Group Relative Policy Optimization (GRPO)-style post-training improves large language models (LLMs) by comparing responses sampled for the same prompt without training a separate value model. From an evolutionary-computation perspective, each update is a short-lived population-selection process: responses are generated, evaluated, selected, and partially inherited by the next policy. Because the sampled group disappears after one update, rare but reusable response modes may be absent from later groups and receive no survival mass. We call this failure group-relative selection myopia. We propose ASPIRE, an archive-protected survival-selection method for response evolution. ASPIRE combines fresh rollouts, variation offspring, and retrieved mode elites in a heterogeneous response population, then solves a constrained Kullback-Leibler (KL) survival projection. Quality determines relative preference, violation suppresses invalid candidates, and feasible archive-supported modes missing from the fresh group receive mode-level survival floors. The resulting survival weights guide policy learning from selected responses, while feasibility-first competition refreshes the mode-elite archive. The projection separates within-mode quality-violation weighting from cross-mode protected mass allocation. Compared with DARLING, ASPIRE improves AIME25 frozene-valuator mode coverage by 10.5%. On constrained tool calling, the Berkeley Function Calling Leaderboard average rises from 32.8 to 34.5, while the violation rate falls from 10.3 to 8.5.
Automatic Modulation Classification (AMC) plays a critical role in spectrum sensing and cognitive radio. Although deep learning has greatly advanced AMC, most methods still struggle under low signal-to-noise ratio (SNR) conditions. Multi-view fusion provides complementary information for more robust signal classification, yet existing approaches typically rely on fixed, task-independent fusion strategies. Evolutionary fusion can automatically explore fusion strategies, but it often produces candidate individuals whose quality is highly uneven, leading to unnecessary computational cost on low-value fusion strategies. To address these limitations, we propose EvoAMC, an evolutionary fusion with knowledge learning method for AMC. EvoAMC produces multiple candidate offspring in each iteration, allowing the search to explore more possible fusion structures. We also utilize a Graph Convolutional Network (GCN)-based knowledge library model, which mines historical architecture-performance knowledge and transfers it across generations, to filter out unpromising candidates before training and guide the evolutionary process. Experiments show that EvoAMC achieves better accuracy under low-SNR conditions and a smaller parameter size, outperforming compared multi-view methods in both effectiveness and efficiency.
Graph neural networks (GNNs) provide a powerful learning paradigm for combinatorial optimization (CO). However, many real-world CO problems involve multiple entity and relation types, whereas existing approaches predominantly use manually designed architectures or target homogeneous graphs. To address this limitation, we propose an evolutionary graph neural architecture search framework for heterogeneous graph-based CO, termed EHG-CO. The framework constructs a heterogeneity-aware search space combining flexible topologies with heterogeneous aggregation operations. A node-wise edge-pair encoding jointly represents architectural topology and edge-specific aggregation operations, supporting compact representation and evolutionary optimization. To reduce architecture evaluation costs, an online multi-level rank-based predictor estimates candidate quality, while a rank-aware uncertainty filling strategy selects informative and promising candidates for actual evaluation and predictor refinement. Experiments on six benchmark heterogeneous graph-based CO problems demonstrate that EHG-CO achieves superior solution quality and competitive search efficiency compared with representative methods. Ablation studies further confirm the effectiveness of the search space, encoding scheme, predictor, and filling strategy.
Airport ground movement (AGM) optimization plays a crucial role in improving the operational efficiency of airport surface systems. The multi-objective multigraph model has been developed to optimize taxi time and fuel consumption with variable speed profiles. However, existing studies based on this model optimize aircraft sequentially under first-come-first-served principles, which inherently restricts the attainable solution space of the entire fleet. To overcome this limitation, we propose a global multi-objective multigraph AGM model that simultaneously optimizes all aircraft. This global formulation results in a high-dimensional combinatorial optimization problem with tightly coupled spatiotemporal constraints. Therefore, we develop a Q-learning-enhanced multi-objective evolutionary algorithm. First, a heuristic-guided decoding procedure incorporating topological information is designed to reduce path infeasibility, and a spatiotemporal repair mechanism is further developed to recover the remaining infeasible solutions. Second, four mating pools are designed with different parent partition schemes and variation operators to explore different search directions in the high-dimensional space. Finally, a Q-learning-based mechanism is developed to adaptively select active mating pools at each generation based on population diversity and convergence status, thereby concentrating computational resources on the most effective search directions. On most of the 16 real-world instances, MPRA-QL outperforms state-of-the-art methods in hypervolume and inverted generational distance.
Feature selection is a complex multiobjective optimization problem that requires balancing the number of selected features and classification performance. Redundant features and complex feature interactions pose significant challenges for traditional multiobjective feature selection algorithms. Existing methods typically employ either discrete or continuous encoding schemes, both of which have inherent limitations, i.e., discrete encoding struggles to capture complex feature interactions, while continuous encoding incurs high computational costs when handling redundant features. To address these issues, this paper proposes a new evolutionary multiobjective feature selection algorithm with a joint encoding mechanism, incorporating a novel joint encoding region partitioning method, an effective initialization strategy, and a new search strategy. Specifically, the joint encoding region partitioning method groups features into discrete or continuous search regions based on their importance and label relevance. An effective initialization strategy is designed to generate high-quality initial solutions in both regions. In addition, a new search strategy dynamically adjusts the number of features in discrete search regions to improve exploration. Experimental results on 16 datasets show that the proposed approach surpasses nine evolutionary multiobjective feature selection algorithms. The proposed approach effectively enhances population diversity while achieving a better balance between the number of selected features and classification performance.
