A neural network approach to constructing commercial broadcast traffic advisories

This work presents a method for constructing an archive of broadcast radio traffic report content from Web advisories using neural networks. Broadcast traffic reports are free and widely used as a source of traveler information. However, there has been no study done to establish what impacts, if any, these traffic reports have in terms of improving listener travel reliability. We developed an analytical technique to quantify travel reliability impacts and conducted a preliminary case study for the Washington, DC, metropolitan area, using radio traffic reports recorded from a local radio station and manually coded for 37 weekdays. However, as coding of radio traffic reports is highly labor-intensive, we used neural networks to construct a database of radio traffic advisories from an existing archive of Web traffic advisories. This paper presents the model developed using feed-forward neural network with back propagation of error that can, given a list of Web advisories, predict roadway segments that would also have an advisory mentioned on the radio. The overall accuracy during the morning peak period was 72%, implying that a commuter listening to constructed advisories would have a 72% chance of listening to an actual advisory mentioned on the radio. During the afternoon peak period, the accuracy was 78%. The missed prediction rates in the morning and afternoon peak periods were 28% and 23%, respectively. Given that we can construct a full year of radio traffic advisories we are able to conduct a more representative study for a longer period of time since traffic conditions on 37 weekdays cannot be used to generalize typical trip experiences of a commuter. Thus, neural networks proved to be a viable low-cost approach to solve the problem of lack of data.