Katharina Parry

Massey University

Traffic network-based modelling

Network-based models of traffic systems provide vital insights for road planning and management. The focus here is on MCMC inference for models of networks with a tree structure based on traffic count data observed on a set of network links.

Many of the model parameters are concerned with travellers' route choices, for which inference would be straightforward if we were to observe path flows. It is therefore natural to sample these latent variables within the MCMC algorithm. A critical problem is that there are very large numbers of possible route flows that could have led to an given observed set of link counts. It becomes essential to find a suitable sampling scheme which avoids the evaluation of the full set of feasible route flows.

We present a solution using first-order Markov model of traveller behaviour to generate candidate route flows in a Metropolis-Hastings sampler.