7.1. Future Work 97model can also be trained with the proposed optimization procedure of thisframework.· Furthermore, it would be interesting to analyse the effect of the AP placementon the error rates. More APs lead to better performance of the system, butis there a reachable optimum if the number of APs is limited? This optimalplacement of the APs can probably also be found automatically.The next topics relate to the localization problem and can help to further increasethe accuracy of the presented algorithms:· Lifting the constraints on the knowledge pool and fusing additional sensorinput into the RSSI measurement stream is probably the most promising wayto increase the location accuracy. And, as modern smartphones are more andmore equipped with sensors like gyroscopes, accelerometers and compassesthere exists a rich set of options. It should pose not much effort to adapt theHMM and PF algorithms to these information sources.· The presented HMM model can be extended to process more of the measure-ment history by introducing higher order models. On current hardware, atleast second order models are computable, and will probable lead to a moreadaptable recognition systems. For example, this will enable the possibility tomodel direction changes using the transition probabilities.· The PF algorithm can be extended to handle the degradation of the particleset under bad prediction conditions with a more elaborate technique than thecurrently implemented random resampling. A backtracking based proposalfound in the related literature to the topic seems to be suited for this problem.