SIAS Transport Planners

Traffic Models for Real People

My next door neighbour is facing increasing difficulty driving to work each morning. I go by bike, a journey which does not attract the attention of traffic modellers, but my neighbour has quite a time of it. She's a peak period traveller, and the modellers say she must complete her journey in exactly one hour. During that hour she must appear simultaneously at all points along her route, and must sit there, immobile, in the company of other travellers unfortunate enough to be out in that same hour. At least she gets to work in one piece. My neighbour on the other side isn't so fortunate. He wants to go shopping, but the modellers slice him up before spreading him over the network to sit in simultaneous queues for precisely six hours before being allowed home. Or is it before being allowed to shop? ... he's very confused, because there appear to be slices of himself coming home with the shopping in the opposite direction, although it's difficult to tell. Something to muse on as he festers in a series of elongated boxes devised by the modellers to represent his journey.

For my neighbours, and everyone else stuffed into those boxes statically distributed about the road network, modellers have a perplexing way of representing the true nature of road traffic. A plumber would spot the analogy straight away, but still have a job relating it to his own driving experience. While the calm order of the assignment plot is not intended to be more than an expedient device for delivering some simple planning information, it is alarming to note its otherwise limited practical use. The most astonishing feature of contemporary traffic planning methodology is to claim that, because this representation constitutes a `calibrated' and `validated' model of events on the road network, it follows that we can use it to predict future traffic conditions likely to prevail under circumstances not explicitly identified.

But never mind the future, what about the disregard of current conditions? A traffic model, which is a combination of computer software and data on traffic movements to describe traffic flow, is an idealised representation which ignores illegally parked cars, bus stops, incidents, irate drivers, lollipop ladies, roadworks, the marching band of the Highland Light Infantry, bad weather, malfunctioning signals, and any other event likely to be encountered on a real journey. We can get away with the pipe flow analogy in a wide context because strategic models are intended to be no more than convenient repositories of data to be used in a broad manner, but the method does not translate to the congested urban network. Ironically, it is when the traffic is not actually moving that those static modelling systems finally break down, although this is not generally recognised.

Calibration Made Easy

Modellers use some neat tricks to ensure that our models measure up to what passes for reality, flow or no flow. The travel boxes are sliced into more bits to help explain the workings of inconvenient things like junctions, and there is a device to guarantee that we never get it wrong, called "matrix estimation". This started life as a worthy recognition that a model road network can only be bent so far before the possibility that the demand data needs some attention must be faced. Somewhere along the line the better intentions got lost (eg. maintaining the integrity of the trip cost distribution), the result being that the underlying philosophy approximates to the following complex expression :

X = ( X * Y )/Y

Where X is the answer you want (the measured traffic flow) and Y is the answer you've got (the traffic flow assigned by the model)

This need not be wholly dishonest if applied sparingly, although there is no unique solution. It can lead to 80% of a motorway arterial to a major urban area ending up in the nearest supermarket car park (name and address of perpetrator withheld). Nobody notices, because the model hides it away in its neat boxes. The solution is `validated' because the basic criterion has been met. There are models in operation throughout the UK where such horrors lurk unseen.

Matrix estimation is the last resort for the traffic planning fraternity, which had previously spent decades developing methodologies to enable traffic models to `work' without need of such remedies. These are largely `deterministic', which means that the outputs can be precisely calculated by the inputs by way of mathematical intervention. Some are devised to produce the right answer irrespective of the real mechanisms involved, which is one reason why model prediction is so precarious. There are some reasons why we do it this way :

It's effectively a black box methodology, understood only by experts who know what they're doing. There is comfort in that.

It's verifiable. It is quite easy to take the simplistic output and check against equally simplistic field observations.

It can deliver a simple economic "yes" or "no" answer, much loved by the Treasury.

Because we always have.

Public Perception

At the back end of the 20th century, to a general public which takes sophisticated computer systems in the living room for granted, the approach to transport planning is both curious and disappointing. Modellers continue to insist that some obvious problems do not exist, and the general public, making no sense of their own representation within traffic models at public exhibitions, now brand traffic planners as being somewhere between politicians and journalists on the scale of social acceptability. Readers will be familiar with the derisive response to the revelation of their professional identity, "You're a WHAT?" being the common response. They know what the problems are, and, although they have no solution, consider our general approach to be just plain daft. And before we get too haughty and technical, we might recall that our cherished methodology was borne out of an entirely unrelated response to the requirement to generate IBM 360 computer sales to town halls in the 1960s. What you see in Figure 1 could have been drawn then, and its derivation has hardly changed since.

