We consider the fault-tolerant facility location problem with penalties. In this problem, we replicate each client v_j times. Some of these replicas are connected to opened facility, while others remain unconnected and are penalized. By employing a natural LP-rounding technique, we provide a 2-approximation algorithm for this problem. In this algorithm, we balance the connection cost and penalty cost by setting thresholds.
In an adversarial environment, agent clusters encounter tasks with uncertain arrival times and structures. As the current agent clusters present the characteristics of multiple networks, the previous methods applied to single-network collaboration cannot adjust the strategy in time according to the current environment, which greatly reduces the collaborative effect. In this paper, we propose a multi-agent collaboration method for time-sensitive tasks in multiple networked confrontation environments. By using hierarchical reinforcement learning, the collaboration problem is divided into the network-agent layer and task-subtask layer. A lot of experiments and theoretical analysis show that the method proposed in this chapter can better accomplish the assignment between the task and the agent, and improve the task income.
In this paper, a convolutional neural network parameter training method based on Hausdorff difference is proposed to solve the problems of gradient vanishing and local optimum in the momentum algorithm. A momentum algorithm is proposed based on Hausdorff difference, introduced similar to the conception of Hausdorff derivative. Furthermore, the proposed algorithm is improved by using two new methods for adaptive nonlinear adjustment of the order. We analyse the influence of the order on the training results of the network parameters and verify the effectiveness of the proposed methods in improving the recognition accuracy and convergence speed of convolutional neural networks by the fashion-MNIST dataset and the CIFAR-10 dataset.
Heterogeneous computing with accelerators has emerged as an effective approach to high-performance computing. Directive-based programming models such as OpenMP and OpenACC simplify parallel programming for GPU accelerators by enabling compilers to translate directive-annotated code into GPU-optimized code automatically. This paper presents accparser, a standalone and unified OpenACC parser built using ANTLR 4. Designed for both C/C++ and Fortran, accparser provides a complete grammar for OpenACC 3.3, making it a valuable tool for compiler developers who aim to implement OpenACC program transformations and code generation. It supports the syntax and semantic verification of OpenACC constructs and helps to interpret the OpenACC standard. Additionally, it can be leveraged as a tool to assist in creating compiler passes that convert OpenACC programs into OpenMP programs, enabling full utilization of OpenMP’s compiler optimizations and runtime support. The source code for accparser is available under the 2-Clause BSD License at https://github.com/passlab/accparser .
With the rise of precision agriculture, efficiently identifying fruit tree locations and optimizing operational paths have become increasingly challenging. To address these issues, we developed an integrated path automation system that leverages the collaboration between unmanned aerial vehicles (UAVs) and ground robots. The system captures multi-angle images of fruit trees using UAVs, which are then processed by the Multi-MiniGPT-v2 model, which is an intelligent multi-agent system incorporating MiniGPT-v2 and MetaGPT, to determine tree coordinates and assess fruit ripeness from both top-down and side views. This coordinate data is input into the Adaptive Reflective Evolution (AREvo) model, which utilizes the large model and Reflective Evolution algorithm for automatic path planning and generates optimized robot paths. Experimental results indicate that this method significantly enhances the operational efficiency of orchard robots, improves the accuracy of fruit tree status detection, and reduces costs. This technology alleviates manual labor, boosts agricultural productivity, and optimizes resource utilization, providing strong support for advancing precision agriculture.
The intermittent and fluctuating solar irradiance makes photovoltaic (PV) power generation unstable, which brings great challenges to the power grid system. Existing deep learning-based PV power generation prediction models usually employ Long Short-Term Memory (LSTM) or Convolutional Neural Networks (CNNs) to predict the future PV power generation. However, the recurrent structure-based methods usually failed to model the long-term dependence due to the accumulated error caused by backpropagation through time. To overcome these issues, an improved Transformer-based prediction model, which is called GLformer, is proposed in this work. Specifically, the proposed GLformer decomposes an input series into global pattern subsequences and local pattern subsequences, where the global pattern subsequences represent the trends and seasonality and the local pattern subsequences are used to describe the fluctuations. The global and local features are extracted and fused using a Transformer-based global-local feature extraction module, which enhances the model’s capability of modeling long-term dependence, and obtaining complete time features for prediction. Extensive experimental results show that the proposed method is superior to most SOTA forecasting models.
