This approach yields device-level descriptions of the input topology that, for each device, focus on describing its relation to every other device in the topology. By means of thesedescriptions, an unfamiliar overall topology can be broken down into devices that are subject to the same constraints as a device in one of the training topologies.