86 6. EvaluationOnline and Offline LEs in m for Noise: = 14 dBm(forward+ backward)PathHMM/off HMM HMM/avg LMSE PF/off PF PF/avgeg-room-changeog1-classicog1-egog1-eg-right1.76 2.34 2.08 4.97 2.51 3.00 2.511.09 1.66 1.41 3.60 2.11 2.46 2.041.60 2.44 2.23 5.00 3.04 3.54 3.221.51 2.27 2.04 5.38 3.03 3.46 3.15og1-long-rooms1.83 2.53 2.33 6.05 3.07 3.67 3.28og1-long-straight1.38 2.07 1.87 4.20 2.14 2.61 2.28og1-room-change1.63 2.05 1.84 3.61 2.45 2.74 2.34stairs-upward1.20 2.31 2.11 5.98 2.69 3.09 2.76Mean in m 1.50 2.21 1.99 4.85 2.63 3.07 2.70Stdev in m 0.25 0.32 0.34 0.91 0.39 0.44 0.47Table 6.6 Online and offline localization errors under = 14 dBm. The differentlydefined paths show significant differences in their error rates.degraded set and restart the sampling at some point in the modified history. Butthe analysis of this approach is reserved for future experiments and has not beencovered in this thesis.Another interesting result is derived from table 6.6. The defined paths, that repre-sents different localization scenarios, show significantlys different error rates. Thesedifference can have the following causes: Either the path is covered with more APswhich leads to more valuable emission probabilities. Or the path is complex withrespect to the number of turns or direction changes which is an unmeasurable eventfor the RSSI only sensors. The same argument holds for an eventual stopping of themovement which was the case in the eg-room-change, og1-long-rooms, and the og1-room-change scenario. It can be assumed, that a more elaborate transition model,which would probably depend on additional sensors, leads to a better adaptation tovariable movement conditions.A more detailed overview of the results on synthetic data is given in the appendixA.3. There, the results for 5 different noise levels are listed with additional for-ward/backward path separation. From comparing the forward and the backwardcase, it can be seen, that they show a significant difference in some cases. A probableexplanation for this observation can be derived from the sensibility to the startingconditions of the HMM and PF algorithm. If the starting point of the path liesin an area with less AP coverage, the initial location hypotheses are distributedwidespread, which can, for example lead to early sample impoverishment in the PFcase. Another reason originates in the handling of the sequential nature of the track-ing problem. The HMM/PF algorithms are able to retain accuracy in zones withless AP coverage by depending on the history of the signals. At an uninformativestarting point such valuable history is not available.It will be interesting to see if these results are also applicable to the case of realworld measurements which are inspected in the next section.