Branch-and-Bound (B&B) is the dominant exact solution method for Mixed Integer Linear Programs (MILP), yet its exponential time complexity poses significant challenges for large-scale instances. The growing capabilities of machine learning have spurred efforts to improve B&B by learning data-driven branching policies. However, most existing approaches rely on Imitation Learning (IL), which tends to overfit to expert demonstrations and struggles to generalize to structurally diverse or unseen instances. In this work, we propose Tree-Gate Proximal Policy Optimization (TGPPO), a novel framework that employs Proximal Policy Optimization (PPO), a Reinforcement Learning (RL) algorithm, to train a branching policy aimed at improving generalization across heterogeneous MILP instances. Our approach builds on a parameterized state space representation that dynamically captures the evolving context of the search tree. Empirical evaluations show that TGPPO often outperforms existing learning-based policies in terms of reducing the number of nodes explored and improving p-Primal-Dual Integrals (PDI), particularly in out-of-distribution instances. These results highlight the potential of RL to develop robust and adaptable branching strategies for MILP solvers.

Coordinating autonomous agents in Multi-Agent Reinforcement Learning (MARL) is hindered by challenges of scalability and effective information sharing, particularly in partially observable environments. While Graph Neural Networks (GNNs) model agent interactions, they often fail to leverage network structure, treating all connections as equally important. This paper introduces Connectivity-Influence Weighting (CIW), a mechanism that injects a structural inductive bias into the message-passing pipeline. CIW dynamically modulates message importance based on node connectivity, allowing agents to prioritize information from either highly connected or peripheral neighbors through two modes: Hub-Weighted (HW) and Fringe-Weighted (FW). We integrate CIW into a Transformer-based GNN with a decentralized actor–centralized critic architecture. Empirical evaluation on navigation tasks across varying agent densities shows that HW excels in sparse settings, while FW is superior in dense ones. CIW-enhanced models consistently outperform a graph-based baseline and approach oracle performance, highlighting the value of structurally guided information filtering for scalable and robust multi-agent coordination.

The Ball-Pivoting Algorithm (BPA) is a crucial technique for 3D surface reconstruction from point clouds, which relies heavily on selecting an appropriate ball radius. The effectiveness of BPA is significantly influenced by the point sampling density. In areas of low sampling density, a small ball radius can create gaps, while a large radius in high-density regions may oversimplify the surface and miss finer details. This paper addresses this challenge by introducing Auto-DBPA, an Automatic radius selection with Density-aware BPA. Auto-DBPA adapts dynamically by adjusting the ball radius according to the local sampling density. Our approach offers a scalable solution for reconstructing complex scenes and objects with varying levels of detail. Unlike conventional methods that require partitioning the point cloud into clusters and merging the reconstructed parts, which can cause visible seams and discontinuities, our unified reconstruction pipeline dynamically adjusts radii across the entire point cloud. To achieve this, we use hierarchical clustering and compute Fast Point Feature Histograms (FPFH) for each density cluster, capturing local geometric properties. We then leverage these geometric features to predict the optimal radius values for each cluster, adequately adapting the ball radius to local density variations. We also address the non-differentiability of BPA, which arises from its geometric computations and lack of gradient information, by introducing an innovative solution based on contextual bandits. Our approach employs the contextual bandits framework to effectively select the optimal ball radius based on local geometric features, significantly enhancing reconstruction quality. Our method is scalable and particularly effective in scenarios with varying density levels where a single-radius solution is inadequate. Results on the ABC, FAUST, SceneNN and ScanNet datasets demonstrate that our method successfully handles 3D reconstruction from varying point cloud densities, outperforming manual tuning, classic methods, and learning-based approaches. Our code will be available athttps://github.com/houda-pixel/Auto-DBPA.

Artificial potential field is a widely used path-planning algorithm with robotics, simulation, and traffic management applications. One of the main drawbacks associated with this technique is its susceptibility to falling in local minimas, with this problem only being exacerbated when having multiple agents. This paper proposes a novel decentralized approach for solving local minima problems for APF methods in multi-agent settings. Our approach leverages local observations of the agents, enabling real-time decision-making without requiring full global knowledge. We evaluate our method against established baselines in a controlled simulation setting. Our results demonstrate that the proposed approach outperforms baselines in terms of agent success rate (reaching goals) by up to 10% while recording the lowest amount of collisions. Furthermore, our method exhibits good scalability with increasing numbers of agents and good stability shown by the low variance in the results obtained from the randomized test environments.

