Testing the state trainer and predictor app

Testing the state trainer and predictor app
Our first app used only raw signal data and the Bayes network machine learning algorithm to build the model. We trained the models for three minutes per aisle by walking around within the bounds of the aisle. We then tested the app by walking up and down the aisles and checking which aisle our model placed us in. The predictions were not very accurate using this method. We experimented with some other models such as IBK and random forest. We also tried moving the beacons to different walls and training it by standing in for set periods of time instead of moving around. None of these significantly improved the accuracy of prediction.

Next steps
It was clear that if we wanted our app to function well, we would have to try everything in the hopes that one method will work well. So similar to our last assignment we decided to write an app that could simultaneously test many different models and features (raw, mean, median, RMS). Improving on what we did last time, we implemented a sliding window approach for collecting the mean, median and RMS as well as the batch method used previously. We hope this improves our models by giving us a more complete data set and more data per unit time spent collecting data. We will also make sure to train our models for a much longer period of time than our last assignment, perhaps 10-20 minutes per aisle.