Practical concerns of implementing machine learning algorithms for W-LAN location fingerprinting

In the past, fingerprinting algorithms have been suggested as a practical and cost-effective means for deploying localisation services. Previous research, however, often assumes an (idealised) laboratory environment rather than a realistic set-up. In our work we analyse challenges occurring from a university environment which is characterised by hundreds of access points deployed and by heterogeneous mobile handsets of unknown technical specifications and quality. Our main emphasis lies on classification results for room detection. We analyse the problems caused by the huge number of access points available and by the heterogenous handsets. We show that standard techniques well-known in machine learning such as feature selection and dimensionality reduction do work. We also provide evidence that pre-processing techniques suggested previously in a laboratory set-up do not improve accuracy.