82 6. Evaluation6.1.2.2 Multiple DevicesIn this experimental setup, three propagation models were trained by using measure-ments from the training corpus for different device combinations. The first propa-gation model was only trained on the Iconia tablet, the second model includes themeasurements of the Nexus smartphone and the last model uses measurements fromall four Android based devices. The reached RPE under the parameter optimizationare reported in the following table:Device Iconia Iconia/Nexus Iconia/Nexus/Galaxyp/Nexus1RPE in dBm 4.3 3.6 5.0Although the combination of the four devices in the last column leads to a highRPE of around 5 dBm, the theory, that more employed devices lead to worse prop-agation models is contradicted by the first two results. The combination of theIconia/Nexus measurements exhibits an even better converging optimization pro-cess than the Iconia-only case. It also shows satisfying results with respect to theevaluation of the localization algorithms. Such an evaluation is conducted for theIconia/Nexus model in the following section 6.2.3.2.6.2 LocalizationThis section presents the performance evaluation of the three presented localizationalgorithms. The HMM and the PF are compared with respect to the baseline giventhrough the LMSE approach. After describing the properties of the evaluation siteand the different evaluated localization paths, the algorithms are initially tested onsynthetic data. This will allow to compare them under idealized conditions withrespect to errors originating in the radio propagation model. The required artificialsequences of RSSI vectors are therefore obtained from the SSMs generated by thePHOTON raytracer.After the synthetic evaluation, the performance of the localization framework on realworld data will be analysed. Experiments with two different devices classes, a tabletand a smartphone, were conducted. The real world experiments are composed ofmultiple localizations paths, that were manually tracked and annotated to establishthe evaluation corpus for the localization problem.6.2.1 Scene and SetupThe evaluation was conducted in the UMIC building, described in 6.1.1, over twoandroid devices, the Iconia tablet and a Galaxy Nexus smartphone. Eight pathsthat lead through the building with various degrees of complexities were defined andsubsequently used for taking the measurement streams. Each path has a forwardand a backward variant which effectively doubles the number of tracked paths tosixteen. The tables 6.4 and 6.5 show the different properties of the defined paths indetail.