766.1 Radio Propagation Model6. EvaluationThe evaluation of the framework starts with a detailed explanation of the environ-ment that is used for the experiments. The introduction of the setup starts bydescribing the different properties of the manually created 3D geometry that is uti-lized to drive the PHOTON raytracer. Afterwards, the placement of the differentAP groups and the resulting coverage conditions is analysed. In the next step, thecomposition of the training corpus that drives the optimization process is presented.A relevant number of measurements over 100 locations in the UMIC scene weretaken with four different Android devices. These measurements will represent thelargest part of the training corpus that is employed in the evaluation.After the introductory part, it will be seen how plausible the trained propagationmodels adapt to the true nature of the environment by using the optimization crite-rion, the RPE, as a measure2. An experiment that investigates the training processover measurements of different devices and another one that tries to understandhow important the granularity of the 3D geometry is, will finalize this part of theevaluation.6.1.1 Scene and SetupThe evaluated localization scenario is located in the UMIC Building3of the RWTHAachen. It is a four level building with a rectangular floor-area of around 900 m2.The first three floors were modelled in Blender leading to a volume of around 8100 m3,as the height of a level is around 3 m. It is composed of"reduced to functionality"office architecture and was erected in the year 2009.6.1.1.1 3D-Model and MaterialsThe 3D-model consists of 11 different object classes given in the form of triangularmeshes. The class of an object is defined by the assigned material. The choice ofthe materials was made by reasoning about their importance of their impact on theradio propagation. Probably the most important ones are the Concrete and theLightWalls class of objects. 2D-Maps of the building at the resolution of these twomaterials were available and are used as a base for these two classes of objects in the3D-model. Further modelled detail is given by doors with the three material classesIronDoor, normal Doors and GlassDoor. The limits of the building are modelledwith the two materials Fascade and GlassWindow. Additional furniture details arecaptured by the materials Cupboard and Table. Also included for completeness is themetal based railing of the stairs and hardware components that are located next tosome of the APs or components that exhibit a massive structure, like the hardwarearray of the server room.2This can lead to a serious overfitting of the model, but since the training corpus had only asize of 100 locations it has been concluded that they are to valuable to sacrifice some of them toallow the better cross-validation technique.3Google Maps reference: http://g.co/maps/q3mp6