96 7. Conclusion1- 2 m, whereas a localization in the difficult zones lead to significant errors of 4 mfor the PF and of 6 m for the HMM algorithm.Du to the history-awareness of the PF and the HMM algorithms, the reported errorrates are distinguished in the online and the offline variant. The online error relatesto the localization result that is available under knowledge about the precedingmeasurements, and the offline error relates to the prediction of the location underthe knowledge about all following measurements as well. This distinction has notbeen made in the reported literature, although this is a relevant information asboth error variants differ by about 0. 4 m accuracy. Whereas the use-case of theonline variant represents a singular localization request under live conditions, theoffline variant is, for example, more valuable for analysing the movement patternsfor a group of people. The offline prediction is therefore useful to learn from theenvironment, whereas the online variant is for use-cases like user navigation.With respect to the real world evaluation results, the HMM algorithm outperformsthe PF based approach on the same available knowledge about the environmentunder offline conditions, whereas under online conditions, the PF leads to slightlybetter estimates. Furthermore, the PF approach is less computational demandingthan the HMM approach which, for example, makes it more portable to smaller andless powerfull devices.The presented localization framework has shown to be a viable platform to efficientlyevaluate the localization algorithms on the information derived from the raytracinggenerated radio propagation models. The training procedure, devised for the un-known raytracer material parameters in form of a genetic algorithm leads to wellbehaving RSSI predictions that are usable as a foundation for the localization algo-rithms.7.1 Future WorkThis section presents a brief collection of future research topics that were identifiedduring this thesis with respect to the two main topics radio propagation models andWi-Fi localization algorithms.In order to achieve more accurate propagation models using raytracing generatedRSSI predictions the following open topics remain:· The variance of the RSSI predictions, if modelled by Gaussian as in this thesis,originates partly in the multipath effects of the radio propagation behaviour.The PHOTON raytracing algorithms can be configured to report the signaldistribution for different recursion levels of its tracing algorithms. It would beinteresting if this can be interpretable as the described multipath effect andtherefore can be used as a further source of information for the localizationalgorithms in terms of signal variance for the individual locations.· Another feature of the raytracer are antenna patterns that are able providea more realistic signal distribution in contrast to the simple isotropic modelthat was used for this thesis. The additional free parameters of the antenna