The overarching goal of this project is to combine MARL with control theoretic approaches for the development of new strategies for the control of autonomous multi-agent systems and their application to the problem of designing cooperative agents able to perform challenging SAR operations in uncertain environments. The specific objectives of the research program are summarised below:

[O1] Identification of a set of benchmark MARL strategies for complex multi-agent systems in the current literature and their application to SAR scenarios

This objective will address the need of investigating the control and learning performance of existing solutions in the literature with the aim of comparing their performance using a set of commonly defining metrics so as to allow contrasting the performance of the
methodologies developed within the project with other benchmarks from the literature.

[O2] Development of efficient learning-based control strategies for complex multi-agent systems

This objective is aimed at synthesising innovative strategies for controlling the collective behaviour of cooperative multi-agent systems performing a joint desired task by combining MARL and control theoretic approaches to steer the dynamics of complex systems while rendering more data efficient the learning process. We will explore the design a variety of centralized and execution schemes, by evaluating distributed and centralized solutions. Key aspects that will be considered is credit assignment and reward structures that are critical for efficient and effective deployment of these solutions in a real-world scenario. This will be instrumental for the development of new techniques to solve SAR problems in [O4].

[O3] Convergence and robustness analysis

This objective concerns the investigation of the convergence and robustness properties of each of the strategies developed within the project on a set of multi-agent benchmarks problems. This will allow for the comparison of the methodologies developed in MENTOR with those already existing in the literature, which will be studied as part of [O1].

[O4] Application to search and rescue applications in natural disasters

The focus of this objective is the application of the strategies developed within the project to multi-agent Search and Rescue problems in uncertain environments. We will perform extensive simulations of the proposed algorithms in a data-driven fashion by modelling real-world case studies of interest and carry out preliminary experimental tests using a proof-of-principle case of study.