Multi-agent reinforcement learning: coordination through abstractions, trust and world models

This thesis explores advancements in cooperative multi-agent reinforcement learning (RL), with a focus on coordination, sample efficiency, and trust in cooperative systems. A significant contribution of this work is the formalization of trust within multi-agent RL, which is pivotal to establishing cooperativeness. Trust, a key determinant of cooperative behavior, is systematically examined through a set of influencing factors. These factors are analyzed to understand how trust shapes agent interactions and supports robust collaboration in dynamic environments. Additionally, this thesis introduces a novel approach to combine the strengths of traditional tabular solutions and Deep RL solutions by constructing discrete abstractions of continuous environments. The use of abstraction allows guiding the learning process in the layers below in the hierarchy, which is particularly useful in the case of environments with very sparse rewards. The solutions are tested on one of the most prominent applications within the RL domain, which is cooperative multi-UAV systems. Central to this work is the integration of model-based RL techniques, utilizing world models to enable agents to reason about future outcomes. By leveraging these learned representations, agents can anticipate the intentions of others, facilitating consensus-building and collective decision-making. The effectiveness of these approaches is demonstrated empirically in different scenarios.