Integrated production-transport scheduling is increasingly required in flexible manufacturing systems where automated guided vehicles (AGVs) transport parts between machines. A persistent limitation in much of the literature is that AGVs are often modeled as continuously available resources, while battery state-of-charge (SoC) limits and explicit charging decisions are omitted or treated through simplified feasibility checks, which can yield infeasible or overly optimistic transport plans. We introduce DFJSP-EAGV, a dynamic flexible job shop variant that jointly optimizes machine schedules and AGV transports under battery depletion and charging decisions, and formulate it as a tri-objective problem minimizing makespan, total machine workload, and total AGV operating time. We propose Bi-NSGA-II-DQN, a bi-level evolutionary–reinforcement learning framework in which NSGA-II performs Pareto-based exploration using a tri-chromosome encoding, while a Deep Q-Network (DQN) selectively refines only AGV decisions of non-dominated solutions under feasibility masks enforcing timing and energy constraints. The framework is evaluated on 24 static instances derived from the Brandimarte, Fattahi, and Kacem benchmarks and compared with NSGA-II, NSGA-III, and SPEA2 under a common experimental protocol. Across the complete benchmark set, Bi-NSGA-II-DQN achieves the best aggregate mean performance in both normalized hypervolume ( HV_n ) and IGD^+ , with particularly strong results on the Brandimarte instances. Learning diagnostics provide evidence of active online DQN training and a progressive transition from exploration toward exploitation. An ablation analysis further examines the contribution and computational trade-offs of the learning-based refinement mechanism. Finally, disruption scenarios based on job insertion, job cancellation, job replacement, machine breakdown, and AGV breakdown illustrate the applicability of snapshot-and-restart rescheduling for generating feasible post-event schedules under updated resource availability.
Multimodal sequential recommendation aims to mine the temporal dependencies within user interaction sequences by fusing heterogeneous data such as text and images, so as to accurately model users’ dynamic preferences. However, existing methods remain insufficient in the deep fusion of structured attributes and visual semantics, and are prone to semantic drift and inference hallucination in long-sequence modeling, which limits their ability to cope with information overload in real-world scenarios. To address these issues, this paper proposes a Multimodal Knowledge Graph Retrieval-Augmented Generation framework (MKGRAG) for multimodal sequential recommendation. The framework first builds a multimodal knowledge graph from the textual and visual metadata of items, and then extracts local knowledge subgraphs related to candidate items through a subgraph retrieval module, which are converted into natural language descriptions by a graph-to-text conversion module. On this basis, a multimodal large language model is employed to extract the deep semantics of visual content, so as to construct item summaries that integrate structured knowledge and visual semantics. For user preference modeling, we design an LLM-based recurrent user preference summarizer that captures the temporal dynamics of user preferences by progressively refining user interest representations. In addition, a Low-Rank Adaptation (LoRA) strategy is adopted to parameter-efficiently fine-tune the large language model, reducing the training and deployment costs. Experimental results on three public benchmark datasets—MicroLens, Amazon-Games, and Amazon-Baby—show that MKGRAG improves AUC by 2.39, 1.90, and 3.56 percentage points over the strongest baseline MLLM-MSR, respectively, verifying its effectiveness in dynamic preference modeling and efficient information processing.
The transition to sustainable energy systems has heightened global interest in solar energy as a clean and renewable power source. Central to this technology is the photovoltaic (PV) cell, which directly converts solar radiation into electricity through the photovoltaic effect. Despite the relative simplicity of PV cell structures, accurately modeling their electrical behavior remains a complex task due to the nonlinear, high-dimensional, and multi-modal characteristics of the underlying optimization problems. To address these challenges, this study proposes an Improved Multi-Operator Differential Evolution (IMODE) algorithm tailored to address these challenges. IMODE incorporates multiple mutation strategies within the Differential Evolution framework and employs a self-adaptive mechanism that dynamically assigns strategies to individuals based on their historical success rates. The effectiveness of the proposed IMODE is first validated using the CEC2022 benchmark test suite. Subsequently, its performance is evaluated across six benchmark PV modeling problems, comprising three circuit-based models (single diode, double diode, and triple diode) and three commercial PV modules (Photowatt-PWP201, STP6-120/36, and STM6-40/36). Comparative analysis with seven state-of-the-art DE variants reveals that IMODE achieves superior estimation accuracy with root mean square error of 9.8602187789E–04 for the single-diode model, 9.8248485178E–04 for both the double-diode and triple diode models, 2.4250748680E–03 for Photowatt-PWP201 module, 1.6600603125E-02 for STP6-120/36, and 1.7298137099 STM6-40/36 for the STM6-40/36 module. In addition, IMODE demonstrates faster convergence, achieving optimal solutions with an average runtime of approximately 1 s. Furthermore, statistical validation through the Friedman and Wilcoxon tests highlights the robustness and superior overall performance of IMODE, positioning it as a reliable tool for accurate PV model identification.
The energy-efficient distributed flexible job shop scheduling problem (EE-DFJSP) requires the simultaneous optimization of makespan and total energy consumption. Due to conflicting objectives and the combinatorial complexity of the decision space, existing metaheuristic algorithms often exhibit premature convergence and insufficient adaptability in both local search and energy-saving strategies. To address these challenges, this paper proposes a hybrid memetic algorithm, termed QNiche-MA, which is built upon the integration of adaptive niche selection and Q-learning. A hybrid chaotic–heuristic initialization strategy is first proposed to improve the quality of the initial population. An adaptive niche-based parent selection mechanism is then introduced to preserve diversity and guide the evolutionary process. In addition, an adaptive multi-intensity energy-saving strategy is designed to reduce non-productive energy consumption by dynamically adjusting optimization intensity according to the population state. Furthermore, a Q-learning-based local search is incorporated to select neighborhood operators. Experimental results on benchmark instances show that QNiche-MA achieves competitive performance in terms of HV, IGD, and I_ϵ^+ , demonstrating its effectiveness in solving the EE-DFJSP.
Neural solvers for vehicle routing problems (VRPs) have emerged as powerful alternatives to traditional optimization heuristics, demonstrating superior computational efficiency and minimal dependence on manual rule design. However, while attention-based architectures excel at sequence generation, they predominantly rely on implicit coordinate projections. This coordinate-centric approach frequently leads to severe attention dispersion as the problem scale expands, failing to fully exploit the explicit geometric priors inherent in routing tasks (i.e., pairwise distances). To overcome this bottleneck, we propose the Distance Augmented Attention (DAA) architecture, a novel scheme that explicitly injects dense distance matrices into the solver’s decision-making process. Our approach introduces two key innovations: (1) a decreasing nested grouping strategy that dynamically shrinks the spatial receptive field across layers, elegantly guiding the model from establishing a broad global context to refining localized routing paths; and (2) an adaptive multi-head gating mechanism that allows different attention heads to autonomously regulate their reliance on these geometric priors. This synergistic design significantly enhances the solver’s spatial reasoning capabilities with negligible computational overhead. Extensive experiments demonstrate that DAA achieves state-of-the-art cross-scale generalization (scaling efficiently up to 1000 nodes) and exhibits remarkable robustness against spatial distribution shifts. This work provides a highly scalable and extensible framework for fusing geometric priors into neural combinatorial optimization.