Another Way

However, very different pictures are being drawn now, in Edinburgh, Paisley, Lanark, Perth, Dundee, Cumbernauld, Hamilton, Sheffield, Leeds, London, Salisbury, Birmingham, Tokyo, Los Angeles, Minneapolis, Singapore, and Buenos Aires, to name just some of the places which constitute rather more than a stirring in the woodpile. Figure 2 is an example of what is going on there, and is an alternative representation of the model portrayed in Figure 1. It is the output of a modelling system which thinks and drives like my neighbours, encountering the same problems that they do. The neighbours don't take convincing of this, because, when not on the printed page, they can see the vehicles in Figure 2 actually move in real-time, some behaving in a more familiar manner than others. The crazy guy at no.26 is in there, and the elderly resident of no.10 can be seen making uncertain progress through some traffic signals, causing the expected level of consternation to others of a more normalised behavioural disposition. That is the crucial factor within what has become known as microsimulation, that the moving components within the system, vehicle drivers, perform according to their aggression and awareness, two neat distillations of the myriad of influences on driver characteristics which have proved sufficient to describe the behaviour of most drivers. Gap acceptance, acceleration, top speed, headway and propensity to change lane are examples of parameters that vary according to these two behaviour parameters.

Although the distribution of levels of aggression and awareness amongst the population of vehicles on the network can be varied at any time during a simulation, these are generally held to stem from national or regional characteristics. The reader may wonder how the data for such an important cornerstone of the methodology could be readily measured, but it is only necessary to assume that the data are normalised, with a small standard deviation, the absolute values being of no concern. SIAS's experience suggests that local data regarding these parameters are not generally required. Provided the road network and all the physical influences on it are properly represented within the modelling system, drivers' progress will be both realistic and accurate. Fanciful and fantastic? Planners in Edinburgh, Paisley, Lanark, Tokyo ...etc. don't think so, and are now unlikely to return to modelling in the abstract.

Principles of Microsimulation

If we never had had a system of traffic modelling, and were to start inventing something today, what would we come up with? It is difficult to imagine that this would not be centred on the modelling of individual vehicle movements as in contemporary microsimulation. Here, behind the apparent randomness of road traffic, lies a complex order based on simple rules of car following, gap acceptance and vehicle kinematics. These can produce complex behaviour over a wide area when traffic densities are high. One reason for this complexity is that high density traffic is prone to chaotic processes which are sensitive to small disturbances. These can produce large effects, as in the `butterfly effect' of weather systems, and the familiar shock wave phenomenon of congested motorways. In one recent study conducted by SIAS it was discovered that a twenty minute incident cleared at one point on a road network could generate a queue within a shock wave over an hour later at some 5km distant, an event which microsimulation systems can represent and evaluate.

Although the evolution of such traffic congestion effects may be complex, they are predictable. The catch is that this is true only if the variables on which their representation depend are known with infinite precision, which is not possible. In practice, any measurement contains errors, and in a weather model this can amplify into gross errors. A road traffic network is a much simpler chaotic system, and so heavily constrained that, under the free flow conditions of the past it has been possible to make a reasonable stab at producing predictive models. These have generally worked within the pipe flow analogy, but if the water stops flowing the system breaks down, a not unusual circumstance. Microsimulation offers a way out of the analytical jam, and opens up the potential for including other aspects of human travel behaviour, including trip generation and modal split.

The S-Paramics Modelling System

At present, microsimulation systems receive their stratified time-based demand from external sources. SIAS's system, S-Paramics, is no exception, although it has proved useful in identifying demand data problems within other modelling systems. For instance, Figure 1 concerns a model which contains 3000 vehicles queueing on a half-mile link, a feature which passed unnoticed until the model was translated into a microsimulation representation. It then became clear that the matching of network and demand was clearly untenable, despite claims of successful validation. This highlights the first problem for models which, while readily transferable into a microsimulation format, are sometimes exposed as flawed. Although the evidence may have been there before, it is not easy to spot, buried within a mound of hard-copy output. The moving pictures of the microsimulation reveal the problem immediately. The advantage of visual microsimulation systems is that they cannot hide errors, while other systems seem too good at concealing them.

"Paramics" is an acronym derived from Microscopic Simulation on Parallel Computers, although such machines are only necessary for the representation of very large areas, such as Tokyo. S-Paramics was designed from scratch to take specific advantage of modern computer architecture. Most contemporary transport planning software is written in a general or high-level language (eg. FORTRAN) and often derived from earlier work dating back several years.

Microsimulation has been attempted by many computer-literate traffic engineers in the past, including this author, but it was not until the recent rapid acceleration in processor speed and data storage capacity that wide area microsimulation at an affordable price could become a reality. Low cost high performance computers now enable microsimulation systems to directly model all the components within traffic systems, to provide a unified approach across all sizes of road network from single junctions to national road networks. SIAS's system, S-Paramics, includes a sophisticated microscopic car-following and lane-changing model, dynamic and intelligent routeing, inclusion of intelligent transport systems, and an ability to interface to other common macroscopic data formats and real-time traffic input data sources. It takes full account of public transport and its interaction with other modes at bus stops and through bus priority measures. In addition to its comprehensive visualised real-time environment, S-Paramics provides very high speed batch mode operation for long term statistical studies. Through microsimulation, traffic engineers are able to distinguish between minor sub-optimal design variations without resorting to deterministic proxy. All known components likely to significantly effect traffic flow are represented, across the full range of road network types.

