Urban Simulation Models and ‘Digital Twins’
It’s taken something like 60 years, but British geographers have almost discovered American regional planning.
tl;dr
- ‘Digital twins’ are the same thing as [integrated] urban simulation models1
- Pretending they are something new or different does a disservice to the discipline, and forces “rediscovery” within an established body of work
- Good urban modelers should read less computer science and more urban planning, urban economics, and civil engineering
- ‘Digital twin’ is obviously a silly misnomer…
- If you find yourself using the phrase ‘digital twin’, read more Britton Harris and Lew Hopkins.2
Urban Models
Legacy, history, and ecosystems are all important for making scientific progress. Without history and legacy, people keep reinventing the wheel. Without ecosystems, the left hand does not know what the right hand is doing. Disciplinary silos, self-interested funding mechanisms, and geography all contribute to fragmentation. “Urban analysis” faces this condition to the extreme. Many disciplines claim the “urban” field, and very few scholars are seriously committed to publishing and staying abreast of multiple literatures. That leads to groupthink and insularity.
This has always been a problem in urban studies because of its multidisciplinary nature, but it’s become acutely weird in today’s desperate cash grab for “AI” funding, and the race to become the scholar who colonizes “[insert here] data science”, and all its attendant fame. The most obvious case of chaos today is the disconnection between the American tradition of urban modeling, and the British enterprise they’re calling “digital twins”. And it’s really counterproductive.3
Even in 1960, it was clear that
“During the past few years land use and traffic models have been developed to the stage at which they are now actively being applied in professional practice. This event has met with both favor and disfavor, ranging from enthusiastic acceptance to unqualified rejection” (Harris, 1960).
Which led to decades of discussion and debate about the use and abuse of urban models (Acheampong & Silva, 2015; Harris, 1985, 1994; HUNT et al., 2005; Lee, 1973, 1994; Waddell, 2011; Wegener, 1994). Note by “urban models” here, I mean sets of multiple, interacting models rooted in economic theory (what is sometimes called the Land-Use Transportation Interaction or LUTI concept of urban simulation)(Moeckel, 2017; Wegener, 2004), and each of Wegener’s nodes in the diagram below is a distinct model whose outputs feed into the next.
This massive literature, spanning more than half a century, seems to have gone unnoticed by the modern generation of quantitative geographers who are just dying to discover something new with all their algorithms and computational power. Only a few days ago, several prominent geographers confidently proclaimed
“For those who may not have noticed, [digital twins] are having a moment… [DTs] have captured the imaginations not only of academic researchers, but also policymakers and practitioners. Enthusiasm spans academic disciplines, domains, and practice. DTs have come of age at a critical juncture, able to leverage the advantages of 21st century data (big, smart, and frequent), computing (power, speed, and distributed systems), and modelling (dynamic, integrated, and flexible) to understand and resolve complex challenges, whether related to transport, infrastructure, or land use.” (Malleson et al., 2024)
That’s… troubling… because urban modeling captured the imaginations of researchers and practitioners decades ago, and it has a long history in regional science and urban planning that is being ignored by these folks. It’s been used to solve ‘complex challenges related to transport, infrastructure, and land-use’ for quite some time. Apart from the frustration of re-hashing old arguments and repeating past mistakes, it is impossible for a discipline (urban social science, here) to make progress with a memory this short and vision this myopic. This is why we’ve been ‘lost in space’ for 30 years (Isserman, 1993).
There are lots of folks in the UK today working on some kind of model funded by the Turing Institute’s Urban Analytics Programme. One particular strand of this work is focused on developing “digital twins” (“DTs”), which they describe as
“a precise computer representation, or virtual copy, of a machine or system. Unlike typical computer models, a ‘twin’ is likely to be composed of numerous interconnected digital objects that are combined to form a coherent virtual representation of the underlying system (Wagg et al., 2024), often supported by real-time data streams…
Well, for those who may not have noticed, urban simulation models have been here (and have been used for operational urban planning) for more than half a century. They’ve had several up and down moments over a long career, and there is a large literature that probes how best to use them in real decision-making contexts. If the goal of this work is ‘to implement a’“ten-year plan to make Britain a global AI superpower”’, then it would be good to build upon the global literature in integrated urban modeling. The current discourse appears woefully unaware of the constant and widespread use of integrated land-use/transportation models in American regional planning (or in France, Belgium, Switzerland, Germany… (De Palma et al., 2005; de Palma et al., 2014; Zhang et al., 2024)).
