6.2. Localization 89Patheg-room-changeog1-classicog1-egog1-eg-rightog1-long-roomsog1-long-straightog1-room-changestairs-upwardMean in mStdev in mHMM/off HMM HMM/avg LMSE PF/off PF PF/avg1.61 2.50 2.27 2.24 1.41 2.04 1.991.44 2.30 2.11 2.31 1.86 2.25 2.102.23 3.87 3.70 3.77 3.03 3.62 3.332.06 2.73 2.48 3.05 2.88 3.02 2.741.80 3.03 2.92 3.45 2.46 2.86 2.641.94 3.48 3.37 2.77 2.21 2.89 2.901.87 2.13 1.81 2.35 2.13 2.37 2.143.07 4.75 4.31 6.02 3.78 4.92 3.902.00 3.10 2.87 3.24 2.47 3.00 2.720.48 0.94 0.91 1.18 0.73 0.90 0.66Table 6.8 Localization error at the individual path level. The forward and backwardvariants of the paths have been combined.the relative differences between the defined path scenarios remain the same as canbe seen by comparing table 6.8 and the results without the adaptation step in table6.7.The observed positive result, the effect of the adaptation function can either beinterpreted as a better understanding of the true nature of the propagation modelor it can originate from a device specific property. In the former case, it can beunderstood as an additional training step which follows the basis training describedin 6.1.2. But the latter case, the interpretation of the effect as an unknown devicespecific property, was the initial assumption that had motivated the approach duringthis thesis.The main problem of this technique is the need for device and location annotatedRSSI measurements. This problem can be approached by investigating whether theneeded information can be obtained by using localization results as a"trustworthy"source. If this"trustworthyness" is measurable, only results with a high accuracyhave to be fed back into the system which would probably lead to a better perfor-mance.6.2.3.2 Multiple DevicesIn this part of the evaluation, the performance of the system over two devices isinvestigated. In the first setting, a radio propagation model is used that was trainedwith measurements of both devices. For each device, 160 paths from the evaluationcorpus are evaluted. The results for the HMM and the PF algorithm are shown infigure 6.5.The average error of the predictions for the Nexus device is comparable to the Iconiadevice although the latter performs slightly better. The variability of the results forthe Nexus is visibly higher than that of the Iconia device. An evidence for thisresult can be obtained by inspecting the radio propagation training corpus. As canbe seen in table 6.2, the standard deviation of the measurements for the Nexus issignificantly larger than the standard deviation of the Iconia device. A cause for