Modeling Freeway Traffic with Coupled HMMs

We consider the problem of modeling loop detector data colle ted from freeways. The data, which is a vector time-series, contains the peed of vehicles, averaged over a 30 second sampling window, at a num ber of sites along the freeway. We assume the measured speed at each loc tion is generated from a hidden discrete variable, which represe nts the underlying state (e.g., congested or free-flowing) of the traffic a t th t point in space and time. We further assume that the hidden variables o n y depend on their spatial-temporal neighbors. Such a model is called a coupled hidden Markov model (CHMM). We can fit the parameters of this m odel using EM. However, since exact inference is intractable, we consider two different approximation schemes: one based on a sequential Monte Carlo technique (particle filtering), and the other based on the Bo yen-Koller (BK) algorithm. We show that both algorithms perform well, c ompared to exact inference, and that the resulting learned model cap tures many important features of the data. Such a macroscopic model cou ld prove useful for fault diagnosis, and in predicting future traffic patterns, particularly in response to causal interventions.