Perhaps more importantly, apart from models themselves, there is also a large American literature on planning support systems, scenario planning, and how to use models in practice (Avin et al., 2022; Chakraborty & McMillan, 2015; Goodspeed, n.d., 2019; Goodspeed & Hackel, 2019; Hopkins & Zapata, 2007). That literature stresses two things, in particular. First that models need to be communicative, not simply technical. Planning is an exercise in collective government, so it is imperative that the model can be explained to people. That doesn’t mean every citizen needs to understand the math, but that the model can be explained conceptually and is not some AI black box. Regional science learned this lesson the hard way, and my guess is the “GeoAI” and “digital twin” people will soon, if they keep pretending this path hasn’t been tread.
Metropolitan Planning Agencies needed models to assess the consequences of alternative transportation plans and policies on urban development and travel patterns. Some wanted to evaluate the effects of land policies such as the use of urban growth boundaries, or policies to promote transit-oriented development. Most wanted to be able to address these kinds of policy analysis questions with models that were behaviorally clear and as transparent as possible, avoiding the problems identified three decades ago by Lee’s critical assessment of the state of large scale urban simulation (Lee 1973), and the more general skepticism of “black-box” models that were so complex that their logic could not be explained to policy-makers or the public.
– Waddell (2010)
Second, the literature stresses that neither the model itself, nor its projections provide a guide for doing actual planning. That is, it’s a bad idea to use a model, generate outputs, then plan for the optimal scenario–you don’t know what shocks will hit the system in the future, so it is unlikely you will be able to stick to that optimal route, even if the model is perfect! Hopkins’s metaphor (almost ubiquitous in certain circles) is canoeing across a river… the plan you began with when you left the first bank may not be appropriate by the time you’re amidst the stream (Hopkins, 2001b). Thus the purpose of the modeling exercise (for scenario analysis) is often to understand different courses of action in an unpredictable system. We call these “robust plans” (Chakraborty et al., 2011).
American Sims and British Sims
In my view there are two big reasons for the different trajectories in urban simulation models developed in the U.S. versus those in the U.K. The first comes from a difference in the modeling frameworks themselves and the kinds of scientists who develop them, and the second comes from funding sources (and the model’s intended use). In the U.S. models are developed by urban economists, regional scientists, urban planners, and civil engineers. In the U.K., they seem to be developed by geographers and computer scientists (Malleson et al., 2022)4. In the U.S. funding for large-scale models comes primarily from the federal government in the form of transportation dollars. I won’t pretend to know the U.K. planning or funding context, but as best as I can tell, most modeling efforts are exclusively research projects. Together those differences end up yielding very different results.
The American Model
Urban simulation models in the American tradition were developed by economists and based on land markets, so agents’ behaviors are rooted in urban economic theory (Holmes & Sieg, 2015). One canonical example is the Detroit Simulation Model, built by NBER who, like the Turing Institute, made a major investment in large-scale urban models (just, you know, 50 years ago, when urban modeling was a novel frontier) :P. In this tradition, we try to model a series of interconnected choices (at the household-level), all of which are based on the principle of utility-maximization for each step in the choice process. It turns out, this is a good way to do policy analysis because the model’s behavioral foundation means it generalizes well to “out-of-sample” scenarios. E.g. if you raise a transit fare, will more people switch to commute by automobile?
Daniel McFadden won the Nobel prize for exactly that reason: the Bay Area model, based on discrete choice theory, was far more capable of simulating the actual response to unseen policy scenarios than anything built before. And that framework, by construction, requires inter-communication among several models (where does each household live? does it own a car? how many cars? Does the commute occur with the car?), each of which is based on its own theoretical considerations. That is, integrated models are “numerous interconnected digital objects that are combined to form a coherent virtual representation of the underlying system,” where “the system” in this case, is a set of interacting agents consuming consuming goods in space, forming land and transportation markets.
