48 4. Designantenna gain. Thus, the possibility was introduced to adapt or normalize over theseproperties with some function that accepts raw RSSI values as input and results innormalized RSSI values for output. The configuration of this function can be learnedduring the optimization process or alternatively later in a postprocessing step withan optimization algorithm that compares device specific measurements to the radiopropagation results.Since such an adaptation is device specific, a prerequisite for the optimization is theavailability of separate measurements with known locations for each device or at leastfor each different device class that share a similar hardware configuration. Insteadof using manual collected and therefore costly measurements, it can also be possibleto extract these location annotated measurements from the localization results of anaccurately working system. Such a system needs a form of confidence measure fordistinguishing between good and bad localization results and is exemplary for themethod of reinforced learning.If individual Device Specific Adaptation is not used due to the presented limitations,a simple global adaptation of the raytracer results is applied. The simple adaptationis a mapping of simulated RSSI values that are smaller than- 100 dbm into the range- 100 dbm to- 90 dbm since all evaluated devices share- 100 dbm as an upper limitfor reported RSSI readings.4.3 Positioning and TrackingIt has been explained in the background chapter that positioning and tracking canuse the same set of algorithms, since the former is understood as a special case ofthe latter. Both work on a sequence of measurements that collapses to a single mea-surements at the start of a positioning or tracking attempt. Although it is possibleto differentiate between the two cases if additionally sensorial input from the devicescould be delivered, for example in the form of acceleration information3, this wasnot evaluated in this thesis. Therefore, both use-cases employ the same component,referred to as Localizer, see context in figure 4.1, in an identical configuration forsolving the associated localization problem.The core of the localization system employs the chosen algorithmic backend that isgiven either in the form of a Viterbi Decoder for the HMM based approach or thesampling algorithms for the PF. The Viterbi Decoder processes prepared sequencesof RSSI readings by executing the Viterbi Algorithm on a HMM. The emission andtransition probabilities are parametrized with data managed by the Environmentcomponent. The same principles are used to drive and configure the PF algorithms.The RSSI readings for all available APs are collected at each device by queryingthe local Wi-Fi-APIs in a configurable interval and pushing them to the Server byusing the Webservice-API. The Server collects these RSSI readings as a sequence ofmeasurements that are further normalized in a signal preprocessing step(see figure4.4) before fed into the Viterbi Decoder for the HMM model or into the PF decoder.3It would be reasonable to expect that such information could be used for manipulating thetransition probabilities. In the simplest case that means distinguishing between moving and stand-ing.