We introduce a fuzzy–stochastic optimization framework for large-scale multi-agent platforms, where user engagement is uncertain, graded, and dynamically responsive to incentives. Unlike binary or purely probabilistic models, we employ fuzzy logic to capture continuous degrees of agent willingness, integrating both intrinsic motivation and incentive satisfaction into an interpretable engagement function. The system’s objective is formalized as a multi-agent Hamilton–Jacobi–Bellman (HJB) equation under global budget constraints, leading to closed-form policies and a principled estimator of the marginal value of engagement. We demonstrate scalability to high-dimensional agent populations, symbolic and numeric tractability in multi-agent settings, and practical recoverability of the optimal policy through a fuzzy actor–critic reinforcement learning implementation. Case studies validate robustness and fairness-aware adaptability for applications such as crowdsourcing, gig economy marketplaces, and mobility platforms. This work contributes a theoretically grounded, ethically interpretable, and computationally efficient framework for incentive control in web-scale stochastic environments, directly advancing the challenges of fairness, transparency, and adaptive resource allocation.
Federated Learning (FL) describes a model training technique in which the process is distributed among separate clients, holding private training data. During an FL round, locally-trained models from clients are aggregated, forming a new global model, which is then redistributed amongst the clients. During this process, the global model indirectly utilizes information from clients’ private data. However, accurate and adaptive measurement of the client’s contribution to the global model is an important aspect of the FL system, especially in na on-IID scenario. In this paper, we propose a fuzzy logicbased aggregation policy that dynamically estimates the learning demand of the global model and weights the importance of clients accordingly. The server evaluates class-wise precision and recall and infers per-class learning demand that reflects whether the model needs more training exposure to the samples of each class. Client-side class distribution and dataset size are then used to determine client’s contribution in the aggregation process. Unlike learned scheduling policies, the proposed approach provides transparent and easily interpretable decision rules using Sugenotype inference. The proposed methodology was tested on MNIST, EMNIST and CIFAR-10 benchmark datasets, showing promising results in a non-IID setting.
Nowadays, Large Language Models (LLMs) are widely included in many daily applications and services, where they play a crucial role in supporting users for automated decision-making tasks. As their adoption grows, so does their vulnerability to prompt injection attacks, which use natural language instructions to bypass safety constraints and also manipulate model behavior. To address these issues, the paper proposes a hybrid framework for prompt injection detection that combines domain-adaptive pretraining (DAPT) of RoBERTa encoder with a set of rule-based text patterns to detect suspicious adversarial interaction modes within prompt inputs. The proposed approach is evaluated through an empirical analysis across multiple prompt injection datasets. Experimental results show that the hybrid model outperforms text-only baselines and maintains reasonable performance under distributional shift. An ablation study is conducted to evaluate the individual component contribution.
Intelligent traffic cognition requires modeling spatiotemporal dynamic shifts not only to forecast traffic states but also to explain evolving patterns. Existing methods mainly optimize prediction accuracy and provide limited mechanisms for: (1) shift-aware node-wise sequential modeling, (2) gradual transitions of node-specific role, and (3) intrinsic interpretability for cognitive reasoning. In this paper, we propose FuzzMamba, a fuzzy-based Graph-Mamba framework that integrates: adaptive node-centric perception sequences learning, time-varying fuzzy role characterization, and explainable inference. FuzzMamba represents a novel paradigm for intelligent traffic cognition that effectively leverages the uncertainty modeling and interpretability strengths of fuzzy theory together with the sequential learning capability of Mamba. Experiments on four real-world traffic datasets demonstrate that the proposed approach provides competitive forecasting performance and interpretable insights into node-role evolution, thereby advancing intelligent traffic cognition beyond pure prediction.
Grouping and overlap operators are widely exploited in aggregation theory and information fusion to model disjunctive and conjunctive characteristics among the data inputs. In this study, we propose a new class of fusion operators, termed as the activation-based grouping and overlap operators. These operators are framed by introducing a threshold-based activation mechanism into the classical ideology of general grouping and overlap operators. The threshold parameter $\tilde{\tau}$ controls the point at which the activation mechanism becomes functional in the aggregation process and is chosen according to the necessities of the decision system. The main idea of the proposed framework is to retain the original behavior of the grouping or overlap operator in the non-activated region, while modifying the response of output near the boundary values by using a suitable activation function. Several construction methods for such operators are discussed, and their mathematical properties are examined. The class of activation-based overlap operators have been studied as a dual class of activation-based grouping operators. The proposed framework offers a flexible way to design a class of fusion operators that are sensitive towards the threshold parameter chosen close to the boundary points, and can be efficiently adapted to different decision scenarios. Due to their adaptive mechanism, such operators can be utilized in a wide range of applications, including decision-making, information fusion, classification problems, and reliability analysis.