The second reason is funding sources for building urban models. Large-scale models are expensive in both computational power and intellectual power, so building and maintaining them requires a lot of resources. But, critically, integrated models are not an intellectual pursuit in the U.S.; they are built into the policy framework at the national level (Knaap et al., 2020).
In a review of ‘digital twin’ implementations, Ferré-Bigorra et al. (2022) discuss the history of different modeling strategies including the early transportation and land-use models developed in the U.S. Curiously, however, their review does not include any coverage of integrated land-use/transportation models (which meet all the criteria set out for ‘digital twins’), rendering their claim that “urban infrastructure models were typically developed separately as they had a completely different background and used different methods” out-of-date, if not wildly inaccurate. This appears to be a misunderstanding of the metropolitan planning process in the States, and following, why models are built and how they are used.
The whole point of integrated land-use/transportation modeling in the U.S. is to make infrastructure decisions(!)5 (e.g. Hanley & Hopkins (2007)) (which is why American modeling courses are taught in civil engineering and urban planning, not geography). Every large metropolitan planning organization (MPO) in the U.S. has an entire staff of modelers (then pay for more) who focus exclusively on developing these models because doing so is essentially mandated for a region to receive infrastructure funding from the Federal Highway Administration. Most things revolve around transportation modeling in the U.S., because that’s where the money comes from.6
When doing these large infrastructure investments (like adding a lane to a highway or building a new transit station), the environmental protection agency (EPA) also requires an environmental impact statement, which shows how the development will affect things like pollution in the air or water (Fan et al., 2017; Liu, 2003), so integrated land-use transportation models are connected to things like nutrient-loading models, or pollutant-emitting models that provide estimates for these reports7. We also have a host of other policy measures like impact fees and adequate public facilities ordinances that require urban infrastructure like schools, water, sewers, and firefighters have adequate capacity to accomodate the proposed growth before a development project is approved–and large scale models provide all these metrics.
During the Obama administration, HUD also had a requirement similar to the EIS, called the Fair Housing Equity Assessment, which required HUD grantees to show that housing money would be used equitable in development processes. We at NCSG8 (and many other regions around the country) spent a long time using integrated models to develop those indicators e.g. for Baltimore (or Denver). This is also a realistic way to look at things like gentrification (Dawkins & Moeckel, 2016; Knaap, 2022)
The emphasis on the use of integrated models in actual planning exercises means that many of the innovations are developed by MPOs themselves or modelers at various state or federal agencies, and may not make it into the academic literature, but instead end up as TRB presentations or federally-sponsored technical reports. In other cases, prominent scholars have left academia entirely to focus on modeling platforms in the private sector.
The British Modell
Batty (2005) is the UK’s premier modeler (and long has been (Batty, 1972)) and he’s a legend. But land-use modeling in the British tradition has always looked a bit different from the U.S. tradition. It was dominated at first by spatial interaction, then by cellular automata (CA), and has been infatuated with ABM for a few years. Unlike the U.S., this tradition is strongly influenced by computer science rather than economics and draws on notions of fractals and complexity, rather than utility and rationality. Further, unlike American models, British models haven’t really been used to do actual urban planning. So when Malleson et al. (2022) claim that “in Urban Analytics, ABM is gaining popularity as a valuable method for understanding the low-level interactions that ultimately drive cities, but as yet is rarely used by stakeholders (planners, governments, etc.) to address real policy problems,” it makes clear they are not discussing the much wider field of urban models.
