42 3. Related Workfor the positioning problem. The chosen scene for the experiments was a 130 m2small 3-room floor with three APs. The number of positioning attempts that wereevaluated is not documented.Another framework, that is presented in[13], generalizes the LMSE selection crite-rion to a clustering based approach. Of the two known clustering strategies k-meansclustering and hierarchical clustering, the former is reported to lead to better results.The unknown initialization variable of the k-means strategy is given by the numberof target clusters. This number was determined empirically. Furthermore, the clus-ter history is exploited without formalizing the model to the state space model ofFigure 2.11. The basic idea is given by rejecting clusters at the time frame t if thebest clusters of t- 1 are too far away. The reported averaged LE is given by 2. 8 m.The contained tracking error is probably an online 2D-error with history constraints.In[9] a simple localizer, based on the LMSE criterion, is used to test the qualitythe radio propagation model that is generated by the sophisticated 3D-Raytracerdescribed in section 3.1.2 of this chapter. The evaluation of the localization algorithmis performed by estimating the positions of 20 RSSI vectors that are collected foreach one of 11 predefined locations. No dependency between the 20 readings ofa single location is assumed. A pretty good LE of 1. 61 m is reported, although thesetting is only a smallish single-room scene of 80 m2with nearly no NLOS conditions.3.3 SummaryThe following observations were made during the evaluation of the presented litera-ture. There is no defined set of rules for determining the quality of radio propagationmodels or for comparing the performance of localization algorithms. Various errormeasures are in use and sometimes the evaluation is conducted on unclear environ-ment conditions. If needed, the presented errors have therefore been normalized tothe measures of this thesis, described in 2.1.3 and 2.3.2, for easier comparison. Itwould be helpful, if the research community could agree on some common set ofprinciples that should be mandatory for the evaluation of both topics. For example,it should be mandatory to report the number of used APs in an evaluation scene forRSSI based localization algorithms. That was not always the case.