiosis is closely associated with obesity and related metabolic diseases including type 2 diabetes (T2D) and nonalcoholic fatty liver disease (NAFLD). The gut microbial features and biomarkers have been increasingly investigated in many studies, which require further validation due to the limited sample size and various confounding factors that may affect microbial compositions in a single study. So far, it lacks a comprehensive bioinformatics pipeline providing automated statistical analysis and integrating multiple independent studies for cross-validation simultaneously. RESULTS: OBMeta aims to streamline the standard metagenomics data analysis from diversity analysis, comparative analysis, and functional analysis to co-abundance network analysis. In addition, a curated database has been established with a total of 90 public research projects, covering three different phenotypes (Obesity, T2D, and NAFLD) and more than five different intervention strategies (exercise, diet, probiotics, medication, and surgery). With OBMeta, users can not only analyze their research projects but also search and match public datasets for cross-validation. Moreover, OBMeta provides cross-phenotype and cross-intervention-based advanced validation that maximally supports preliminary findings from an individual study. To summarize, OBMeta is a comprehensive web server to analyze and validate gut microbial features and biomarkers for obesity-associated metabolic diseases. AVAILABILITY AND IMPLEMENTATION: OBMeta is freely available at: http://obmeta.met-bioinformatics.cn/.

Massively multiplayer online games (MMOGs) are a new type of large-scale interactive applications providing a seamless virtual world for millions of gamers from all over the world. Traditionally, MMOGs are implemented using a client-server architecture. With the maturity of cloud computing, many MMOG providers such as Blizzard Entertainment and Jagex Limited have begun to utilize virtualized machines to serve their consumers due to the elasticity, adaptability, and cost advantages of clouds. It is a major challenge for MMOG providers how to dynamically allocate resources in an on-demand pattern in order to reduce the energy cost associated with operating a MMOG cloud while providing good-enough quality of experience (QoE) for MMOG gamers. In this paper, we propose a dynamic resource allocation scheme for MMOGs in heterogeneous cloud computing systems, which makes use of both dynamic virtual machine (VM) consolidation among physical machines (PMs) and long short-term memory based VM resizing at physical machine level to achieve the energy efficiency and desired QoE requirements. Our scheme considers multiple types of resources, characteristic of AFK (away from keyboard) gamers, heterogeneous PMs and VMs, strict QoE requirements and overheads incurred due to migrating VMs. Furthermore, a novel hybrid algorithm based on differential evolution and modified first-fit heuristic is presented for dynamic consolidation of VMs in heterogeneous cloud data centers. The experiment results show that, compared to the traditional over-provisioning policy, our resource allocation scheme can achieve up to 44.8% energy savings while ensuring the QoE requirements of MMOG gamers under rapidly changing workloads.

Prompted by the remarkable progress in wireless communications technology and the explosive growth in the number of mobile devices (MDs), there is an increasing need for performing computation-intensive tasks in mobile edge computing. However, limited to the ability of MDs, MDs are difficult to meet the computational demands of the tasks. So how to offload and process computationally intensive tasks is a key problem in the field of mobile edge computing (MEC). In this paper, we propose a load balancing aware task offloading method in mobile edge computing environment, which aims to maximize computing speed and minimize server selection time. Firstly, we model the wireless channel gain and offloading strategy as Markov Decision Process. Secondly, we propose an offloading strategy generation algorithm based on deep reinforcement learning to generate binary offloading strategy. In addition, due to the limited load capacity of edge servers(ESs), we consider how to balance the load of global ESs and forward tasks to other ESs. Therefore, we propose a Particle Swarm Optimization(PSO) algorithm based on server’s load to achieve the optimal edge server selection goal. In order to evaluate the effectiveness of the proposed method, we use real data sets to do simulation experiments. The experimental results show that the proposed method can achieve lower latency in comparison with the state-of-the-art approaches applied to similar problems.

With the rapid development of the mobile Internet and the Internet of Things, there is an increasing need to execute compute-intensive tasks on mobile devices. But, mobile devices have limited resources, which makes it difficult to provide fast response times. In order to overcome such difficulties, a new computing paradigm called mobile edge computing (MEC) has been proposed to extend cloud-computing capabilities to the edge of the network. In MEC systems, mobile devices can offload compute-intensive tasks to resource-rich edge servers for execution, which effectively reduces the task processing latency. Nowadays, there are many researches on independent task offloading in MEC. However, many applications are composed of dependent tasks in real-world scenarios. Therefore, dependent task offloading has become a hot topic in MEC. In this paper, we study multiple workflows offloading in MEC. A model-free algorithm based on deep reinforcement learning is proposed to learn the optimal multiple workflows offloading strategy so as to minimize the number of workflows whose deadlines are not satisfied. Simulation results show that the proposed multiple workflows offloading strategy performs better than the state-ofthe-art approaches applied to similar problems.

