Memetic Computing最新文献

Molecular property prediction is an important step in the drug discovery pipeline. Numerous computational methods have been developed to predict a wide range of molecular properties. While recent approaches have shown promising results, no single architecture can comprehensively address all tasks, making this area persistently challenging and requiring substantial time and effort. Beyond traditional machine learning and deep learning architectures for regular data, several deep learning architectures have been designed for graph-structured data to overcome the limitations of conventional methods. Utilizing graph-structured data in quantitative structure–activity relationship (QSAR) modeling allows models to effectively extract unique features, especially where connectivity information is crucial. In our study, we developed residual graph attention networks (ResGAT), a deep learning architecture for molecular graph-structured data. This architecture is a combination of graph attention networks and shortcut connections to address both regression and classification problems. It is also customizable to adapt to various dataset sizes, enhancing the learning process based on molecular patterns. When tested multiple times with both random and scaffold sampling strategies on nine benchmark molecular datasets, QSAR models developed using ResGAT demonstrated stability and competitive performance compared to state-of-the-art methods.

Over the past decade, the surge in global container port throughput has heightened the demand for terminal efficiency, with the container yard operations being central to the overall port performance. However, the unpredictable arrival of external trucks poses significant challenges for yard cranes which must simultaneously schedule operations for both internal and external tasks. Traditional yard crane scheduling methods often rely on outdated assumptions that fail to account for the dynamic impact of external tasks. In response, container terminals increasingly model the yard crane scheduling as an online problem. A notable advancement in online scheduling is the online rollout method, which evaluates the decisions based on the potential outcomes of their future rollout schedules rather than immediate priorities. Although this method outperforms the previous approach, it faces two main issues: the rollout simulation is time consuming, and decisions based solely on objective value of rollout schedules may not align with long-term scheduling objectives. To overcome these limitations, we have developed a two-stage adaptive rollout decision model. In the first stage, less desirable tasks are dynamically filtered out to reduce the number of rollout simulations required, while the second stage employs a genetic programming evolved evaluation function to infuse more refined forward-looking insights into the scheduling process. This approach has proven to significantly enhance yard scheduling efficiency and performance, as confirmed by experimental validation. Given the dynamic nature of yard crane operations, we believe this method could be beneficially applied to other dynamic scheduling contexts.

Evolutionary Algorithms (EAs), including Evolutionary Strategies (ES) and Genetic Algorithms (GAs), have been widely accepted as competitive alternatives to Policy Gradient techniques for Deep Reinforcement Learning (DRL). However, they remain eclipsed by cutting-edge DRL algorithms in terms of time efficiency, sample complexity, and learning effectiveness. In this paper, aiming at advancing evolutionary DRL research, we develop an evolutionary policy optimization algorithm with three key technical improvements. First, we design an efficient layer-wise strategy for training DNNs through Covariance Matrix Adaptation Evolutionary Strategies (CMA-ES) in a highly scalable manner. Second, we establish a surrogate model based on proximal performance lower bound for fitness evaluations with low sample complexity. Third, we embed a gradient-based local search technique within the evolutionary policy optimization process to further improve the learning effectiveness. The three technical innovations jointly forge a new EA for DRL method named Proximal Evolutionary Strategies (PES). Our experiments on ten continuous control problems show that PES with layer-wise training can be more computationally efficient than CMA-ES; our surrogate model can remarkably reduce the sample complexity of PES in comparison to latest EAs for DRL including CMA-ES, OpenAI-ES, and Uber-GA; PES with gradient-based local search can significantly outperform several promising DRL algorithms including TRPO, AKCTR, PPO, OpenAI-ES, and Uber-GA.

