Volume 34, Number 4 (2026)

---

Special Issue: Complex Systems and Intelligent Infrastructures
Mohamed Nemiche, Souad Tayane and Mohamed Essaaidi

In an era increasingly defined by uncertainty, interconnectedness and systemic transformation, complexity science has emerged not merely as a theoretical framework, but as an essential lens for making sense of the world. From network dynamics and artificial intelligence to climate tipping points and cultural epistemologies, the field of complex systems continues to expand its relevance and application. This special issue brings together six diverse yet interwoven contributions that collectively explore how intelligent infrastructures and emergent systems are reshaping our understanding of resilience, learning and adaptation in the twenty-first century.  

Cite this publication as:
M. Nemiche, S. Tayane and M. Essaaidi, “Special Issue: Complex Systems and Intelligent Infrastructures,” Complex Systems, 34(4), 2026 pp. i–ii.
https://doi.org/10.25088/ComplexSystems.34.4.i

---

Complexity in the Twenty-First Century: From the Limits of Growth to the Growth of Limits
Reda Benkirane

Complexity, a term that is both ambiguous and multifaceted, is used widely today. Various legitimate definitions can be proposed for it, as is the case with “ample” notions such as intelligence, consciousness or culture. The recurrent mention of this term can be attributed to the transformation of our societies and their artifacts, as well as the acceleration of time brought by the digital revolution—a technological upheaval comparable to the invention of writing and the printing press.

Cite this publication as:
R. Benkirane, “Complexity in the Twenty-First Century: From the Limits of Growth to the Growth of Limits,” Complex Systems, 34(4), 2026 pp. 387–400.
https://doi.org/10.25088/ComplexSystems.34.4.387

---

A Network Agent-Based Model for Moroccan Inbound Tourism: Incorporating Social Influence in the Decision-Making Process
Smahane Jebraoui, Sidati Khabid and Mohamed Nemiche

Tourism plays a critical role in Morocco’s economy, supported by its rich cultural heritage and diverse attractions. Understanding the decision-making processes of inbound tourists is essential for enhancing destination appeal and promoting sustainable growth. This paper proposes a network agent-based model for Moroccan inbound tourism, incorporating social factors into tourist decision-making processes. By simulating interactions within a network of agents, the model highlights the role of social influence in shaping tourist behaviors. It provides a comprehensive framework for examining how individual decisions emerge from the interplay between personal preferences and social environments.

The paper also explores various scenarios to evaluate the effects of social influence, the promotion of lesser-known destinations and the repercussions of negative reviews. The simulation results offer valuable insights into strategies for sustainable tourism development, emphasizing the importance of leveraging social dynamics to optimize tourism policies and enhance the visitor experience.

Keywords: Moroccan tourism; social influence; social network; tourist decision-making; agent-based model

Cite this publication as:
S. Jebraoui, S. Khabid and M. Nemiche, “A Network Agent-Based Model for Moroccan Inbound Tourism: Incorporating Social Influence in the Decision-Making Process,” Complex Systems, 34(4), 2026 pp. 401–424.
https://doi.org/10.25088/ComplexSystems.34.4.401

---

Research on Self-Learning Method for Key Nodes Identification in Heterogeneous Networks
Luyao Wang, Zhiwei Yang, Kewei Yang and Libin Chen

Identifying key nodes in heterogeneous networks is both theoretically important and practically valuable. Traditional methods require precise parameters and constraints, limiting adaptability and autonomy. To address this, we propose the deep reinforcement learning–based heterogeneous network key nodes identification (DRLKHN) method, a self-learning method for identifying key nodes. DRLKHN autonomously learns strategies for identifying key nodes, utilizing a graph convolution network (GCN) for feature extraction and designing action and state space vectors. Experimental results show that DRLKHN outperforms traditional methods like high degree adaptive (HDA), high eigenvector adaptive (HEA), high closeness adaptive (HCA) and high PageRank adaptive (HPA) in simulated networks. In the real-world force, intelligence, networking and C2 (FINC) network, DRLKHN improves performance by 28.6%, 32.2%, 12.7% and 36.3% over HDA, HEA, HCA and HPA, respectively. Despite its relatively high time complexity, DRLKHN effectively integrates the GCN and reinforcement learning to manage complex relationships in graph data, providing intelligent decision support for identifying key nodes in real networks.

