6.1. Radio Propagation Model 81A total of Numphotons=5 · 105photons are used during raytracing, which has beenshown to give a nice balance between accuracy and computing time. The Densityparameter represents the variance of the Gaussian, used for the final smoothing step.The AP Power value of 10- 5indicates, that the genetic algorithm searches over therange[10- 6.. 10- 4] for the antenna gain PHOTON parameters of the 4 different AP power groups. The Resolution in combination with the Bounding Box determinesthe size and the voxelization ratio of the scene. With the given numbers, a voxelhas a homogeneous edge length of 20 cm.The Genetic Parameters define the initial size of the population with the Startpopparameter. 80 organisms lead to 1760 GPU-node driven jobs for the 22 APs of theUMIC scene. In each generation Childcount=80 new organisms are created by ran-domly chosen parents. The fittest Childcull=30 children determine the populationof the next generation. The Mutprob and the Mutamt parameters control, how oftenmutations happen( Mutprob) and how strong the induced variability is( Mutamt).These chosen values have been found to lead to good convergence for the UMICscenario under the constraints of 50 GPU nodes and around 10 hours cluster usagefor each experiment.6.1.2.1 Granularity of the 3D geometryAn interesting evaluation can be conducted by choosing different levels of granular-ity for the 3D-geometry that is used in the raytracer. Lesser needed detail for the3D-geometry is translated in lesser model acquisition costs, which is therefore worthinvestigating. For this evaluation, only the measurements of the Iconia device areused to minimize eventual device specific distortions. The default parameter set forthe optimizer was used and combined with different 3D-geometries. The Basic ge-ometry consists of the mesh triangles for the Concrete and LightWalls materials thatare listed in table 6.1. The 3D geometry for these 2 material classes is probably thecheapest to obtain, as the basic information is available in 2D form for most build-ings. From this 2D map, the composition of a 3D scene can be approximated, whichhas for example been done in the work of El-Kafrawy[8]. Under the assumption,that the doors are also available on an eventual 2D map, the Basic+Doors setupwas chosen. The Full setup represents the geometry with all materials of table 6.1.In the following table, the optimization results in form of the RPE are given:MeshMaterialsRPE in dBmBasic1 Full1 Basic2 Full2 Basic+Doors5 Full1111225.7 5.8 4.4 4.154.2114.3It can bee seen in the first and the second column, that the reduction to one freematerial parameter leads to worse convergence of the training process.After distinguishing between the two materials, Concrete and LightWalls in theBasic2 scenario and Concrete with the combined material Other in Full2, the RPEdrops significantly. The 5-material setup remains in the same RPE range as doesthe full-detail model of the last column. In the corresponding section 6.2.3.3 for thelocalization evaluation, it is investigated whether these results have an impact onthe performance of the localization algorithms.