In modern warfare, decoy targets are often set up to interfere with enemy reconnaissance and operations. To avoid being disrupted by decoys during combat, it is necessary to consider the implementation of robust combat strategies. In this paper, we consider the problem of collaborative target assignment for unmanned aerial vehicle (UAV) swarms with decoy targets. We model it as a deletion-robust submodular maximization problem, and develop a two-stage algorithm that attains a 1/(4+1/β ) -approximation guarantee, where β≤ 1 is the approximation ratio for the unconstrained submodular function maximization algorithm.
In order to improve the economic benefits of photovoltaic power generation systems, accurate prediction of photovoltaic power generation revenue has become particularly important. However, traditional prediction algorithms such as CNN and LSTM suffer from overfitting issues, while the latest emerging KAN neural network also has problems with parameter tuning difficulties and sensitivity to data quality. Based on these issues, this paper proposes a CNN_LSTM_KAN photovoltaic power generation revenue prediction algorithm based on genetic algorithm. The experimental results show that the algorithm outperforms traditional methods in both prediction accuracy and robustness, providing an effective solution for predicting the revenue of photovoltaic power generation systems.
Modern supercomputing nodes are typically highly heterogeneous and exhibit Non-Uniform Memory Access (NUMA) characteristics. Automatically mapping tasks on such NUMA architectures while maintaining high operational efficiency poses significant challenges. In this paper, we propose a novel automatic mapping algorithm called NAAM (NUMA-Aware Auto Mapping), which accounts for both intra-node and inter-node heterogeneity. NAAM dynamically adjusts mapping strategies for various tasks across heterogeneous NUMA systems by performing a dynamic search to identify optimal mappings for typical heterogeneous hardware configurations. Finally, we conduct extensive experiments to validate the effectiveness of our method, with results showing that NAAM outperforms baseline approaches by up to 8
Existing research on multi-component radar signal recognition widely adopts recognition frameworks based on time-frequency transformation (TFT) and convolutional neural networks (CNN). To address the issue of the vulnerability of traditional TFT-generated time-frequency representations (TFR) to noise under low signal-to-noise ratio (SNR) conditions, we propose a new TFT scheme, called Multi-Scale TFT Network (MTFT-Net). Specifically, MTFT-Net learns diverse and comprehensive basis functions to obtain various TF features of time-domain multi-component radar signals. It then uses subsequent aggregation modules to concentrate and reconstruct the energy of the TF features, ultimately outputting the TFR. Experimental results show that MTFT-Net generates better TFRs with superior noise resistance under low SNR conditions compared to traditional TFT methods. Moreover, MTFT-Net can mimic the styles of various traditional TFTs. Finally, we compare its performance with the advanced TFA-Net to demonstrate the effectiveness of the proposed method.
Existing open-source text language identification tools are not ideal for recognizing short texts in Chinese, Japanese, and Korean, and some of them cannot distinguish between Simplified and Traditional Chinese. To improve identification efficiency, this paper proposes a two-stage language identification algorithm for Chinese, Japanese, and Korean. The first stage is based on a character script identification algorithm, according to the Unicode encoding range of Chinese, Japanese, and Korean, and determines the language. If the identification result is Chinese, the classification algorithm developed using the bidirectional long short-term memory (BiLSTM)-Attention model is used in the second stage to determine whether the language is Simplified or Traditional Chinese. The experimental results show that the proposed method improves the identification efficiency of Chinese, Japanese, and Korean short texts compared to existing open-source language identification tools.