Keywords: heterogeneous network; key nodes identification; graph representation; deep reinforcement learning

Cite this publication as:
L. Wang, Z. Yang, K. Yang and L. Chen, “Research on Self-Learning Method for Key Nodes Identification in Heterogeneous Networks,” Complex Systems, 34(4), 2026 pp. 425–454.
https://doi.org/10.25088/ComplexSystems.34.4.425

---

A Reinforcement Learning–Based Approach for Smart Lighting and Shading in Buildings
Fadwa Lachhab and Mohamed Bakhouya

Lighting systems in commercial and residential buildings constitute a major source of the world energy consumption. Optimizing energy efficiency through lighting management requires an optimal control strategy in order to balance daylighting requirements while maintaining visual comfort in illuminated spaces. This paper introduces a reinforcement learning (RL)–based approach using the Q-learning algorithm to optimize lighting and shading control, maintaining constant illuminance with maximum visual comfort. A prototype was developed in a laboratory to test the scenario, using internet of things (IoT) and artificial intelligence (AI) technologies, for lighting and shading control. AI techniques are integrated to enable a smart conversation between lighting and shading systems in order to maintain the required light level. A real-time chatbot based on natural language processing (NLP) is integrated with IoT techniques in order to provide a user-friendly building automation system. Experiments have been conducted for validation purposes and obtained results show the effectiveness of the proposed solution by maintaining the ideal level of lighting with efficient consumption. In fact, the proposed control is capable of optimizing energy consumption by more than 45% against a normal lighting operation while maintaining occupants’ visual comfort within a suitable illuminance.

Keywords: lighting control; shading control; energy efficiency; Internet of Things; artificial intelligence; visual comfort

Cite this publication as:
F. Lachhab and M. Bakhouya, “A Reinforcement Learning–Based Approach for Smart Lighting and Shading in Buildings,” Complex Systems, 34(4), 2026 pp. 455–478.
https://doi.org/10.25088/ComplexSystems.34.4.455

---

Analysis and Estimation of the Probability of Failures in an Electric Lighting Network Using Mamdani Inference and Scaled Conjugate Gradient
Mohammed Amine Jouahri, Zakaria Boulghasoul and Abdelouahed Tajer

This paper presents an intelligent fault detection system (FDS) for public lighting networks, designed to improve diagnostic accuracy and system reliability. The proposed system integrates the Mamdani fuzzy inference method and scaled conjugate gradient (SCG) neural networks to detect four key fault types: power, lighting, cloud cover sensor and road flow sensor. Inputs such as traffic flow, cloud cover, power supply and lighting intensity are used to ensure precise diagnostics. The Mamdani method offers strong interpretability and robustness in handling uncertainties, while the SCG algorithm enhances performance through efficient learning. Simulation results show fault detection probabilities exceeding 85%, confirming the effectiveness of the system. This paper demonstrates the potential of combining fuzzy logic and neural networks for reliable and intelligent monitoring of public lighting infrastructure.

Keywords: public lighting networks; fault detection; power faults; lighting faults; sensor faults; Mamdani inference; scaled conjugate gradient

Cite this publication as:
M. A. Jouahri, Z. Boulghasoul and A. Tajer, “Analysis and Estimation of the Probability of Failures in an Electric Lighting Network Using Mamdani Inference and Scaled Conjugate Gradient,” Complex Systems, 34(4), 2026 pp. 479–507.
https://doi.org/10.25088/ComplexSystems.34.4.479

---

Exploring Federated Deep Learning for Internet of Things Cyberattacks
Abderahmane Hamdouchi and Ali Idri

The Internet of Things (IoT) connects billions of devices that operate autonomously, increasing the risk of cyber threats, such as theft and manipulation of personal data. This has increased interest in utilizing deep learning (DL) methods to develop intrusion detection systems (IDS). In general, DL-based IDS rely on centralized approaches, which require IoT devices to transmit data to central servers for analysis. However, these centralized methods raise privacy concerns, prompting the adoption of federated learning (FL) as a promising alternative. This paper evaluates and compares various FL configurations using dense neural networks (DNNs) and convolutional neural networks (CNNs) as base models. The research explores three aggregation methods (FedAVG, FedPROX and FedSGD), three device counts (5, 15 and 30), two data setups (raw and balanced) and two feature selection methods (analysis of variance and chi-squared) with two feature thresholds (50% and 100%). The evaluation was conducted on the NF-ToN-IoT-v2 and NF-BoT-IoT-v2 datasets, using the Scott–Knott test and the Borda count method to analyze 144 FL configurations. The results indicate that FedAVG and FedPROX outperform other aggregation methods, with DNNs identified as the most effective base model for attack detection in FL environments. The top-performing models, using only 17 features, were DNN_R50_PROX_30 (accuracy of 97.80%) and CNN_R50_PROX_5 (accuracy of 99.87%) for NF-BoT-IoT-v2 and NF-ToN-IoT-v2, respectively.

Keywords: intrusion detection system; federated learning; deep learning; NetFlow; IoT; cybersecurity

Cite this publication as:
A. Hamdouchi and A. Idri, “Exploring Federated Deep Learning for Internet of Things Cyberattacks,” Complex Systems, 34(4), 2026 pp. 509–542.
https://doi.org/10.25088/ComplexSystems.34.4.509