4.2. Radio Propagation 47Figure 4.2 Schematics for decentralized parameter training on cluster. The mainloop of an optimization iteration is emphasized with fat green lines and consists ofinteractions between the Optimizer, the Simulator and the GPU-nodes on the cluster.Attached is also the data flow for the initialization phase when measurements ofdifferent devices at different locations are collected in an indoor scenario.completes overnight instead after multiple weeks. Without these resources, onewould have to sacrifice either resolution at the raytracer level by decreasing thenumber of simulated rays or reduce the number of included APs. It can be expected,that the prepared infrastructure will also be useful during training of the localizationmodels2.With the described mechanics, evaluating a population of organisms leads to spawn-ing of Npop× Napsjobs that are pulled by scripted worker GPU-nodes, located in thecluster, over HTTP. A job is defined by a set of material parameters, a set of powerparameters for each AP class, a 3D-model and a list of locations in the 3D-scene.The worker node executes the raytracer with the given input data and evaluates theresulting SSM at the defined locations. The list of RSSI values of each location ispushed back as a result for a given job-ID to the optimizer. If all jobs, the resultsof the individual APs, of an organism are available to the Optimizer, the fitness ofthe organism is calculated by using the RPE as a measure. The Optimizer tracksthe organism with the best fitness, so that if a new optimum is reached, the systemstores the parameter set for later use. Convergence is assumed, if either a predefinedgeneration count is reached or the change of fitness between two successive optimais less than 1 percent of the absolute fitness value.4.2.3 Device Specific AdaptationAlthough the scale of the measured RSSI readings over the different devices can beinterpreted as a dBm value which represents the dampening of the emitted signal,these readings are not normalized over device specific properties like deviations in the2But as the computational requirements of the tracking algorithms can be handled by a modernmulticore CPU and no real big training corpus of tracked measurements was available this has notbeen fully explored.