92 6. Evaluationof section 6.1.2.1. It can therefore be concluded, that modelling the environmentwith one material leads to an insufficient propagation model with respect to thelocalization problem. Even increasing the complexity of the scene geometry doesnot leads to improvements. On the contrary, the Full1 propagation model performsless accurate than the simpler Basic1 model.The Basic2 propagation model performs very good with regard to its simple meshstructure. Only the concrete and light walls are employed in the 3D geometry whichis a relatively easy to obtain information about a building. The model leads tobetter error rates over all algorithms compared to the Full2 model. The Full2 canbe understood as an upgraded variant of the geometry from Basic2. They share thescene objects composed of concrete material, but the Full2 combines all remainingavailable triangles to the secondary material, whereas the Basic2 only uses the lightwalls for that. Concluding from the error rates, this seems to be a bad approach.It is an interesting result, that the Basic+Doors5 geometry outperforms the Full11model on the HMM/PF algorithms quite significantly and leads to the overall bestresults. The additional knowledge that is included in the most complex model Full11seems not to be an exploitable information source for the raytracer under the setupof this thesis.6.3 SummaryFrom the overall results presented in this chapter, it can be summarized that the de-signed framework is able to solve the localization problem with acceptable accuracyfor PHOTON driven radio propagation models. The proposed method of trainingthe free model parameters with genetic algorithms leads to material and AP spe-cific coefficients that are usable to generate RSSI estimates that diverge only about4 dBm from the true reality represented by the training corpus. By further devicespecific fine tuning with the presented Device Adaptation scheme, the divergencecan be reduced to 3. 5 dBm.The analysis of the performance of the three different localization algorithms hasbrought up the conclusion that the HMM approach outperforms the PF approachover most scenarios. On synthetic data, the gap between the error rates of bothalgorithms is more significant than under real world conditions. On an assumednoise level of 14 dBm, the offline error of the HMM approach is given by a promising1. 5 m whereas the Particle Filter reaches only an offline accuracy of 2. 6 m. As wasexpected from its algorithmic properties, the simple LMSE approach leads not tocompetitive results on synthetic data, as it reports a localization error of around 5 m.The relations between the error rates of the algorithms shift quite significantly ifthe evaluation is conducted on real world measurements. The HMM and the Par-ticle Filter algorithm perform now very similar although the HMM shows a visibleadvantage under offline conditions. The average error rates over all defined pathsare in the range of 2- 3 m for most evaluated scenarios. It was observable, that theerror rates between paths of different complexity show a high variability. Especiallythe stairway path can only be inferred with an accuracy of less than 4 m. Removingthis path and another outlier shifts the offline/online error rates down to 1. 5 m/ 2. 0 mover the remaining six easier scenarios.