The future of models

What will transport models be like in the next 10 to 20 years

Professor Mike Batty

Professor Mike Batty

The dilemma is this: the more we enrich the model hoping to improve its simulation of the existing situation, the more problematic is its validation and the more unlikely we are to replicate the present to the extent we would like

If we pose the question, what will transportation models be like in the next 10 to 20 years, I have two responses that reflect very different assumptions: what do we think models will actually be like, and what would we like models to be? I will address the second question first because my response builds on what has been happening to such models during the last 20 years, and my response begins by sketching some of this background.

Since transportation models were first developed in the early 1950s in the United States, they have become richer and more detailed. They began as models of how populations generate trips in aggregate terms in different zones of the city and the models then worked out how such trips were distributed and assigned to the transportation network, noting how capacities and congestion could be handled during what came to be called the ‘four step process’ – trip generation, trip distribution, modal split, and then assignment. Aggregate models quickly evolved into disaggregate where the activities of individual trip makers could be simulated, and then this was relaxed even further when such trip making decisions became embedded and considered within the wider patterns of daily activities which represented motivations for why trips are generated. These latter activity models tend to represent trip makers individually subject to various constraints and rules about trip making and take a rather different approach to simulation. They cast trip makers as individual agents who interact with one another in competing for space to make trips, modelled using processes based on micro simulation.

Thus transport models began with aggregate models of how populations in small areas make trips on different modes, they then moved to examine how these same kinds of model could be formulated in terms of the generic trip maker (which were called discrete choice models reflecting the way different choices of travel were made and building on a strong stream of economic theory), then they embraced the paradigm of travel activities and time budgeting which were encapsulated in techniques of micro simulation, ultimately moving on to even richer and more detailed structures which can be loosely called agent-based models. As we have moved from aggregate spatial interaction through discrete choice to activity-based and now to agent-based models, the data and computational requirements have got greater and greater while in parallel, data itself has got easier to collect and computation has become ever more tractable.

However there is still a major trade-off between more aggregate traditional models that are easier to use and much more routine versus these more elaborate and detailed ones which are much harder to build and whose data and computational needs always stretch what is available and possible. The other key feature that in a sense those of us in modelling have always been aware of because ‘models are abstractions, simplifications of the real thing’ is that when we make models richer by disaggregation and by adding detailed processes of trip decision-making, we make them much harder to validate. A consequence of this is that in general, aggregate, more traditional models probably perform better than their newer, richer and more complex counterparts and even if they do not perform differently, neither kinds of model ever perform as well as we would like. 

The dilemma is this: the more we enrich the model hoping to improve its simulation of the existing situation, the more problematic is its validation and the more unlikely we are to replicate the present to the extent we would like. But we know that the things that we put into these newer models are important and the paradox is that if we leave them out and revert to simpler models, we are negligent of not using information that we know is important to the way people make trips and use transport. In short, as we have improved our models in terms of structure and detail, they have not got any better in terms of their performance.

The big question for me is whether such models can ever get better in terms of their performance. I fear not for many reasons. First the very systems that we are attempting to simulate are getting more complex. It is as though we need to run with our models to keep standing still. We must develop ever-richer models to deal with this complexity but there is no stability in the system we are modelling. Take trip distribution: 50 years ago the biggest single trip making activity was the journey to work, now it is no more important than trips to school and health care and so on. Take retail activity. There are massive changes going on in the online world that are hard to observe and even harder to simulate using traditional models which are changing how we purchase and deliver goods. In fact this notion of how cities are structured in terms of new information technologies – the use of email, social media, web access, search and so on – is complicating the process of transportation and our mobility in ways that are very hard to represent and model in traditional as well as even  the newer ways of simulation.

The second reason is that we are finding it increasingly difficult to produce good comprehensive data sets that cover all the elements we wish to model. Data is getting better but much of what we now need is invisible to us or too costly to acquire. 

The third reason relates to modelling itself and the fact that “ … essentially, all models are wrong, but some are useful” as George Box, the famous English statistician, once said. We seek simplifications so that we can distil the essence of the questions we are seeking to address but in transport systems, it seems that any simplification at all is too great a simplification and we lose the essence of what we are interested in.  

The fourth question is that for any model to be tractable, we need to close the system. In a global world where transport flows, many of them critical to the functioning of the system of interest, cross the boundary between the city or system of our concern and those that are beyond our remit or even interest, we can never close the system and our models are bound to be incomplete. In short, we are increasingly breaking the rule that for a good model, we need to close the system effectively. In truly open systems which cities have become, this is now impossible.