Predictive Maintenance (PdM) is crucial in manufacturing, enabling early detection of equipment failures and optimized maintenance to improve reliability and reduce costs. Yet, extracting accurate and explainable predictions from complex time-series data remains challenging. Deep learning models such as Long Short-Term Memory (LSTM) networks offer strong predictive power but act as ‘black boxes,’ while fuzzy systems provide interpretability but lack robustness in dynamic industrial settings. To address these limitations, recent studies have integrated fuzzy inference with LSTMs. In this paper, we propose FiLSTM (Fuzzy Rule Induction for LSTM), a hybrid model that leverages Decision Trees to generate fuzzy rules and membership functions directly from data. This approach enhances both interpretability and prediction, achieving higher accuracy, faster computation, and lower execution time compared to FLSTM. FiLSTM advances PdM systems toward real-time, scalable, and transparent industrial applications.

Current AI-driven Question-Answering (QA) systems face significant challenges in delivering accurate, domainspecific responses across diverse fields. While Large Language Models (LLMs) excel at text generation, they often face difficulties in producing precise answers without access to external knowledge sources. Retrieval-Augmented Generation (RAG) addresses this limitation by connecting LLMs to external information sources. However, RAG systems often struggle to maintain the relevance and reliability of the retrieved information. Finetuning rerankers on domain-specific data could improve the selection of relevant text chunks before generation, but adapting them for multi-domain RAG settings remains challenging. Additionally, the cost of fine-tuning becomes significant, particularly when dealing with large models and extensive datasets. To address this limitation, we propose Domain Adaptive Document Reranking for Retrieval Augmented Generation. This new approach leverages the Mixture of Experts (MoE) paradigm when finetuning rerankers within RAG systems. Our framework employs a routing mechanism that dynamically directs queries to domainspecific experts, improving performance without sacrificing crossdomain generalization. Experimental results in the agricultural and medical domains demonstrate that our framework outperforms baselines, underscoring the critical role of domain expertise in refining retrieved information for accurate LLM-based QA in multi-domain settings.

The Ball-Pivoting Algorithm (BPA) is a notable technique for 3D surface reconstruction from point clouds, heavily reliant on the ball radius. In practical application, determining the optimal radius for BPA often necessitates iterative experimentation to achieve better reconstruction quality. BPA entails geometric computations like iterative pivoting, inherently lacking differentiability. In this paper, we tackle the dual challenges of radius selection and non-differentiability in BPA. Inspired by contextual bandits, we propose an innovative approach that learns the optimal radius based on local geometric features within point clouds. We validate our method on the ModelNet10 and ShapeNet datasets, showcasing superior surface reconstruction compared to manual tuning and other classic methods both for low and high point cloud densities. Our code is available at https://github.com/houda-pixel/Auto-BPA.

This research introduces a robust detection system against malicious network traffic, leveraging hierarchical structures and self-attention mechanisms. The proposed system includes a Packet Segmenter that divides a given raw network packet into fixed-size segments that are fed to the HPAC-IDS. The experiments performed on CIC-IDS2017 dataset show that the system exhibits high accuracy and low false positive rates while demonstrating resilience against diverse adversarial methods like Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), and Wasserstein GAN (WGAN). The model's ability to withstand adversarial perturbations is attributed to the fusion of hierarchical attention mechanisms and convolutional neural networks, resulting in a 0% to 10% adversarial attack severity under tested adversarial attacks with different segment sizes, surpassing the state-of-the-art model in detection performance and adversarial attack robustness.

In the era of 5G and beyond, Unmanned Aerial Vehicles (UAVs) have emerged as a promising technology to extend and improve wireless communication infrastructure. This is thanks to their flexibility, low cost, and high probability of establishing line-of-sight connections. As the UAV s' missions increase in length and complexity, battery life becomes the main bottleneck for deploying fully autonomous UAV systems at larger scales. In practice, battery replenishment operations are inevitable to avoid increased costs. However, carefully scheduling these operations is crucial to prevent hurting the mission's perfor- mance. In this work, we propose a fully autonomous scheduling solution based on Proximal Policy Optimization (PPO) that aims to reach the required mission duration without losing a critical number of UAVs and maximizing the aggregate throughput in a free path loss scenario. Extensive simulations have shown that the proposed algorithm outperforms baselines by 119.55% and 84.66% in maximizing the lifetime metric of UAVs and the minimum throughput for users, respectively.

