Methods and apparatuses for training and using a graph neural network

A computer-implemented method of using a graph neural network, GNN, to determine recommended configurations for a plurality of nodes, wherein the GNN comprises the plurality of nodes and a plurality of edges, the method comprising:

• i) for each first node in a candidate set from the plurality of nodes:
• utilizing the GNN to select a selected configuration from a plurality of test configurations for the first node, whilst considering only changes to the configuration of the first node in conjunction with current configurations for the other of the plurality of nodes,
• ii) inputting, in conjunction, the selected configurations for the candidate set into the GNN;
• iii) outputting, from the GNN, for each of the plurality of nodes, first estimates of one or more performance metrics;
• iv) determining, based on the first estimates, one or more improved configurations from the selected configurations that improve a collective estimated performance of the plurality of nodes;
• v) updating the current configurations for the plurality of nodes with the one or more improved configurations;
• and
• vi) repeating steps i) to v) with the current configurations for the plurality of nodes.

A computer-implemented method of using a graph neural network, GNN, to determine configurations of a plurality of nodes, wherein the GNN comprises the plurality of nodes and a plurality of edges, the method comprising:

• inputting, in combination, one or more initial performance metrics and current configurations for the plurality of nodes into the GNN;
• outputting, from the GNN, one or more predicted performance metrics for each of the plurality of nodes;
• and
• for each node in the plurality of nodes:
• calculating a gradient by determining a loss function based on the one or more predicted performance metrics and the one or more initial performance metrics;
• and
• updating the current configuration for the node based on the gradient.