Graph Neural Networks For Decentralized Controllers

This paper describes the application of graph convolutional neural networks (GCNNs) and graph recurrent neural networks (GRNNs) to the problem of learning a decentralized controller. Determination of an optimal decentralized controller is challenging since an agent’s action is determined by its local neighborhood rather than all agent states. In the GCNN framework a nonlinear map is learned from the system state to the control action and in the GRNN framework the dynamical evolution of the system is learned. In both settings training is performed by immitation learning. To evaluate the utility of the proposed models, the authors consider flocking behavior.

References

Graph Neural Networks For Decentralized Controllers [pdf] [code]