Target recognition from underwater acoustic signals is a major challenge in surveillance systems, especially in military and defense fields. Deep learning models are increasingly used for the automatic classification of underwater signals, but many challenges remain due to the complexity of sound navigation and ranging networks, the noise present in the signals, and the difficulty of collecting large amounts of data for efficient training. In this paper, we propose two new architectures for underwater signal classification based on Spatio-temporal modeling. In experiments, evaluations on two real datasets show that the proposed approach achieves a classification accuracy of 98% which outperforms the state-of-the-art methods. In addition, the proposed end-to-end network is considerably faster than MFCC-based networks such as Yamnet and VGGish.

3D point clouds are a simple and compact data format that represents the surface geometry of 3D objects. The output of the data acquisition process often yields incomplete shapes. Hence, it is crucial to infer the missing regions of 3D objects from incomplete ones for many real-world applications. By leveraging a framework of 3D point cloud completion architectures, the proposed inception module is an intermediate layer that aims to extract the hierarchical features, recognize the fine-grained details of point clouds and avoid overfitting. We conduct comprehensive experiments on three state-of-the-art datasets: ShapeNet-55, ShapeNet-34, and PCN. The experimental results demonstrate that the enhanced architectures outperform the state-of-the-art point cloud completion methods.

Agent-based modelling has been widely studied for observing moving targets and patrolling.However, in general, the studies are interested either in observation in a continuous environment or patrolling in a graph representation.In order to deal jointly with observation and patrolling, a common representation of the environment is required.In this paper, we firstly proposed a new environment representation's formalism, merging both agent-based distributed patrol and observation method.Secondly, we implemented a new approach called I-CMOMMT to cope with a trade-off between observation and patrolling using our new formalism.The obtained results are compared with other methods to show the efficiency of our approach.

This work proposes an architecture for sharing devices' resources in the Internet of Things providing real sensor data for its users.The main idea is based on the fact that users such as developers and researchers do not always have access to the necessary hardware and resource sharing should impact these persons activities.Taking advantage of the Sensors as a Service model, we propose an architecture where several sensors and actuators can be coupled to environments and they also are represented virtually in a web system becoming available to be consumed by users and platforms.The architecture is composed of three layers and a model representing devices, the cloud, and clients, and how they interact with each other.A study case for testing the whole approach is also presented.

This paper proposes an architecture for sharing IoT Objects’ resources in the Internet of Things providing a model for its owners to expose devices, which can be consumed by clients inspired by the Sensor-as-a-Service model. The main idea relies on the fact that users, such as developers and researchers, do not always have access to the necessary hardware and resources. Exposing devices in IoT should impact these persons activities. Then, we present the Resource Management Architecture, where several IoT Objects endowed with sensors and actuators can be added to environments that are represented virtually in the architecture. The IoT Objects become available to be consumed by users through the use of applications. The architecture is composed of three layers: one representing devices, the cloud solution, and applications, and how they interact with each other. We also present a study case for testing the whole approach in a smart city scenario.

We propose a conceptually simple new decentralised and non-communicating strategy for the multi-agent patrolling based on the LSTM architecture. The recurrent neural networks and more specifically the LSTM architecture, as machines to learn temporal series, are well adapted to the multi-agent patrol problem to the extent that they can be viewed as a decision problem over the time. For a given scenario, a LSTM neural network is first trained from data generated in simulation for that configuration, then embedded in agents that shall use it to navigate through the area to patrol choosing the next place to visit by feeding it with their current node. Finally, this new LSTM-based strategy is evaluated in simulation and compared with two representative strategies, a cognitive and centralised one, and a reactive and decentralised one. Preliminary results indicate that the proposed strategy is globally not better than the representative strategies for the aggregating criterion of average idleness, but better than the decentralised representative for the evaluation criteria of mean interval and quadratic mean interval.

The increasing number of moving vehicles along roads and the lack of supporting infrastructure is a wellestablished problem. Major consequences are augmenting of traffic jams, accidents, fuel consumption and pollution. Vehicular Ad hoc NETworks (VANETs) represent opportunities to deal with the aforementioned problems. In VANETS, efficiency and safety to applications are provided using communication support. In efficiency applications, each vehicle is aware of its location. Using this information and communication support, vehicles collaborate to reduce travel time and to improve mobility. In contrast, safety applications aim to reduce or even avoid accidents, and must obey strong timing constraints. In this context, VANETs applications can benefit from Computational Intelligence (CI) and adaptive approaches to implement the required demands. Thus, the contribution of this paper is twofold: $( i)$ we discuss how VANETs can benefit from CI and Artificial Intelligence techniques to make transportation networks more efficient regarding to safety applications, and, $( ii)$ we report our current work and new directions in the development of efficiency applications to VANETs using adaptation and CI techniques.