With the continuous expansion of comprehensive road transport networks in major cities, accurate traffic flow prediction—an essential component of intelligent urban transport systems has become increasingly challenging. Spatio-Temporal Graph Neural Networks (STGNNs), which combine time-series models like the Transformer with spatial relationship learning models such as Graph Neural Networks (GNNs), have shown significant promise in addressing this challenge. However, they still face issues like insufficient exploitation of long-term dependencies, high structural and computational complexity, and underutilization of periodic patterns in traffic data. To address these limitations, we propose LL-MLP, a spatio-temporal model based on an encoder-decoder architecture. Our approach extends the look-back window to better capture long-term dependencies, introduces learnable parameter matrices to enhance spatial information extraction, and incorporates periodic information as a dynamic covariate to further leverage temporal patterns. Extensive experiments on four real-world traffic datasets demonstrate that LL-MLP achieves performance on par with or exceeding that of existing model architectures.
High-performance computing (HPC) is crucial for obtaining solutions to complex scientific and engineering problems described by partial differential equations (PDEs). However, most of these simulations are very sensitive to one major cause of performance bottlenecks: the communication overhead between nodes. While the Message Passing Interface (MPI) is a standard library for communication on HPC systems, the best-suiting MPI communication function is very hard to decide due to numerous variations of algorithms and their dependency on multiple factors. In this paper, we present a machine learning-enabled approach to predict optimal MPI communication functions for large-scale PDE simulations. Our model encompasses many diversified features, like mesh size, number of variables, compiler type, number of cores, and variable types, in its construction to take a more holistic approach toward the optimization of communications. In fact, the model predicts the most suitable MPI function for a given simulation environment by taking into account the aforementioned parameters. Experiment results show clearly that incorporating these features into the prediction model yields the best communication strategy which is highly accurate, reliable, ensuring better scalability and performance. To the best of our knowledge, this is the first thorough study addressing the use of machine learning techniques to optimize MPI communication for PDE simulations; therefore, it is an important contribution for leveraging the predictive capability of machine learning models for the co-design of large-scale numerical simulations.
In recent years, the International Air Transport Association (IATA) has introduced the New Distribution Capability (NDC), which has enabled airlines to implement continuous dynamic pricing. However, there is currently a lack of research on continuous dynamic pricing for multiple products. To fill this gap, this paper utilizes the Soft Actor-Critic (SAC) algorithm from reinforcement learning to study the continuous dynamic pricing problem of multi-class airline tickets. Additionally, the paper proposes a reward shaping method called Balanced Inventory Soft Actor-Critic (BISAC) to prevent certain ticket classes from being sold too quickly due to pricing errors during the sales process, outpacing other classes. Experimental results indicate that the performance of the BISAC algorithm significantly outperforms the original SAC algorithm and existing algorithms used for dynamic pricing of multiple products.
Pellet performance plays a crucial role in the subsequent smelting process in the furnace, so an im-proved multilayer perceptron (MLP) pellet metallurgical performance prediction improvement model is proposed. The improved model mainly consists of a multilayer perceptron and a gradient boost-ing framework (XGBoost), which ensures that the integrated model has the advantages of the XGBoost iterative training model that effectively improves the prediction accuracy, and also ensures the flexibility and extensiveness of the MLP model in dealing with nonlinear problems and largescale data sets. The model is trained with a large amount of raw pellet raw material ratio and pellet ore metallurgical performance data to quantitatively characterise the relationship between pellet raw mate-rial ratio information and pellet ore metallurgical performance. Combined with the simulation results, the prediction results of RDI, ΔT, RI and RSI are better than some existing prediction algorithms, and the prediction of the metallurgical properties of pellet ores is achieved.
This study proposes a novel decentralized federated learning algorithm, DKFL, which combines knowledge distillation technology and has been successfully applied in the fields of image classification and demand forecasting in the industrial chain. This algorithm effectively improves the performance of the model through knowledge distillation, while retaining the privacy protection advantages of decentralized federated learning. In image classification tasks, our algorithm not only performs well on the CIFAR10 dataset, but also demonstrates significant improvement in classification accuracy on the automotive parts dataset in the automotive industry chain. In demand forecasting tasks, this algorithm was applied to the supply chain dataset of the industrial chain, achieving more accurate forecasting results. The experimental results show that the proposed method can effectively improve the generalization ability and prediction accuracy of the model while ensuring data privacy, providing new ideas and technical support for the promotion of federated learning in the application of the industry chain.
