Using decision tree algorithms for modelling driver distraction situations.

Abstract

Algorithms based on decision trees are widely used in artificial intelligence research. In this thesis we apply two different of these algorithms to data gathered in a driver distraction study with 25 participants. The majority of studies in cognitive modelling only start gathering data after formulating a hypothesis, based on theory in the domain. We starts with the data however and analyse it to try and find interesting features. The idea behind this approach is to try and discover less obvious patterns. First we extracted 7 features from the original data to get a dataset that we could apply our algorithms to. Our algorithms were able to correctly predict 54% and 57% respectively of the users either focussing on the road or on their phone. We found that inherent properties such as age and gender are not strong predictors of performance in driving. On the other hand looking at the actual driving performance can be used to see if a participant is focussing on the road or their phone. It is important to note for further research that the original data assumed that participants were following instructions to focus on the road or the phone, but did not verify it. To gain better insights from decision tree algorithms it is important that the data is checked to ensure that the data is consistent. For improving the accuracy of the algorithms it would also greatly help to have a larger number of participants. This task had participants look at their phone to fill in a number. For further research it would be useful to look at tasks with distractions in other modalities with the same algorithms.