78 6. EvaluationFigure 6.1 Access Point coverage of the building after placing the moveable APsfor the training procedures. The histogram is extracted from raytracer simulations.The total volume of the modelled UMIC building is given by 8100 m3. The averageAP density at a location in space, from at most 22 APs, is given by 7. 6 APs pervoxel.6.1.1.3 DevicesMultiple devices were evaluated by taking measurements at locations with differ-ent ratios of completeness with respect to 100 defined locations. Investigating theperformance of the framework over more than one device leads to a better under-standing about the generalization capabilities of the system. This means, it can beseen if the trained propagation model adapts only to device-specific and not to theenvironment-specific properties, which are the more interesting here.The primary two Wi-Fi capable devices were an Acer Iconia Android tablet anda Nexus 4 smartphone. Some experiments were also conducted with the WISPYspectrum analyzer. The remaining two devices are again Android smartphones.One is a Galaxy P and one a Nexus 1. And the last device was an external TP-LinkUSB-Adapter connected to a linux driven netbook. As can be seen in the next table,the first two android devices have been analyzed more thoroughly.DeviceIconiaNexusGalaxypNexus1MessbookWispyType API NlocsNapsNallNrssisTablet Android 3.2 86 22 924 25715Phone Android 4.0 94 22 919 20428Phone Android 2.3 68 22 496 7644Phone Android 2.3 73 22 515 8055USB-Adapter Libpcap 44 20 492 4478Spectrum Analyzer Custom 19 3 41 41The different numbers represent the following values: Nlocsis the number of vis-ited locations. Napsthe number of unique APs for which measurements have beencollected. Nallis the number of(location, AP) tuples with measured RSSI values.Some combinations have no values due to their distance. And finally Nrssiis thetotal number of RSSI readings that were taken at for the Nallcombinations.The measurements for the Android devices were taken by interfacing with the OS-level APIs. For the linux netbook a Python wrapper over libpcap was used.