Memetic Computing最新文献

Oriented object detection has garnered significant attention for its broad applications in remote sensing image processing. Most oriented detectors perform dense predictions on a set of predefined anchors to generate oriented bounding boxes, where these anchors require classification (cls) and localization (loc) labels for detector training. Recent advancements in label assignment utilize the overall quality score of cls and loc predictions to determine positive and negative samples for each oriented object. However, these methods typically establish the overall quality score by assigning fixed weights to cls and loc quality scores. This approach may not be optimal, as fixed weights fail to dynamically balance cls and loc performance during model optimization, thereby constraining detection efficacy. Motivated by this observation, this paper proposes a Dynamic Weighting Label Assignment (DWLA) algorithm. DWLA dynamically adjusts the weights of individual quality scores based on the current model state to continuously balance cls and loc performance. Additionally, to mitigate the impact of unreliable predictions and achieve more stable training, this paper proposes a level-wise positive sample selection scheme and an object-adaptive scheme for constructing initial candidates of positive samples, respectively. Comprehensive experiments on the DOTA and UCAS-AOD datasets have validated the effectiveness of the proposed DWLA.

The complexity of next-generation wireless communications, especially Beyond 5G and 6G communication systems, will be handled by artificial intelligence-based management paradigms. The joint selection of routes and functional split levels involves critical decisions that network infrastructure providers need to make to support requests from virtual Mobile Network Operators (vMNOs). These decisions comprise the assignment and configuration of physical network resources, which must comply with the specific quality of service restrictions of each vMNO request. Recent work defined a detailed mathematical model for this complex challenge, its formulation as a constrained, discrete optimization problem, and the first algorithmic approaches. It was also found that an evolutionary algorithm delivers higher-quality solutions than an ad-hoc heuristic, and faster running times compared to a well-known commercial solver. This paper introduces a memetic algorithm that exploits the strengths of the former evolutionary method while incorporating several key innovations: a domain-specific recombination operator; a specialized repairing procedure; an enhanced fitness evaluation scheme; and a multiobjective archiving strategy that preserves promising solution trade-offs. We conduct a comprehensive evaluation of the performance and behavior of this proposal, as well as the contribution of each specific design component. The results highlight that our memetic algorithm consistently outperforms previous approaches from the literature, providing better trade-offs in terms of solution quality and the rate at which vMNO requests are successfully fulfilled.

Semantically segmented aerial laser scanning (ALS) pointcloud is crucial for remote sensing applications, offering advantages over aerial images in describing complex topography of vegetation-covered areas due to its ability to penetrate through vegetation. While many ALS pointcloud segmentation methods emphasize the importance of color information for accurate segmentation and colorize the ALS pointcloud with aerial images, they often overlook the fact that some points in vegetation-covered areas are occluded and cannot be observed in aerial images. Consequently, these methods may assign inaccurate colors to these points, resulting in degraded segmentation performance. To address this issue, this paper proposes a Hierarchical Heterogeneous Graph Learning (HHGL) algorithm. HHGL tackles the problem by treating the colors of occluded points (referred to as “color-missing points”) as missing values and compensating for them based on the local and global geometric relationships among color-missing points and color-observed points. Specifically, the proposed algorithm first models the local geometric relationships as a heterogeneous graph, which aggregates the features of adjacent color-observed points to make up for the missing colors. Additionally, the global geometric relationships are represented as a hierarchical structure, refining the aggregated features and capturing long-range dependencies among color-missing points to facilitate segmentation. Experimental results on real-world datasets validate the effectiveness and robustness of the proposed HHGL algorithm.

Molecular representations have essential roles in bio-cheminformatics as they facilitate the growth of machine learning applications in numerous sub-domains of biology and chemistry, especially drug discovery. These representations transform the structural and chemical information of molecules into machine-readable formats that can be efficiently processed by computer programs. In this paper, we present a comprehensive review, providing readers with diverse perspectives on the strengths and weaknesses of well-known molecular representations, along with their respective categories and implementation sources. Moreover, we provide a summary of the applicability of these representations in de novo molecular design, molecular property prediction, and chemical reactions. Besides, representations for macromolecules are discussed with highlighted pros and cons. By addressing these aspects, we aim to offer a valuable resource on the significant role of molecular representations in advancing bio-cheminformatics and its related domains.

Quantum neural network (QNN) is a research orientation that combines quantum computing and machine learning. It has the potential to solve the bottleneck problem of shortage of computing resource in deep learning, and is expected to become the first practical application scheme that demonstrate application level quantum advantages on current Noise Intermediate scale Quantum (NISQ) devices. However, limited by the current scale of NISQ devices, QNNs have fewer quantum bits and quantum circuits cannot be too deep. Currently, there is no clear design strategy for the architecture of QNN. Designing QNN architectures arbitrarily not only has high circuit complexity but also often poor network performance. Similar to classical convolutional neural network, in this paper, a quantum evolution-based optimization algorithm is proposed for design of quantum convolutional neural network (QCNN) architecture. The design of QNN architecture is viewed as a combinatorial optimization problem, and the quantum evolution algorithm is used to adaptively design the QCNN architecture with its global search ability in a large discrete search space. Comprehensive experimental results indicate that the proposed method can effectively reduce the complexity of QCNN circuits, reduce the difficulty of deploying quantum circuits, and further improve the expressibility of QCNN.

