60 4. DesignFigure 4.11 Example results of a brute force optimizing run with the Evaluator.Empirical values for the emission and transition are determined by looping overcombinations of values for the transition and emission probabilities of the ParticleFilter model. Different error types lead to different optima. The range of =(5, 5)yields acceptable results for both error variants.outliers as localization failures to recognize the transition from one building level toanother or other pathological cases.The results of an evaluation are given in plain HTML-reports for the different possi-ble aggregations of the errors. Such aggregations contain sub-errors over the differentconfigured paths. Results for individual tracked paths are drawn on a 2D-map andcombined with a HTML-report of the algorithmic steps in resolution of a time frame.The results can be afterwards accessed via the HTTP-Server.The Evaluator component was used to empirical fine-tune the different propertiesof localization algorithms. For example, the variance of the emissions and the tran-sitions of the PF were determined by a brute force search over the parameter spaceas visualized in 4.11. A typical efficiently configured evaluation run for the 160 col-lected paths for a single device takes around 2 minutes on current Hardware. Thus,low dimensional brute force searches are feasible.4.7 SummaryThe presented architecture of the localization framework captures all necessary as-pects that are needed for solving the positioning and tracking problem. A cen-tral HTTP-Server, build around loosely coupled components, provides the differ-ent communication channels between the consumers and producers of information.These information sources contain moving Wi-Fi capable devices, GPU-nodes andFat Clients.After allocating a 3D-model of the indoor localization environment with positionconfigured APs, the PHOTON raytracer is employed to build a radio propagation