Ubiquitous systems consider the use of electronic components for enhancing daily objects with some kind of computational intelligence for aiding users in their tasks pervasively.Ambient Intelligence (AmI) is a branch of ubiquitous computing that provides an environment full of interconnected devices and it can provide data communication, inference mechanism based on context information and collaboration among system's devices.Similarly, the Internet of Things (IoT) provides uniquely identified devices or things in a network for helping users in their activities.Multi-Agent Systems (MAS) are intelligent systems where agents are responsible for reasoning, competing and using resources to achieve desirable goals pro-actively and autonomously.Agents have been employed in some approaches and works during the last years, but none of them considered embedded MAS responsible for smart devices in an AmI system running over an IoT network.Besides, it is also interesting that agents of the embedded MAS can interact, sharing information with agents situated in another embedded MAS using the IoT network to learn from user's experiences.This paper proposes an architecture for the development of AmI systems using embedded MAS for interfacing with sensors and actuators in a heterogenous network using an IoT middleware.

Nowadays industrial process are mainly distributed, and their supervision systems are still centralized. Consequently, when communications are disrupted, it slows down or stops the supervision process. To allow the anytime supervision of such systems, we propose a distributed approch based on a multi-agent system where each supervision agent autonomously handles both diagnosis and repair on a given location. We demonstrate the advantage of our proposal and evaluates ADS2 using an industrial case-study. Experiments demonstrate the relevance of our approach with an overall reduction of the supervised system down-time of 34%.

In this paper, we propose an architecture that uses a tender protocol, the Contract Net Protocol (CNP), to let human operators express their consent about the allocation of goals to Unmanned Aerial Vehicles (UAVs) in a constrained environment. The CNP has several good points: it has an appropriate level of automation, it is simple, it spares the bandwidth. But if bids are evaluated solely on the base of a numerical payoff, the CNP cannot fully convey human preference in complex situations. Thus, we extended and implemented the CNP within our framework, Aerial, to enable a more subjective human feedback. We detail how we build the bids and how we filter them to not flood the user. We also explain how we enable a dependable commitment in a dynamic context when the award time is not accurately foreseen.

The multi-agent patrolling task constitutes a challenging issue for Artificial Intelligence and has the potential to cover a variety of domains ranging from agent-based simulations to crises management. Several techniques have been proposed in the last few years to address the multi-agent patrolling task with a closed-system setting. A few centralized strategies were also described to address the open-system setting, in which the agents can enter or leave the patrolling task at will. In this article, we propose two decentralized, cooperative, auction-based strategies in which agents trade the nodes they have to visit. These strategies are inspired from the computational social choice theory and allow the agents to reason on the performances of the group rather than on their own. We show that these strategies perform at least as well as the state-of-the-art centralized performances, and better on specific criteria.

In this paper, we present an approach to coordinate two temporal plans, where one plan is for the achievement of a reactive goal, and thus has more priority, while the other is for the achievement of a proactive goal. As both plans have been computed independently from each other, conflicts can arise during their execution. We present a plan merging algorithm to coordinate the actions of both plans. Unlike many other plan merging approaches, which use only the actions present in both plans to coordinate them, our approach is supported by a sound plan repairing approach, to deal with the situations where both plans can not be merged as they are i.e. when one plan uses the same resource needed by the other plan and then does not release the resource. The plan repairing algorithm uses an extension of the well known planner Sapa, to repair a plan in the presence of another plan, and thus taking into account the constraints imposed by the latter. Our approach is generic and can be used in different situations of coordination between multiple agents and within a single agent.

This article addresses collaborative learning in a multi-agent system: each agent revises incrementally its beliefs B (a concept representation) to keep it consistent with the whole set of information K (the examples) that he has received from the environment or other agents. In SMILE this notion of consistency was extended to a group of agents and a unique consistent concept representation was so maintained inside the group. In the present paper, we present iSMILE in which the agents still provide examples to other agents but keep their own concept representation. We will see that iSMILE is more time consuming and loses part of its learning ability, but that when agents cooperate at classification time, the group benefits from the advantages of ensemble learning.

A challenge for pervasive computing is the seamless integration of computer support with users' activities in a very dynamic setting, with deep human and resource mobility. Portable devices such as notebooks, PDAs and smartphones are becoming more and more popular, as their computational power increases and their prices fall. Moreover, the quick spread of wireless networks has permitted the exchange of data in places as diverse as university campus, airports, coffee houses and family homes. In this paper we introduce campus, a framework that supports the development of multi-agent, context aware, pervasive computing applications. Campus is designed to provide the necessary infrastructure for ambience intelligence applications.