The fifth question which is some respects is the hardest of all relates to prediction in human affairs. We know that we cannot predict a future that we assume we are active in determining. Karl Popper told us this many years ago and he argued that such ‘historicism’ as he called it was a misguided aim in social science and social affairs. We all know this deep down but our inclinations are to assume the best and consider that somewhere there might be a world where our predictions turn out to be true. And, of course, there is the somewhat weaker position that many of us take where we consider that some of what we might predict will turn out to be true, notwithstanding the fact that most of what we attempt by way of forecasting is demonstrably false even in the short term. 

It is sometimes thought that short term forecasting is more reliable than long term but even this is problematic. So if we cannot build models that enable us to forecast, then what? Even predicting ranges of answers, alternative scenarios are plagued with problems for at some point a decision maker grappling with the future must make a decision: one from many possibilities must be chosen however wrong that might be. And wrong it will be, as George Box kept impressing on us.

This is a dilemma of the first order. Many have commented on this over many years and will continue to do so for it is in the nature of things – in the nature of human systems,  and that this is the situation we must live with. The big question then is if we hold the view that the future is unpredictable, what then do we do? Do we use models that we know will always give us the wrong answer? And of course we will only know this in hindsight. Or do we persist and suspend reality when we develop models, assuming that the predictions that we make will turn out to contain some truths? But even then, we seldom know which truths are more likely then others. 

Having sketched all this background, I will return to the two questions I posed at the onset: to remind you of these, I said first: in the future, ‘what do we think models will actually be like?’ while second ‘what would we like models to be?’, and I will address the second question first. I think that we need models because they impose a framework, an order on our thinking about the present and the future that imposes a discipline on how we try to address problems of transportation rigorously. 

Without some such discipline, we can only resort to intuition and to responses that are influenced by the politics of the situation, by our prejudices.  Models are thus there to inform the debate and there is still the prospect of using them conditionally to address ‘what if’ scenarios that must then be used to bound the debate. These sound like very fine words I know, and if we cannot demonstrate that any prediction can be borne out from the simplest or the most complex models, then it is entirely understandable that the effort of building them may not be judged to be worthwhile. It is thus an act of faith on model builders and policy analysts alike, on the part of any professional involved in the process of inventing a better future.  

At the end of the day I am reminded of the words of John Gillespie in 1969 when he was the Assistant County Planning Officer of Cheshire and I was a junior researcher. He said: “If we don’t use models, we might as well predict the future by reading the leaves at the bottom of our teacups”. He was serious. He argued there was no alternative. He had a passionate belief in science, not the science of prediction but the scientific method, its discipline in how to go about developing a more effective understanding. At the end of the day, this is a matter of faith and belief, and if you deny science, then you will deny models of any kind.

This is not to say that we do not need to be discriminating about types of models: we do. There are good and bad models, good and bad applications, good and bad data and so on. This brings me back to the first question: what do we think models will actually be like in the future, meaning the medium term future of the next 10-20 years. I think it will be increasingly difficult to convince policy makers of the value of computer models but what will exist is a love-hate relationship with all those involved accepting and agonising with the tension between science and its use and application in human affairs. 

As cities get more complex and as our choices and behaviours get ever more heterogeneous, as form gets increasingly disconnected from function, as much of the world of transport becomes more invisible through electronic flows which will connect to material and people flows in ever more intricate ways, we will increasingly feel the need for a science to make sense of all this. The demand for models will thus increase but so will the disillusionment because our models will increasingly lag behind the complexity of the system we are grappling with. I cannot foresee how we will handle this dilemma. I think what might happen is that the quest for comprehensive models will disappear with more and more partial models being built of bits of the system that address more specific problems. 

To an extent, what we probably need is many different models of the same phenomena because different models will reflect different perspectives on the same system and problem. How we handle this possible plethora of models that counter one another is another urgent question. The argument that we need several different and competing models of the same system is pretty untenable if each of these models is wrong in the George Box sense. It will take an act of supreme faith to argue that we need many different models to provide a wider perspective on the future when the future seems to be ever more unpredictable as we are increasingly aware it always was.

Michael Batty is Bartlett Professor of Planning, UCL, and Chairman, Centre for Advanced Spatial Analysis (CASA), UCL

www.complexcity.info

Discuss this at Modelling World 2015