Accurate tracing of crude oil origins is essential for thwarting deceptive trade practices, including origin falsification to evade taxes, thereby preventing economic losses and security threats for importing nations. Traditional crude oil origin determination methods require complex sample preparation, expensive instrumentation, and stable testing environments, rendering them impractical for real-time analysis at locations such as ports. This paper introduces a novel approach utilizing near-infrared spectroscopy (NIRS) combined with deep learning algorithms to expedite and enhance the precision of crude oil source identification. To effectively eliminate outliers, an improved Mahalanobis distance is introduced, incorporating regularization principles and global-local concepts. This approach addresses the challenges of inverting covariance matrices in high-dimensional spectra and excludes samples with localized aberrations. Furthermore, the integration of multi-receptive fields perception, Transformer-based global information interaction, and the scSE attention mechanism has led to the development of an MG-Unet model, designed to resolve spectral peak overlap issues and capture long-range feature dependencies. The proposed method achieves state-of-the-art accuracy of 96.92%, demonstrating significant potential for reliable crude oil source tracing.

Self-supervised learning, particularly through contrastive learning, has shown significant promise in vision tasks. Although effective, contrastive learning faces the issue of false negatives, particularly under domain shifts in domain adaptation scenarios. The Bootstrap Your Own Latent approach, with its asymmetric structure and avoidance of unnecessary negative samples, offers a foundation to address this issue, which remains underexplored in domain adaptation. We introduce an asymmetrically structured network, the Bootstrap Contrastive Domain Adaptation (BCDA), that innovatively applies contrastive learning to domain adaptation. BCDA utilizes a bootstrap clustering positive sampling strategy to ensure stable, end-to-end domain adaptation, preventing model collapse often seen in asymmetric networks. This method not only aligns domains internally through mean square loss but also enhances semantic inter-domain alignment, effectively eliminating false negatives. Our approach, BCDA, represents the first foray into non-contrastive domain adaptation and could serve as a foundational model for future studies. It shows potential to supersede contrastive domain adaptation methods in eliminating false negatives, evidenced by high-level results on three well-known domain adaptation benchmark datasets.

The quality of knowledge graphs (KGs) significantly influences their utility in downstream applications. Traditional methods for enhancing KG quality typically involve manual efforts and knowledge pattern learning to detect errors and complete missing triples. These approaches often incur high manual costs. To address these challenges, this paper proposes a novel “data-driven” approach to KG improvement. This method utilizes numerical data records to validate and enhance the information within KGs, overcoming limitations such as the requirement for a robust internal structure of KGs or the scarcity of expert resources. A pioneering technique that integrates Markov Boundary discovery with correlation analysis of data properties is developed in this study. This technique aims to identify and correct errors, as well as to fill in missing components of the KGs. To evaluate the effectiveness of this approach, experimental analysis was conducted, highlighting its potential to significantly improve KG accuracy and completeness. This data-driven strategy reduces reliance on extensive manual intervention and expert knowledge, introducing a scalable way to refine KGs using empirical data. The results from the experiments demonstrate the capability of this method to enhance the quality of KGs, marking it as a valuable contribution to the field of knowledge management.

Evolutionary multi-objective multitasking optimization (MTO) has emerged as a popular research field in evolutionary computation. By simultaneously considering multiple objectives and tasks while identifying valuable knowledge for intertask transfer, MTO aims to discover solutions that deliver optimal performance across all objectives and tasks. Nevertheless, MTO presents a substantial challenge concerning the effective transport of high-quality information between tasks. To handle this challenge, this paper introduces a novel approach named TL-MOMFEA (multi-objective multifactorial evolutionary algorithm based on domain transfer learning) for MTO problems. TL-MOMFEA uses domain-transfer learning to adapt the population from one task to another, resulting in the reproduction of higher-quality solutions. Furthermore, TL-MOMFEA employs a model transfer strategy where population distribution rules learned from one task are succinctly summarized and applied to similar tasks. By capitalizing on the knowledge acquired from solved tasks, TL-MOMFEA effectively circumvents futile searches and accurately identifies global optimum predictions with increased precision. The effectiveness of TL-MOMFEA is evaluated through experimental studies in two widely used test suites, and experimental comparisons have shown that the proposed paradigm achieves excellent results in terms of solution quality and search efficiency, thus demonstrating its clear superiority over other state-of-the-art MTO frameworks.

