Modelling: What has changed in the past ten years

How will abundant new data allow us to explore travel behaviour in more depth? Can we find find correlations that previously were invisible? As correlation is something quite distinct from causation, will correlations (without understanding) suffice, for example in very short term predictions when we just cannot afford to understand before we act?

Tom van Vuren has been the Chairman of Modelling World since its inception in 2006. He started work in transport modelling in the mid-eighties, when he worked at the Universities of Delft and Leeds. For the past 25 years he has worked in consultancy, and since 1999 has been a Divisional Director at Mott MacDonald. He gives his views on the future of models...

http://landor.co.uk/modellingworld/2015/home.php    

What has changed in the past ten years?

On the surface, not a lot has changed in the past ten years.  I don’t think that either the techniques we use today or the supporting software has evolved more than piecemeal.  But I think that the development and sometimes grudging acceptance of WebTAG has strengthened the profession.  Going back to the Modelling World debate a few years back, modelling had become too much of an art without the discipline of science. Despite what some of its detractors say, I think WebTAG allows sufficient space for artists as well, although I prefer to think of us as artisans.

The past ten years have thrown us a few future uncertainties that we have not had to deal with previously. The peak car debate is interesting. For me, the observation that car ownership and usage has so much reduced in young males shook our beliefs that behaviour that we observed in the present can be transferred into the future, a fundamental assumption in our traditional models.  It means two things: first we need to consider the segmentation in our models, and the minimum WebTAG segmentation, which ignores age and gender, will just not cut it. Then, it is inevitable that we need to explore alternative futures; for example one where the recently observed behavioural change is permanent and one where this phenomenon turns out to be just a temporary blip.

Such scenario testing is also going to be necessary for the advent of alternative modes such as autonomous vehicles or shared mobility services such as car clubs. And what about the possibilities of cycling? We are only starting to conceptualise what may happen to mode choice and network operations. I just cannot see how we can afford not to look at scenarios and use them to determine resilient transport decisions.

What needs to change in the next ten years?

Everyone has high hopes for Big Data and I would be very disappointed if in the next ten years we don’t manage to establish robust ways in which to build our models from passively collected data sources such as mobile phones.  But I expect more. These abundant new data should allow us to explore travel behaviour in much more depth than traditional surveys ever could, and hopefully find correlations that previously were invisible. Yes, I mean correlations! And I know that correlation is something quite distinct from causation, but sometimes correlations (without understanding) may suffice, for example in very short term predictions when we just cannot afford to understand before we act.  I am particularly excited about the possibility of unexpected correlations that may lead to new insights, hypotheses that can be properly tested for causality and ultimately lead to new model forms.

What needs to change, but where I see the two worlds (if anything) diverging, is for academic research to feed faster into practice. This has been a bug bear for me for a long time, and in the area of Big Data this is even more important than before. Loads of practitioners are active in the field, but not necessarily armoured with the full set of weaponry required to acquire, store, manipulate and analyse the data efficiently and often hampered by timescales and the pressure to succeed. There are some amazing facilities out there in academia, such as the Leeds Institute for Data Analytics or the Urban Big Data centre in Glasgow, with much better IT and, most importantly, much better skills, but I cannot give you any examples where these have been used for tangible practical benefit.

What is the greatest challenge in modelling?

It is sad, but I am convinced that we still don’t fully understand how people make their travel decisions. We do a more than adequate job by representing them as rational, by assuming they know about their alternative choices, by segmenting them on the basis of income or car ownership or family status. But reality is quite a lot more complex. And when we start to understand about this complexity, we inevitably have to simplify behaviour back to a level that can be accommodated by our current modelling framework, either because it’s so much more convenient or because the implications for run times.

That is our second challenge. When I started my career, a typical model took an overnight to run. These days, 30 years later, we are lucky if they run as fast as that. Any computational speed-ups have been more than counterbalanced by increased model size and further complexity. For us as modellers to contribute meaningfully to debate and decision-making our models must become faster rather than more complex. I would guess that for scenario-testing converged model runs should take no more than 2 hours.

What has been the greatest change in your career?

I started work in 1985.  Network modelling was done on mainframes. Equilibrium modelling was in its infancy – many models were all-or-nothing. Virtually every model was car-only. Outputs were large reams of numerical printouts.

It is easy to forget how far we have come in those 30 years. Younger modellers probably don’t realise that what we now take for granted as good practice was not that common even around the turn of the Century. Urban models by default are now multi-modal.  Junction modelling is a given (BTW – SATURN was ahead of its time in that respect).  No self-respecting modeller now produces outputs without a proper background map using GIS. We have got so much better at explaining our models and model results to non-experts despite many of our colleagues still considering that models are expensive, opaque and often irrelevant for decision-making. We are a much stronger profession than when I started; and in that context it is gratifying that a growing number of modellers obtain the Transport Planning Professional qualification.