Perturbed utility route choice models 

What is it for? We want to improve decision-making that involves the design of traffic networks - at any scale from local to global. This is relevant for many issues of general interest including climate, urban liveability, economic growth, traffic congestion, commuting, housing prices, traffic safety, etc. Relevant decisions that could be better informed range from the planning of bicycle networks over urban and national road pricing systems to the development of global transportation networks.

What do we have? Current traffic models are not that different from the models developed by Nobel laureate Dan McFadden in the seventies. They use a paradigm that was feasible at the time, but which is better suited for low-dimensional settings and small datasets. Behavior is seen as a discrete choice among not so many alternatives (e.g., whether to commute by car or public transport). These models are often prohibitively expensive and time-consuming to work with. They require many unpleasant compromises to be made in order to be feasible with real large networks, which distracts from their validity.

What has changed? The availability of data has increased enormously worldwide. We now have tons of data where cars, trucks, planes, ships, trains, parcels, etc. are GPS logged at very high frequencies. This is much more data than we can handle with current methods. We also now have very rich open-access network data covering very many urban areas.

What do we want? We want a new paradigm that can synthesize the vast traffic data, now available in most of the world, into a holistic picture. The paradigm should integrate the underlying complexity of millions of daily interacting movements through large networks. We want to make this available for policy evaluation in a way that is easy and cheap to implement. We want to deal equally well with cities like Copenhagen as with cities in less developed countries.

How can we do it? In the old paradigm, the basic objects are choices from choice sets and corresponding probabilities. I propose to replace these objects with paths through networks and corresponding flows. This matches the data as it is without compromise.

To develop the new paradigm, we go back to the roots of microeconomic theory to exploit convex analysis, the branch of mathematics that underlies the classical microeconomic theory as well as mathematical optimization. Thus, the new paradigm fits with modern tools from machine learning and optimization that enable us to deal with the extremely high dimensionality of real networks.

Does it work? Yes, we have proof of concept.

·  Using 1.3 million car trips in the Greater Copenhagen area, we estimate an unrestricted route choice model with a network with 50k road links. We do this in seconds using plain regression. Under the old paradigm, it would be basically impossible to achieve anything remotely similar. 

·  Using 300 thousand bicycle trips in Copenhagen, we condense the observed behavior of bicyclists in an extremely detailed bicycle network into generic measures of “bikeability“. We compute counterfactuals to assess the impact of the Copenhagen bicycle network on the amount of bicycle use. This paper is published in PNAS.

The way forward? The research agenda is led by applications covering different geographies and types of networks and dealing with issues of general interest related to climate, environment, and economic welfare. The list of things to do is virtually endless. Fundamental issues in microeconomic theory, statistics, and optimization emerge naturally, whereby the research agenda connects all the way from fundamental research to a multitude of applications.


Papers