In the realm of maritime emergencies, unmanned aerial vehicles (UAVs) play a crucial role in enhancing search and rescue (SAR) operations. They help in efficiently rescuing distressed crews, strengthening maritime surveillance, and maintaining national security due to their cost-effectiveness, versatility, and effectiveness. However, the vast expanse of sea territories and the rapid changes in maritime conditions make a single SAR center insufficient for handling complex emergencies. Thus, it is vital to develop strategies for quickly deploying UAV resources from multiple SAR centers for area reconnaissance and supporting maritime rescue operations. This study introduces a graph-structured planning model for the maritime SAR path planning problem, considering multiple rescue centers (MSARPPP-MRC). It incorporates workload distribution among SAR centers and UAV operational constraints. We propose a reinforcement learning-based genetic algorithm (GA-RL) to tackle the MSARPPP-MRC problem. GA-RL uses heuristic rules to initialize the population and employs the Q-learning method to manage the progeny during each generation, including their retention, storage, or disposal. When the elite repository’s capacity is reached, a decision is made on the utilization of these members to refresh the population. Additionally, adaptive crossover and perturbation strategies are applied to develop a more effective SAR scheme. Extensive testing proves that GA-RL surpasses other algorithms in optimization efficacy and efficiency, highlighting the benefits of reinforcement learning in population management.

This paper investigates a new problem in an identical parallel machine environment called parallel machine scheduling with job family, release time, and mold availability constraints (PMS-JRM), which is highly challenging from the computational perspective as it extends the basic NP-hard problem (P_m||sum C_j). The mold availability notion, first introduced in this paper, represents the availability relationship between jobs and machines. The PMS-JRM model originates from the imaging data collaborative processing in a low-earth-orbit satellite constellation under a time-varying communication network, and it can represent other multi-resource collaborative scheduling problems with discontinuous communication. An integer programming model was proposed to formulate the PMS-JRM. Due to its NP-hardness, two highly efficient heuristic solution approaches were proposed, namely a greedy algorithm with a hybrid first come first serve (HFCFS) dispatching rule (GA-HFCFS) and a Memetic Algorithm with Heterogeneous swap and Key job insertion operators (MA-HK). Extensive experiments were conducted on a set of test cases with various scales, and the results showed that GA-HFCFS outperforms three classical dispatching rules available in the literature. Taking the results of GA-HFCFS as initial solutions, MA-HK achieves optimal solutions for all small-scale cases while providing superior solutions within the same running time compared to two other competitors for large-scale cases. In particular, MA-HK yields better solutions in less running time than the state-of-the-art CPLEX solver. Additional experiments were conducted to highlight the critical ingredients of MA-HK.

Learning-based methods have become increasingly popular for solving vehicle routing problems (VRP) due to their near-optimal performance and fast inference speed. Among them, the combination of deep reinforcement learning and graph representation allows for the abstraction of node topology structures and features in an encoder–decoder style. Such an approach makes it possible to solve routing problems end-to-end without needing complicated heuristic operators designed by domain experts. Existing research studies have been focusing on novel encoding and decoding structures via various neural network models to enhance the node embedding representation. Despite the sophisticated approaches being designed for VRP, there is a noticeable lack of consideration for the graph-theoretic properties inherent to routing problems. Moreover, the potential ramifications of inter-nodal interactions on the decision-making efficacy of the models have not been adequately explored. To bridge this gap, we propose an adaptive graph attention sampling with the edges fusion framework, where nodes’ embedding is determined through attention calculation from certain highly correlated neighborhoods and edges, utilizing a filtered adjacency matrix. In detail, the selections of particular neighbors and adjacency edges are led by a multi-head attention mechanism, contributing directly to the message passing and node embedding in graph attention sampling networks. Furthermore, an adaptive actor-critic algorithm with policy improvements is incorporated to expedite the training convergence. We then conduct comprehensive experiments against baseline methods on learning-based VRP tasks from different perspectives. Our proposed model outperforms the existing methods by 2.08–6.23% and shows stronger generalization ability, achieving the state-of-the-art performance on randomly generated instances and standard benchmark datasets.