Abstract
MENTOR is aimed at developing novel control strategies for coordinating complex multi-agent systems in challenging and uncertain environments, with a particular focus on search and rescue operations during natural disasters. The core objective of the project was to integrate control-theoretic methodologies with machine learning techniques, in particular reinforcement learning, to design scalable and robust strategies capable of steering the collective behaviour of large groups of agents. To this end, the project first conducted a comparative analysis of existing control and multi-agent reinforcement learning approaches applied to complex multi-agent systems. Building on this analysis, new learning-based control strategies were developed to improve efficiency, robustness, and learning speed in complex scenarios. The proposed methods were analysed from a theoretical perspective to assess their convergence and robustness properties and were subsequently tested on representative applications involving coordinated multi-agent behaviour. The project combined complementary expertise in control engineering, machine learning, and complex systems, through collaboration between research units at the University of Naples Federico II and the University of Bologna, with additional international collaborations bringing in expertise in collective behaviour and distributed decision-making. As far as the outcomes are concerned, the project aimed at the Identification of a set of benchmark MARL strategies for complex multi-agent systems in the current literature and their application to SAR scenarios. This was achieved by systematically analysing representative MARL paradigms (e.g., centralized training with decentralized execution, value-decomposition methods, and hierarchical RL), identifying their scalability and robustness limitations in large-scale coordination tasks, and establishing structured benchmark problems inspired by shepherding and swarm coordination as abstractions of SAR operation. Another focus was the development of efficient and scalable learning-based control strategies. Significant progress was achieved through the integration of reinforcement learning with control-theoretic principles. Importantly, these learning-based methodologies were not developed in isolation but were grounded in a strong theoretical understanding of collective dynamics and influence mechanisms.
Dettagli del progetto
Responsabile scientifico: Mirco Musolesi
Strutture Unibo coinvolte:
Dipartimento di Informatica - Scienza e Ingegneria
Coordinatore:
Università degli Studi di NAPOLI Federico II(Italy)
Contributo totale Unibo: Euro (EUR) 88.000,00
Durata del progetto in mesi: 29
Data di inizio
28/09/2023
Data di fine:
28/02/2026
