In this thesis, we argue for the need to expand graph neural network models to more complex graph types. We benchmark common GNN architectures on simple graphs. Then, we introduce local and hard pooling layer to modify graph topology. We apply GNNs to two tasks: Vehicle behaviour prediction on the highway (introducing edge features) and high-voltage power grid control (introducing more complex edge features and heterogeneous nodes). We use the latter's output to warm-start a classical optimizer.      «

In this thesis, we argue for the need to expand graph neural network models to more complex graph types. We benchmark common GNN architectures on simple graphs. Then, we introduce local and hard pooling layer to modify graph topology. We apply GNNs to two tasks: Vehicle behaviour prediction on the highway (introducing edge features) and high-voltage power grid control (introducing more complex edge features and heterogeneous nodes). We use the latter's output to warm-start a classical optimizer....     »