Dataset

The data set can be found at https://archive.ics.uci.edu/ml/datasets/ujiindoorloc. It was already split into train/validation sets. The provided validation set was used as a test set. The training dataset was split into training and validation sets. The split was done based on USERID column so that training set contains 70% of the data and the validation set contains 30%. The resulting validation set was used to tune the hyperparameters of each considered model.

k-Nearest Neighbors

A first simple approach to estimate position from a vector of RSSI values is to use a k-Nearest Neighbors algorithm. We can estimate floor and building by majority vote and longitude and latitude by interpolation.

The contribution of each neighbor can be weighted with a function of its distance to the considered vector. This approach is similar to the WKNN approach described in [2].

	k	Weighting function
Building	5	1/d2
Floor	5	1/d2
Longitude	3	1/d2
Latitude	3	1/d2

Chosen hyperparameters values for the kNN models

Deep learning based methods

Longitude and latitude estimation

Longitude and latitude are jointly estimated using a multilayer perceptron.

Overview of the architecture used for longitude and latitude estimation

Batch normalization was applied to each layer. White Gaussian noise was added to inputs in order to reduce overfitting. Based on the results of experiments done on the validation set, standard deviation of 0.1 was chosen.

Training was done with an initial learning rate of 0.001. The "reduce on plateau" strategy which consists in reducing the learning rate when validation loss stops decreasing was used. The model was trained for 200 epochs.

	Initial learning rate	Learning rate decay	Batch size	Input noise σ
Hyperparameters values	0.001	0.1	128	0.1

Chosen hyperparameters values for longitude/latitude MLP regressor

Building and floor classification

Building and floor classification is done with a simpler architecture, with only one hidden layer.

Overview of the architecture used for building and floor estimation

Training was done with an initial learning rate of 0.0001.

	Initial learning rate	Learning rate decay	Batch size	Input noise σ
Hyperparameters values	0.0001	0.1	128	0.1

Chosen hyperparameters values for building/floor MLP classifier

Results

Longitude and latitude estimation

To evaluate the accuracy of our models on longitude/latitude estimation the mean position error (ME) and root-mean-square error were considered.

	Validation ME	Validation RMSE	Test ME	Test RMSE
k-NN	8.96	15.68	9.26	16.00
MLP	9.86	14.14	9.74	13.66

Results for the longitude/latitude estimation task

Building and floor estimation

	Validation building accuracy	Validation floor accuracy	Test building accuracy	Test floor accuracy
k-NN	0.9732	0.9072	0.9953	0.8920
MLP	0.9984	0.9282	0.9990	0.9298

Results for the building and floor estimation tasks

References

[1] Joaquín Torres-Sospedra, Raúl Montoliu, Adolfo Martínez-Usó, Tomar J. Arnau, Joan P. Avariento, Mauri Benedito-Bordonau, Joaquín Huerta UJIIndoorLoc: A New Multi-building and Multi-floor Database for WLAN Fingerprint-based Indoor Localization Problems In Proceedings of the Fifth International Conference on Indoor Positioning and Indoor Navigation, 2014. https://www.researchgate.net/publication/283894296_UJIIndoorLoc_A_new_multi-building_and_multi-floor_database_for_WLAN_fingerprint-based_indoor_localization_problems

[2] A. Moreira, M. J. Nicolau, F. Meneses and A. Costa Wi-Fi fingerprinting in the real world - RTLS@UM at the EvAAL competition In 2015 International Conference on Indoor Positioning and Indoor Navigation (IPIN) https://www.researchgate.net/publication/308855851_Wi-Fi_fingerprinting_in_the_real_world_-_RTLSUM_at_the_EvAAL_competition