Prompted by the remarkable progress in both Internet of Things and 5G technology, mobile edge computing (MEC) system has been attracting more and more attention from industry to academia. In the MEC system, edge nodes (ENs) deployed at base stations offer computing resources to nearby resource-hungry mobile devices (MDs) to provide high-quality services for edge users. Most researches either study task offloading or resource allocation in MEC system, and few researches consider optimizing both at the same time. In this paper, we propose a MEC system to minimize MDs’ expected long-term delay and energy costs by offloading tasks to the nearby ENs and allocating resources to the tasks. In view of the difference between the two decision problems, we develop an innovative solution, which includes two effective deep reinforcement learning (DRL) algorithms that solve the two problems respectively. We improve the actor-critic (A2C) algorithm to solve the offloading problem effectively, and the resource allocation problem is solved by the deep deterministic policy gradient (DDPG) algorithm. We train the two algorithms alternately. Through this well-designed solution, the problem of task offloading and resource allocation can be solved well. Experimental results show that the performance of A2C-DDPG is better than existing task offloading and resource allocation schemes.

As a popular open-source framework for big data processing, Hadoop Yarn has been widely used by large internet and e-commerce companies such as Amazon, Alibaba, and Facebook. One of the main challenges faced by Hadoop service providers is to optimize the scheduling of MapReduce jobs in order to provide desired performance levels to users and improve cluster resource utilization. In this paper, we propose a deadline-aware preemptive job scheduling strategy, named DAPS, to minimize the job deadline misses. In the proposed DAPS, the job scheduling problem is formulated as an online optimization problem, and a preemptive resource allocation algorithm is developed to search for a good job scheduling policy. We implement the proposed DAPS in Hadoop Yarn clusters and evaluate the effectiveness of the proposed DAPS through several experiments. Our proposed approach has an average job completion rate of 92.33%, which is better than capacity scheduler (62.94%), Earliest Deadline First scheduler (78.76%), and PDSonQueue (87.58%). The experimental results show that the performance of the proposed DAPS is superior to the state-of-the-art strategies applied to similar problems.

Prompted by the remarkable progress in both cloud computing and GPU virtualization, cloud gaming has been attracting more and more attention in the gaming industry. With the cloud gaming model, players do not need to download or install the game on local devices, and constantly upgrade their devices. Despite these advantages, cloud gaming faces several challenges for its success, including long response delay, poor game fairness, and high operational cost. To this end, this article proposes an edge computing-assisted multiplayer cloud gaming system named ECACG to improve multiplayer cloud gaming experiences and operating costs by offloading the game rendering task to the nearby edge server. Based on the ECACG, two decision processes are completed. One is player request scheduling and the other is rendering server allocation. The decision problem is formulated into a constrained multiobjective optimization model. A novel hybrid algorithm based on deep reinforcement learning and heuristic strategy is developed to solve the optimization problem. The effectiveness of the proposed ECACG is evaluated by simulation experiments based on the real-world parameters. The simulation results show that compared with the existing schemes, the proposed ECACG can achieve lower rental costs and better fairness, while providing the good-enough response delay for players.

Prompted by the remarkable progress in cloud computing, more and more game service providers are starting to deploy their gaming applications on infrastructure as a service cloud. To ensure quality of service required by gamers, game service providers need to maintain a large number of cloud servers running game instances requested by players. Therefore, game service providers need not only pay attention to the costs of operating games, but also consider the quality of game services. This paper proposes a quality-of-service aware cost optimization scheme for online gaming services deployed on IaaS clouds which combines dynamic virtual machine provisioning and Long Short-Term Memory (LSTM) based gamer request dispatching to reduce the operating costs while ensuring quality of service. The effectiveness of our proposed scheme is assessed by simulation experiments based on the real-world data sets. The results show that, compared to the state-of-the-art approaches applied to similar problems, our scheme can achieve up to 35% cost savings while providing the just-good-enough quality of service to gamers under rapidly changing workloads.

Hadoop YARN has become a dominant framework for big data analysis and processing. However, the inbuilt scheduler in Hadoop YARN framework is not designed for energy efficiency. To overcome this problem, this paper presents an energy-efficient extension on the existing Hadoop YARN framework. In addition, we formulate the MapReduce scheduling in the heterogeneous Hadoop YARN cluster as an energy consumption optimization problem, and propose a heuristic algorithm to solve this optimization problem. The proposed algorithm takes advantage of both load balancing and dynamic voltage/frequency scaling to improve performance and energy efficiency of the Hadoop YARN cluster. We evaluate the effectiveness of our method by carrying out extensive experiments on a real Hadoop YARN cluster consisting of five servers. The results show that our method can provide significant energy savings and achieve better performance compared with three alternative methods applied to similar problems.