Driver and Vehicle Behaviour

In S-Paramics, the movement of individual vehicles is governed by three interacting models representing vehicle following, gap acceptance and lane changing. Vehicle dynamics are relatively simple, combining a mixture of driver behaviour and some limitations based on vehicles' physical type and kinematics (eg. size, acceleration/deceleration). These models are applied simultaneously at the level of individual vehicles which aggregate to display the characteristic features of congested traffic flow which have hitherto proved difficult to model deterministically.

Microsimulation offers a unified approach to traffic modelling which is entirely consistent with the nature of the problem. Contemporary traffic engineering practice dictates that specific analytical and design methodologies are required to address the different issues of link traffic flow, controlled and uncontrolled junctions, roundabouts and merges. Such a singular approach to separate issues, involving principles established under conditions quite different from the prevailing circumstances, are not readily adaptable to radical diversions from conventional design. Based on often long-term historic research, the deterministic methodologies of most traffic planning software fail to accommodate the behavioural characteristics inherent to congested networks, and consequently present problems of proper representation, validation and integrity.

In reality, there is a de facto unifying element to all aspects of road network and junction design, in the form of the common nature of the demand applied to these components. Vehicle drivers do not approach each network or junction type with a different mind set, although their behaviour is influenced by the prevailing geometric layout and traffic conditions. The unified simulation approach recognizes this, and places emphasis on a single system-wide calibration, embodied within the methodology itself. If microsimulation really does work, it should be able to predict the outcome of all combinations of design and demand scenarios, without the need for calibration to specific circumstances.

Validation

Deterministic macroscopic and mesoscopic models (such as NESA, TRIPS, EMME2, SATURN, TRANSYT, OSCADY, PICADY) have only a limited capacity to accommodate effects which have not been anticipated and previously measured. They can be `calibrated' if the range of their underlying determinism covers the situation to be modelled. It is generally accepted that in a strategic context this constraint may be waived in the interests of expediency, and the calibration ensures that the output of the model is adjusted to fit observation. However, because of the deterministic nature of the modelling system, such models can only "work" if the contributory circumstances do not change beyond the range within which the calibration is certified to be applicable. There is little evidence that forward projection by calibrated deterministic models can work if it involves extrapolation from base network conditions, and indeed there is no reason why it should.

A true microscopic simulation process will achieve proper validation if the road network and demand have been properly described. Although the requirement for calibration to localised circumstance is diminished, S-Paramics remains a simplification of reality, where the calibration process concerns determining the appropriate level of detail sufficient to achieve validation. This is the case for all modelling systems, the difference being that microsimulation systems model the detail directly, and are therefore likely to be closer to reality. As a result, the many SIAS commissions undertaken with S-Paramics since late 1996 have been conducted with greater confidence.

Microsimulation in the Real World

Central Government agencies have commissioned S-Paramics models for reassurance where a departure from standard has occurred, and have been satisfied with junction designs produced with the assistance of microsimulation analyses where other contemporary systems have been difficult to use. Figure 3 shows a still from the S-Paramics analysis of the six-arm roundabout on the A720 Edinburgh City Bypass, and is a typical example of a problem which proved difficult to examine with ARCADY and TRANSYT, where signalisation appeared to offer a solution for a safety issue, but potentially to the detriment of capacity. The microsimulation analysis and design testing enabled signal settings to be fixed for variable demand without serious consequences to capacity, and the design has now been implemented.

Unlike their macroscopic counterparts, microsimulation models can predict events that they have not been designed to anticipate. A simple example is that they can correctly derive the speed of shock waves in dense traffic without explicitly modelling them, a feature which has proved valuable for the assessment of trunk route improvement schemes. The shock waves generated from the incident in Figure 4 travel over the horizon, an effect which can be assessed in operational and economic terms to demonstrate that a reduction in the effect of incidents can dwarf economic returns produced by COBA.

In Conclusion

Traffic planners and engineers are conservative by nature, and many view microsimulation with suspicion. Some consider it a revolutionary approach with insufficient integrity, yet continue to struggle with systems designed to address issues of an earlier age. In fact, there is very little to be suspicious about, because, with microsimulation, what you see is what you get. If a microsimulation model doesn't look right, then it probably isn't, and vice versa. It is little more than applied common sense.

Important handbooks available to SIAS clients and S-Paramics licence holders