4 1. IntroductionIn the HMM case, the locations are assumed to be discrete and enumerable. There-fore, the possible combinations of these locations, the probable solutions to thetracking problem, are enumerable, too. Since these are a huge number of possi-ble location sequences, the HMM model has to make assumptions that restrict thesearch space to a tractable size. Only depending on the quality of the assumptionsthe algorithm predicts the most probable location sequence of all possible solutions.The PF makes use of a continuous location space thus avoiding the error inducedby the coarseness of an eventual discretization of the space. Instead of searching forsolutions by enumerating the search space, the algorithm generates solutions thatare elements of an infinite solution space. Whereas the HMM uses the mentionedconditional probabilities to assign probabilities(or scores) to all candidate sequences,the PF generates a subset of all solutions by simulating the progress of the modelledstochastic process by sampling from the conditionals. Due to this properties, thePF technique is a member of the Markov chain Monte Carlo methods.1.3 Framework and ImplementationFor the realization of this thesis, a localization framework was designed and imple-mented to solve the identified problems. The framework handles the training andsimulation of the radio propagation models and uses the results as a foundation forthe application of the localization algorithms. The final results of the localizationalgorithms are then either used to predict the current location with the track history,during the online stage, or are later processed for an offline evaluation.The system is driven by a central Server process that communicates with the pro-ducers and consumers of the different data streams over HTTP service interfaces.A prominent producer is given by the mobile devices that push their collected RSSIreadings for processing at the Server process, and subsequently consume the resultsduring online tracking. Another producer/consumer is the implemented Fat Clientused for visualizing and debugging the localization system with an OpenGL baseddata analysis toolkit which is especially suited for the 3D nature of the simulatedenvironment. The third component, that is interfaced with the Server over HTTP,are the GPU-nodes that are employed for training the free parameters of the radiopropagation models.The mobile devices and the GPU-Nodes are the simplest of these four primary com-ponents, as they are only responsible for high-level I/O. The device with sensorsHT T P and the GPU-Node with HT T P processor HT T P. Each of the twocomponents consist therefore only of around 100 lines of code. On the contrary, theServer is the most complex component and it depends on a number of subsystemsthat are responsible for the different tasks of the localization problem. The mostrelevant subsystems are the Simulator , the Optimizer and the Localizer which aretherefore briefly described.The Simulator is responsible for the organization of the simulation of radio propa-gation models. It responds to requests for propagation models by dispatching theconfiguration of the model parameters in a job enclosure to the available GPU-Nodesand returns a job specific result. Many thousands of these requests are queued by