Abstract:
This paper is devoted to identifying precursors of freeway breakdown: decreased speed and ow as the result of excess demand. We do so by formulating prediction problems wherein we try to predict whether a breakdown will begin in the next five or ten minutes. We use tree-structured statistical models to select conditions that are likely to precede breakdowns. The data we use are from single loop detectors on Interstate 5 in the Seattle area, and were available essentially in real time, so results such as ours could be useful to indicate that anti-breakdown countermeasures are called for. We also discuss techniques for visualizing the breakdown process which we found critical in our research, and the extensive process by which researchers must clean up detector data of this form before performing data analyses.
tr79.pdf