Abstract Metaheuristic optimization algorithms are essential for solving complex and high-dimensional problems, yet their performance is often constrained by premature convergence and insufficient population diversity. Although the Emperor Penguin Optimizer (EPO) has demonstrated promising results, its single-population structure and static search behavior limit robustness and scalability. To overcome these drawbacks, this paper proposes a Multi-Population Emperor Penguin Optimizer with Joint Opposite Selection (MLPEPO-JOS), an adaptive multi-population optimization framework that introduces a process innovation in how population coordination, diversity preservation, and opposition learning are integrated. The proposed approach integrates three complementary mechanisms: an adaptive multi-population regulation strategy driven by real-time fitness improvement trends, an intelligent modular arithmetic-based division scheme that ensures balanced population partitioning under arbitrary population sizes, and a Joint Opposite Selection (JOS) mechanism that combines Selective Leading Opposition and Dynamic Opposition to enhance exploration-exploitation balance. MLPEPO-JOS is evaluated on 41 benchmark functions from the IEEE CEC2017 and CEC2022 suites, achieving the best mean ranking and competitive performance in terms of accuracy, robustness, and convergence behavior. Statistical analyses indicate significant improvements over many of the evaluated competitors, although the significance varies across individual comparisons.
Abstract Trajectory prediction is a key task in autonomous driving, where a model must forecast plausible future motions of surrounding agents from dynamic histories and scene context. Many recent methods use high-definition (HD) maps to provide explicit road topology and semantic constraints, but such maps are costly to maintain, geographically limited, and difficult to update in changing environments. This paper presents Bird’s-Eye View Trajectory Prediction (BEVTraj), a map-free trajectory prediction framework that directly uses BEV features constructed from raw multi-sensor data. Since sensor-derived BEV features are dense and only implicitly structured, BEVTraj employs deformable attention to selectively aggregate agent-relevant spatial context from sparse locations rather than processing the entire BEV grid uniformly. In addition, a Sparse Goal Candidate Proposal (SGCP) module combines overlapping mode-wise anchor clusters with attention-based goal prediction to generate compact, scene-adaptive goal hypotheses. These goals are decoded and iteratively refined into multimodal trajectories, while trajectory-aware mode scoring and temporal consistency modeling improve confidence estimation and suppress abrupt, physically implausible refinement updates. Experiments on nuScenes and Argoverse 2 Sensor show that BEVTraj achieves competitive results under a map-free setting without using pre-built HD maps, while ablation studies and qualitative analyses illustrate the roles of BEV context aggregation and sparse goal reasoning. The source code is available at https://github.com/Kongminsang/bevtraj.
Abstract Time series classification has achieved strong performance through transformation-based methods such as MiniROCKET, which rely on fast random convolutional feature extraction from raw temporal signals. However, representing time series data in a single domain limits the ability to capture complex temporal–spectral and structural characteristics in diverse real-world datasets. In this study, we propose ManyROCKET, an adaptive parallel MiniROCKET framework that integrates multiple complementary signal representations within a unified and computationally efficient classification pipeline. The framework incorporates seven established transformations spanning time-domain, frequency-domain, time–frequency, analytic, and derivative-based representations. By systematically combining these transformations, we explore a broad set of multi-representation configurations on benchmark datasets from the UCR Time Series Archive to identify the most effective representation subset. The selected configuration is then evaluated at archive scale and directly compared with the Raw Time (Identity) MiniROCKET baseline. To ensure scalability, we introduce an adaptive two-tier parallel execution strategy that enables concurrent processing of multiple representations while preserving the core efficiency of MiniROCKET-style feature extraction. Extensive experimental results demonstrate that multi-representation learning consistently and robustly outperforms the standard raw time-domain MiniROCKET baseline across diverse datasets. The proposed ManyROCKET framework achieves superior classification performance while maintaining practical computational efficiency, highlighting the effectiveness of integrating complementary signal domains within a unified architecture.
Abstract The Escape algorithm (ESC), which simulates crowd evacuation behavior, is a novel algorithm. However, it has certain limitations in terms of convergence speed and the avoidance of local optima. This paper proposes a multi-strategy enhanced variant (ESCSA), which introduces three key improvements: First, it incorporates a dynamic mask probability adjustment mechanism that adapts the dimension update ratio throughout the iteration process, thereby achieving a better balance between exploration and exploitation. Second, it incorporates stagnation detection and Simulated Annealing local search. When the optimal solution shows no significant improvement over several consecutive iterations, the Simulated Annealing optimization mechanism is triggered to generate Gaussian-perturbed neighboring solutions, avoiding local optima via the Metropolis criterion. Third, it implements a worst-elimination strategy, periodically injecting randomly positioned new individuals to replace the worst solutions in the population, thereby maintaining diversity. On the CEC 2017 benchmark suite (10/30/50/100 dimensions), ESCSA achieved the best overall performance among 11 algorithms, ranking first in the Friedman mean rank test across all four dimensions with average ranks of 2.217, 1.383, 1.400, and 1.933, respectively. It obtained the optimal mean fitness on 17, 22, 21, and 17 functions for the four dimensions, respectively. In three engineering design problems (speed reducer design, multi-disc clutch braking, and rolling element bearing design), ESCSA ranked second, third, and first, respectively, demonstrating strong and consistent performance across constrained optimization tasks. When subsequently applied to path planning for mobile robots in static environments, ESCSA was experimentally confirmed to plan robot paths and identify superior solutions efficiently.
Abstract The COVID-19 pandemic overwhelmed hospital capacities, prompting the use of patient isolation units (PIUs) to flexibly expand containment in emergency rooms (ERs). However, existing studies primarily rely on static analyses and rarely quantify the direct infection risk to healthcare workers (HCWs). To address these, we propose an integrated framework to simultaneously minimize the aerosol dispersion index (ADI) and infection risk index (IRI) under dynamic conditions. A multistage methodology integrating computational fluid dynamics, surrogate modeling, and the non-dominated sorting genetic algorithm II was developed to analyze 120 dynamic scenarios. Optimal solutions were derived to quantify the improvements in ADI and IRI relative to general operation. Furthermore, the technique for order preference by similarity to ideal solution algorithm was employed to determine the optimal distributions of the PIU operational parameters and HCW behavioral factors based on weighted priorities. The optimized strategy demonstrated enhanced improvements in both ADI and IRI compared to the general operation. Optimal control minimized instances where IRI exceeded levels observed without a PIU, while simultaneously enhancing ADI. Variable analysis revealed that the optimal fan filter unit operation depends heavily on the specific dynamic conditions. Furthermore, maintaining physical distance emerged as a consistently dominant factor in reducing IRI, whereas respiratory avoidance efficacy was largely negated by turbulence. This study highlights the necessity and effectiveness of adaptive optimization in PIUs. Implementing intelligent strategies and adaptive control algorithms that respond to real-world environmental changes is essential for comprehensive infection control in high-risk environments.
