4.7. Summary 61Model for the expected RSSI values. The free parameters of the model are thentrained with a genetic algorithm that is distributed over an array of GPU-nodes.The RSSI based localization algorithms, the Hidden Markov Model and the Parti-cle Filter, are based on the same basic model assumption given by the state spacemodel of Figure 2.11. The model parameters for the emission probabilities of thetwo algorithms are derived from the trained radio propagation models. Althoughthe transition probabilities for both algorithmic approaches are modelled differently,both rely on the 3D-geometry for restricting invalid movements. Position annotatedmeasurements, for training the model parameters with the classical parameter infer-ence algorithms like the Baum-Welch algorithm in the HMM case, are not assumedto be available.The Wi-Fi capable devices are only dumbed down information providers and haveno higher functionality.The support infrastructure is given by the Fat Client that is used for interactivedebugging of the algorithms and associated data-streams, and finally by the Eval-uator that is utilized to measure the overall performance in terms of error rates ofthe localization framework.