Driver profiling is a widely used tool in fleet management and driver-specific insurance because it differentiates drivers based on their driving behaviors, such as aggressive and non-aggressive, which correspond to different levels of driving risk. However, most existing driver profiling methods require all drivers to drive on the same predefined route or type of roads, simply to make sure their driving behaviors are comparable. This premise makes these methods not be able to profile drivers who drive on arbitrary roads, which constitute the real-world scenarios for most drivers. To enable the profiling of drivers using their naturalistic driving data, i.e., driving trajectories recorded while they were driving on arbitrary roads at their own free will, in this paper, we propose a novel method named cLustering rOads And Drivers Successively (LOADS). Specifically, LOADS first categorizes the roads into different types using the extracted characteristics of all drivers driving on the respective roads. It then groups drivers into different clusters to obtain their profile labels (e.g., aggressive or non-aggressive) using the extracted driving characteristics on each road type. We conduct extensive experiments using two real-world driving trajectory datasets comprising thousands of driving trajectories of hundreds of drivers. Statistical analysis results indicate that the driver groups identified by LOADS have significantly different driving styles. To the best of our knowledge, LOADS is the first method that focuses on profiling drivers who drive on arbitrary roads, showing a great potential to enable real-world driver profiling applications.

There may exist a one-to-many mapping between objective and decision spaces in multimodal multi-objective optimization problems (MMOPs), which requires the evolutionary algorithm to locate multiple non-dominated solution sets. In order to enhance the diversity of the population, we develop a multimodal multi-objective differential evolution algorithm based on distributed individuals and lifetime mechanism. First, every individual can be seen as a distributed unit to locate multiple non-dominated solutions. The solutions with the good diversity are generated by adopting virtual population, and the range of virtual population is adjusted by an adaptive adjustment strategy to locate more non-dominated solutions. Second, it is considered that each individual has a limited lifespan inspired by natural phenomenon. As the search area of individuals becoming adaptively smaller, the individuals with good quality are archived and they can reinitialize with a new lifespan for enhancing diversity of the search space. Then the probability selection strategy is applied in the environment selection to balance exploration and exploitation. The test results on 22 multimodal multi-objective benchmark test functions verify the superior performance of the proposed method.

Constrained multi-objective problems are difficult for researchers to solve because they contain infeasible regions. To address this issue, this paper proposes two cooperative constraint handling techniques that use an external archive. First, two constraint handling techniques, i.e., the penalty function and the constrained dominance principle, are embedded in multi-objective optimization algorithms and work cooperatively on two populations to increase population diversity. Then, an external archive is designed to preserve high-quality solutions that strike a good balance between objectives, values, and constraints throughout the evolution process. Finally, comprehensive experiments are conducted to validate the performance of the proposed algorithm, and seven state-of-the-art constrained multi-objective optimization algorithms are used to compare three test suites and ten real-world problems. The experimental results demonstrate that the proposed algorithm can achieve competitive performance in solving various constrained multi-objective problems. Additionally, the results show that cooperative constraint handling techniques are more robust than single constraint handling methods.

Multi-modal multi-objective optimization problems (MMOPs) involve multiple Pareto sets (PSs) in decision space corresponding to the same Pareto front (PF) in objective space. The difficulty lies in locating multiple equivalent PSs while ensuring a well-converged and well-distributed PF. To address this, a neighborhood-assisted reproduction strategy is proposed. Through interactions with non-dominated solutions, the generated offspring could spread out along the PF, while ineractions with neighbors could improve the convergence ability. Importantly, individuals can participate in multiple neighborhoods, reducing the computational burden. Additionally, a neighborhood-assisted environmental selection strategy is prposed to encourage exploration of diverse solution regions, ensuring a balanced distribution of the population and preservation of multiple PSs. Comparative experiments are implemented on the CEC 2019 MMOPs test suite, and the superior performance of the proposed algorithm is demonstrated in comparison to several state-of-the-art approaches.

Since it was first presented, particle swarm optimization (PSO) has experienced numerous improvements as a traditional optimization approach. PSO becomes more complex as a result of the majority of improvement strategies, which use learning model replacement or parameter adjustment to enhance PSO’s performance. Based on linear system theory, this study proposes a simple and scalable framework for restructuring particle swarm optimization (RPSO) and provides a new example of the RPSO algorithm framework, Q-RPSO. The RPSO framework adopts a single position updating formula instead of the original position and velocity updating formulas, which are unrelated to the PSO’s velocity and the current position. The experiments were carried out to compare with the standard PSO and six PSO variants based on CEC 2013 benchmark functions. The experimental results demonstrate that, whether in terms of global exploration capability or convergence accuracy, Q-RPSO outperforms all competitor algorithms.