Learning communication for real-world multi-agent systems

The coverage problem asks the question of how a given area can be covered as efficiently as possible, i.e. under given
time or specific coverage constraints. Applications for this problem range from household robotics over harvesting to
security or monitoring applications over search and rescue. While the single robot coverage problem can be easily
solved analytically, introducing multiple robots in an environment that changes and contains obstacles introduces
challenges that can potentially be solved more efficiently with machine learning approaches. A promising solution to this
problem are Graph Neural Network. By considering multiple robot as a graph and the neighboring robots (i.e. the robots
in reach of the communication signal) a sub-graph, complex emerging behaviour can be learned with reinforcement
learning, i.e. learning by doing, much like evolution in nature. Not having to specify explicit behavioural rules appears
desirable for multi-agent tasks with limited communication and computation capabilities since multiple features can easily
be combined into a single solution just by adapting the reward policy. It can be evaluated if GNNs can be used with a
reinforcement learning paradigm to find a solution to the coverage problem by learning communication as well as
behaviour that outperforms a locally optimal first-principles solution in terms of speed and computation complexity.