84 6. Evaluationuses an Iconia/Nexus trained model to investigate whether the localization systemgeneralizes over multiple devices. The optimization result of these two models isreported in 6.1.2.2. Furthermore, the distinction between the different devices isonly made during the real world evaluation.Offline/Online ResultsFor both evaluation setups, the three algorithms are compared. The PF and theHMM approach return Offline and Online results. Both adapt to RSSI informationcontained in the history due to their algorithmic nature. The offline result is thepredicted sequence of locations under the knowledge of the full history. The onlineresult is the predicted sequence of locations under the knowledge of the limitedhistory up to a time frame. The latter represents the result that is available for auser of a localization service who expects to be informed about his current locations,and not about the way he came.Therefore, it is reasonable to distinguish between the online and the offline resultand it can be expected, that the offline variant outperforms the online result due toits larger knowledge pool. The LMSE does not use the history of the RSSI values,therefore it reports only an online result.The online error of the PF and the HMM is given in an additional averaged form.This form is defined by using the mean location over the top 50 candidate hypothesisof a time frame t instead of choosing the best hypothesized location.6.2.2 Synthetic MeasurementsIn this section, the performance of the three localization algorithms is evaluated onsynthetically generated RSSI sequences used to simulate readings of a mobile devicewith different noise conditions. For each of the 16 path variants 20 samples weregenerated therefore leading to 320 evaluated path samples. From the lengths ofthe paths given in table 6.4 can be derived that around 8000 m of covered distanceare evaluated. The average movement speed is assumed to be 1 m/s reflecting theproperties of the real world measurements that are concluded from table 6.5. Forevery 0. 75 s, an RSSI vector is generated that represents the target frequency of theRSSI pushing android devices4.The synthetic measurements were generated by using the RSSI values of a trainedradio propagation model as the means of Gaussians with a variable noise . Fromthese Gaussians the measurements are sampled for 0 dBm 18 dBm with astep size of 0. 5 dBm. The results for the algorithmic variants under online andoffline conditions, for a localization space with 40 cm voxel size, can be obtainedfrom figure 6.2 and table 6.6.It can be seen, that the LMSE based algorithm leads to incompetitive error ratesover all noise conditions with respect to the HMM and the PF approach. This wasexpected since it does not use the valuable information contained in the measurement4Although the target frequency was not completely reached in the real world measurements dueto network latency over the HTTP communication channel.