Background subtraction can be treated as the binary classification problem of highlighting the foreground region in a video whilst masking the background region, and has been broadly applied in various vision tasks such as video surveillance and traffic monitoring. However, it still remains a challenging task due to complex scenes and for lack of the prior knowledge about the temporal information. In this paper, we propose a novel background subtraction model based on 3D convolutional neural networks (3D CNNs) which combines temporal and spatial information to effectively separate the foreground from all the sequences in an end-to-end manner. Different from conventional models, we view background subtraction as three-class classification problem, i.e., the foreground, the background and the boundary. This design can obtain more reasonable results than existing baseline models. Experiments on the Change Detection 2012 dataset verify the potential of our model in both quantity and quality.

Recommender systems provide users with ranked lists of items based on individual's preferences and constraints. Two types of models are commonly used to generate ranking results: long-term models and session-based models. While long-term models represent the interactions between users and items that are supposed to change slowly across time, session-based models encode the information of users' interests and changing dynamics of items' attributes in short terms. In this paper, we propose a PLASTIC model, Prioritizing Long And Short-Term Information in top-n reCommendation using adversarial training. In the adversarial process, we train a generator as an agent of reinforcement learning which recommends the next item to a user sequentially. We also train a discriminator which attempts to distinguish the generated list of items from the real list recorded. Extensive experiments show that our model exhibits significantly better performances on two widely used real-world datasets.

With the rising demand on internet online services, more and more individuals and organizations migrate their data and services from local to geographically distributed internet data centers for reliability, management, and cost benefits. From internet data center operators' perspective, profit is one of the most significant factors which are to be taken into account and it is mainly determined by the quality of service and the electricity cost. In this work, we jointly take the diversity of time-varying electricity prices, the quality of service and the power usage effectiveness of data centers into consideration and propose a profit optimization framework that includes three important decisions: requests dispatching, the number of active servers and frequency adjustment. An efficient heuristic algorithm is then developed to provide internet data center operators with the advice on these decisions to achieve profit maximization for geographically distributed data centers. Through extensive trace-driven simulations, we demonstrate the effectiveness and superiority of our solution.

In this letter, we propose an adaptive part-level model knowledge transfer approach for gender classification of facial images based on Fisher vector (FV). Specifically, we first decompose the whole face image into several parts and compute the dense FVs on each face part. An adaptive transfer learning model is then proposed to reduce the discrepancies between the training data and the testing data for enhancing classification performance. Compared to the existing gender classification methods, the proposed approach is more adaptive to the testing data, which is quite beneficial to the performance improvement. Extensive experiments on several public domain face data sets clearly demonstrate the effectiveness of the proposed approach.

Web service composition problem was considered as a planning problem by previous research. However, many factors constantly affect the QoS and results of invocation of web services, thus the environment of web services is dynamic. As result, web service composition problem should be considered as an uncertain planning problem. This paper uses Markov property to deal with the uncertain planning problem for service composition. According to the uncertainty model, we propose a reinforcement learning method to compose web services. Without knowing the transition function and reward function, our uncertain planning method uses an estimated value function to approach a real function and is able to obtain a composite service. The results of experiments show that our method can effectively reduce computing time of the service composition.

How we manage Web services depends on how we understand their variable parts and invariable parts. Studying them separately could make Web service research much easier and make our software architecture much more loose-coupled. We summarize two variable parts that affect Web service compositions: uncertain invocation results and uncertain quality of services. These uncertain factors affect success rate of service composition. Previous studies model the Web service problem as a planning problem, while this problem is considered as an uncertain planning problem in this paper. Specifically, we use Partially Observable Markov Decision Process to deal with the uncertain planning problem for service composition. According to the uncertain model, we propose a reinforcement learning method, which is an uncertainty planning method, to compose web services. The proposed method does not need to know complete information of services, instead it uses historical data and estimates the successful possibilities that services are composed together with respect to service outcomes and QoS. Simulation experiments verify the validity of the algorithm, and the results also show that our method improves the success rate of the service composition.

Local binary descriptors are attracting increasingly attention due to their great advantages in computational speed, which are able to achieve real-time performance in numerous image/vision applications. Various methods have been proposed to learn data-dependent binary descriptors. However, most existing binary descriptors aim overly at computational simplicity at the expense of significant information loss which causes ambiguity in similarity measure using Hamming distance. In this paper, by considering multiple features might share complementary information, we present a novel local binary descriptor, referred as ring-based multi-grouped descriptor (RMGD), to successfully bridge the performance gap between current binary and floated-point descriptors. Our contributions are twofold. First, we introduce a new pooling configuration based on spatial ring-region sampling, allowing for involving binary tests on the full set of pairwise regions with different shapes, scales, and distances. This leads to a more meaningful description than the existing methods which normally apply a limited set of pooling configurations. Then, an extended Adaboost is proposed for an efficient bit selection by emphasizing high variance and low correlation, achieving a highly compact representation. Second, the RMGD is computed from multiple image properties where binary strings are extracted. We cast multi-grouped features integration as rankSVM or sparse support vector machine learning problem, so that different features can compensate strongly for each other, which is the key to discriminativeness and robustness. The performance of the RMGD was evaluated on a number of publicly available benchmarks, where the RMGD outperforms the state-of-the-art binary descriptors significantly.