The traditional single industrial chain structure is gradually evolving into the multiplex industrial chain networks. The design problem of multiplex industrial chain networks with multiple supply cycles (MICND-MSC) is studied in this paper. The impact of cross-chain supply of multiple raw materials by enterprises during the production process is considered in the problem. The mixed-integer linear programming (MILP) model of MICND-MSC is constructed and an artificial bee colony algorithm (ABC) with reinforcement learning mechanism (LDABC) is proposed to address the MICND-MSC. Heuristic methods named HMC, HMR, and PRH are designed to construct potential initial candidates for the population. The neighborhood structures for different production stages are employed in the LDABC to explore the solution space during the evolution processes. The reinforcement learning mechanism is utilized to learn empirical knowledge of neighborhood structures to guide the search process. The experimental results show that the LDABC is a potential algorithm to address the MICND-MSC.
Location-based service (LBS) applications are increasingly popular for travelling. The public transit scenario is very common in urban areas, yet there is a lack of effective privacy protection mechanisms to safeguard users’ location privacy when using services from untrusted third-party LBS providers. This paper introduces Transportation Shift-based Routing (TSR), a privacy protection mechanism for public transportation travel scenarios. It allows users to query public transit routes through applications that comply with this mechanism without disclosing any significant details about the location, and fills the gap in privacy protection mechanisms in public transportation scenarios. The fundamental concept is to shift the query’s starting and ending points to nearby stations and repeatedly perform perturbations that satisfy geo-indistinguishability, submitting a collection of perturbed station pairs to the LBS. TSR computes the set of obtained routes to derive an optimal route plan. The best route calculated can maintain service availability and protect user location privacy, with only a minimal difference from the best route obtained by direct query. TSR also allows users to freely choose privacy parameters for different scenarios and strives to ensure service quality, offering high flexibility. We implement the TSR application and conduct experiments with real-world map services. The experimental outcomes indicate that TSR successfully strikes a balance between preserving location privacy and maintaining service usability.
In this paper, we propose WaitIO-Hybrid, a communication library for coupling multiple MPI programs across heterogeneous systems. It supports sockets (WaitIO-Socket) and shared file system communication (WaitIO-File) across systems with different hardware and software configurations. When tested on the Wisteria/BDEC-01 supercomputer, WaitIO-Hybrid achieved 5.2 times better communication performance than WaitIO-Socket. NICAM/ADA applications ran 1.4 times faster on a combination of Arm CPUs on Odyssey and NVIDIA GPUs on Aquarius. WaitIO-File also performed well on the “Flow” system at Nagoya University.
The rapid growth of cloud computing has brought new challenges in Parallel Batch Machine Scheduling (PBMS), particularly when incorporating malleability and rejection constraints. This has led to the Parallel Batch Machine Scheduling with Malleability and Rejection problem (PBMSMR), which involves malleable jobs whose widths can be adjusted during execution within specified limits and incorporates job rejection subject to a penalty threshold. Based on an analysis of key properties of batch scheduling with job rejection, we develop an approximation algorithm for PBMSMR by employing a greedy approach that reorders and iteratively refines job sets to minimize the objective while respecting rejection constraints. The algorithm achieves a time complexity of O(n^2 log n) and an approximation ratio of (4-2/Km) , where K and m denote the capacities and the number of the machines, respectively. For jobs with identical release times, we fine-tune the algorithm to achieve an approximation ratio of (3 + 2(K-1)/Km) . Additionally, for PBMS with two machines of non-identical capacities and fixed-width jobs, we achieve a ratio of 17/5 with O(n log n) , improving upon the previous state-of-the-art approximation ratio of 5 with a runtime of O(n^2) .
