72 5. ImplementationLike theHMMLocalizer(), thePFLocalizer()decoder returns three lists of posi-tions for the offline, online and averaged online result variant.LMSEThe most simple algorithm, is the LMSE based approach which is implemented intheLMSELocalizer()class and the correspondingLMSE()Cython decoder class.The Python side computes only the ndarray for the emission probabilities for thevoxelized RSSI values for each AP. This ndarray initializes theLMSE()class that isafterwards instrumented via itsdecode(ms)function to evaluate the measurementsequence of the input signals.To efficiently keep track of the top N solutions for the nearest neighbour distancecomputations that are executed over the total localization space, the already in theHMM and the PF employed SkipList structure is used. The top N solutions can beunderstood as a simple clustering of the best locations and has for example in theARIADNE [13] system been enhanced to a k-means clustering model.5.2.3 Fat ClientThe Fat client is the primary debugging tool for the behaviour of the radio prop-agation models and the algorithmic progress of the localization algorithms. It isbuild around the Traits-UI GUI toolkit that conveniently embeds the MayaVi 3Dengine for scalar data visualization of the localization space. The location relatedscalar data has to be given in the form of ndarrays that are already used all over theimplementation of the framework. Most transformations of the algorithm specificndarrays to a MayaVi compatible structure is done by the native functionality ofthe NumPy/SciPy libraries. If the solution could only be awkwardly implementedwith NumPy block operations, Cython functions have been implemented for betterreadability and flexibility.The GUI needs access to the described Server infrastructure that is defined by itsHTTP address. It uses the Server for initialization to a specific 3D geometry andthe corresponding AP placement. Furthermore, it can investigate the training cor-pus for the radio propagation models and the evalution corpus for the localizationalgorithms. The PHOTON generated RSSI predictions can be visualized with theMayaVi engine after beeing processed by the GPU-nodes and subsequently deliveredby the Server. Due to the GPU-driven nature of the PHOTON raytracer, this designchoice makes sense as no special hardware and CUDA driver setup is needed on theclient side.As mentioned, the evaluation corpus for the localization algorithms are also fetchedfrom the Server over HTTP and can be used to investigate the results of the al-gorithms. The algorithms are executed locally as the Fat Client has access to thecode of the corresponding modules from the Server infrastructure. The differentLo-calizer()classes can be instantiated withEnvironment()conditions that can bemanipulated from the GUI. The algorithmic progress from timeframe to timeframeof the top N location hypotheses can be visualized for the HMM, PF and LMSEimplementations. This is very helpful to find weaknesses in algorithmic properties.