88 6. EvaluationPatheg-room-changeog1-classicog1-egog1-eg-rightog1-long-roomsog1-long-straightog1-room-changestairs-upwardstairs-upward-rMean in mStdev in mHMM/off HMM HMM/avg LMSE PF/off PF PF/avg1.71 2.91 2.66 2.31 1.56 2.25 2.161.46 2.85 2.60 2.65 2.13 2.59 2.463.04 4.35 4.13 4.21 3.18 3.89 3.591.85 2.83 2.52 3.19 2.62 2.90 2.562.02 3.54 3.40 3.86 2.82 3.08 2.932.94 4.00 3.84 3.32 2.62 3.28 3.322.58 2.66 2.36 2.59 2.54 2.71 2.543.17 3.49 3.22 5.82 3.50 4.35 3.583.55 5.53 5.08 6.48 4.42 5.59 4.432.37 3.46 3.21 3.54 2.68 3.21 2.950.68 0.87 0.87 1.17 0.69 0.85 0.66Table 6.7 Detailed errors for the different paths from the evaluation corpus. Theaggregation of the results can be seen in figure 6.3. The stairs-upward-r path, thatis addionally shown in the forward/backward variant, is a drastic negative outlier,probably due to complex nature of the stairway scene and the suboptimal placementsof APs.Figure 6.4 Localization error over all algorithms for the Iconia device after apply-ing the Device Adaptation technique. The error rates for the PF and the LMSEalgorithm drop by around 20 cm.6.2.3.1 Device AdaptationThe evaluation is continued by investigating the Device Adaptation method de-scribed in the design chapter 4.2.3. This method assumes, that the reported RSSIvalues from the device are distorted with respect to some unknown function. Theconfiguration of this function can be learned from the radio propagation models iflocation annotated measurements are available for the individual model of the deviceor for the more general device class.Such a function was trained for the previously described and evaluated Iconia setup.As the mentioned location annotated measurements, the training corpus for the radiopropagation model was used. The comparison of the results in figure 6.4 with theprevious results in figure 6.3 reveals that the procedure has a measurable effect onthe error rates. For the PF and the LMSE algorithm, the reduction is about 20 cm.The error rates of the HMM approach are reduced by about 40 cm. Not surprisingly,