1.4. Evaluation 5the Optimizer during the training of the model parameters. The Optimizer imple-ments the genetic algorithm approach to the optimization problem. The resultingoptimized propagation models are employed by the Localizer component which usesthe stored RSSI values as the primary information source for solving the localizationproblem. Under the assumptions of this thesis, the localization problem is givenby sequence of RSSI readings arriving from the mobile device. These readings aresubsequently processed by the Localizer through the application of a localizationalgorithm. Three different algorithms are available, the HMM, the PF and LMSEimplementation.The implementation of the framework is based on the Python programming lan-guage. Since Python is very popular in the research communities it has a wealth ofthird-party libraries that are suited to support the scientific topic of this thesis. Thecore implementation of the localization algorithms is written in a dialect of Python called Cython[3]. This was necessary due to the slow6Python runtime with respectto needs of number crunching algorithms. Cython is a Python -to-C compiler, whichenables the prototyping of algorithms in native Python followed by a transformationinto an efficient C representation. The transformation is supported by providingtype annotations and using dedicated data structures in the form of NumPy arrays.These multi-dimensional NumPy array types provide the basis data structures forthe radio propagation models and the localization algorithms.These design decisions have lead to a flexible framework that can be easily extendedif needed. For example, switching from the PHOTON raytracer to another sim-ulator for radio propagation can be accomplished by simply adapting the currentPHOTON -specific driver script and ensuring a similar 3D voxel representation ofthe simulated signal strengths. Furthermore, using the Python/Cython/NumPystack has lead to fast and memory efficient localization algorithms which has madethe evaluation of the system convenient.1.4 EvaluationThe last part of this thesis is dedicated to the evaluation of the designed, imple-mented and now presented localization framework. After a thorough description ofthe conditions under which the experiments of the evaluation were conducted, thefirst steps of the evaluation will investigate the quality of the PHOTON generatedradio propagation models.It will be analysed whether the proposed training process with genetic algorithmsleads to propagation models that can adapt to multiple device classes. The otherobjective of this part of the evaluation is given by the question how much granularityon the 3D geometry level is needed for the PHOTON raytracer to produce propaga-tion models that represent a good estimate of the unknown true signal distribution.For these tasks that relate to search for the unknown material parameters as trainingcorpus is needed for the optimization algorithm. Such a corpus was collected for 4different devices with RSSI readings from 100 locations of the UMIC building.6Actually, Python is quite fast for most of the common use-cases in software engineering. There-fore, the ratio of Python/Cython code over the implementation of the framework is about 10:1.