Decentralised Coordination and Communication in Multi-Agent Reinforcement Learning Systems

Coordination of actions plays a crucial role in multi-agent systems, as it allows entities to work in a shared environment, together towards a common goal, or individually without hindering each other's progress. In order for this to occur, agents must demonstrate high levels of spatial awareness and collaborative skills that enable them to understand and acknowledge each other's intentions. Added to this challenge are all those constraints related to real-world implementation, such as decentralisation of information and efficiency requirements that cannot be easily ignored. This thesis aims to contribute to the field of research by studying coordination among agents habilitated to exchange information. Existing challenges and solutions are discussed, then an alternative approach is presented to address the problems. Specifically, the paper argues that explicitly allowing agents to choose whether to coordinate with others or to act independently provides them with adaptability to different scenarios while still ensuring an optimal understanding when needed. To support this claim, CoMix is presented as a novel method that reflects this strategy. Extensive tests with a focus on the scalability of the solution show its positive results in different scenarios, and a comparative analysis highlights the ability of agents to learn strategic behaviour.