It also makes for an interesting proposition when Malleson et al. (2022) claim that “only a handful of simulation models have sought to integrate aspects of spatial cognition and bounded learning”–because models in the U.S. are based on the utility maximization principle of human behavior (which I would argue is an actual model of cognition). In an econometric residential location choice model, the decision to move is estimated based on preference parameters, given current: family size, current satisfaction with neighborhood/home/commute/wages, available alternative units, etc.)–not a comparison with a random number generator (Jordan et al., 2014). For example in Llorca et al. (2022):
The land use model SILO (Moeckel, 2017) updates the population on a year-by-year basis from the base year 2011 to the future year 2050. Demographic events, such as giving birth, marriage, leave parental household, death, and household relocation are executed in random order to avoid an artificial path dependency. The interaction between events is accounted for in the following year. For example, if a couple living in a small apartment has a child born, the probability to relocate the next year is larger than without the child. Household relocation decisions are simulated based on dwelling attributes, household attributes, zonal attributes and in particular travel time to work of all workers of a household, as explained in the next subsection. New dwellings are built by developers who attempt to mimic the location preferences of households. Demographic transition, household relocation and updates of the real estate market are simulated in an agent-based environment. Housing prices are updated based on vacancy rates in the neighborhood.
Rather than learning econometric “preference” parameters from choice models, urban simulation models in the British tradition are generally given different algorithms that drive agents’ behavior. In the context of ‘policy analysis’, this means the model is being told what to do:
“Since it is challenging to predict accurately the policy diffusion curve, we utilize data assimilation, that is an “on-line” feed of data to constrain the model against observations” (Oswald et al., 2024)
And that means this model is learning the idiosyncrasies of this particular setup. The Machine Learning people call that overfitting, and it means these models are not useful for scenario analyses because they are only trained to understand a single scenario–they are not actually rooted in individual behavior. Aside–if you watched HBOs Westworld, this is why Bernard works and Delos doesn’t.
Bernard is based on the conceptual model of a particular human and Delos is based the premise of replicating a real person down to the \(n^{th}\) degree. AI-Delos falls apart in the real world because the model can’t handle a reality it hasn’t experienced before. Thus AI-Delos makes for an interesting and useful thought experiment, but a useless practical application. We learn a lot from toy models like Schelling’s, but those are not the same as operational models.
So it’s frustrating that, in a recent review article “Agent-based modelling for Urban Analytics: State of the art and challenges”, the authors omit the entire history of urban modeling in the American tradition. Of the 94 references listed in Malleson et al. (2022), not a single one is used in an operational urban context (i.e. to help actually guide decision-making). Nor is a single model in the citation list based on the premise of economic behavior. None are published in public policy journals. Almost none are published in urban planning journals, but nearly all in computer science and artificial intelligence.
There is a reason the UK simulation models are published in computer science journals (or EPB) and the U.S. models are published in urban planning, public policy, regional science, and transport/land-use journals: because the American models are designed to be used for urban planning. And yet, here are those kindly Brits to offer us sage advice on how to use models for ‘urban analytics’, and lend their vast expertise solving “complex challenges, whether related to transport, infrastructure, or land use.” How generous 😝.
(I’m looking for an argument here… If you think I’m wrong about this, lets fight it out–good urban planning is about engaging disagreement (Hopkins, 2001a) :))
Conclusion
In closing his classic article, Harris (1985) says
“Any retrospective of the history of simulation modeling of urban areas over the last 30 years will discover enormous changes, to which the theoretical orientation of regional science has made a very great contribution. Our review has demonstrated that the fields in which the greatest progress has been made are precisely those in which theory has been most closely tied to model development, and where economic ideas of market clearing and equilibrium have informed model design. The contribution of regional science and geography which is new in this respect is the recognition of diverse behaviors captured in the gravity model and discrete choice theory, and the increasing emphasis on externalities and indivisibilities.
Today, the discourse on ‘digital twins’ contains not a single trace of these ideas, and that’s a step backwards. Regional science may be dead (Berry, 1995), and British geographers have discovered “the wheel should be red” (and boy are they excited) (Fotheringham, 2023). Although I’m picking on the Brits, it’s not just them. American geographers are doing this too, for some reason. It seems clear to me that none of this knowledge has been consolidated before, and that’s why I care about urban analysis.
And I think the way we fix this is by rebuilding the ecosystem of urban analysis researchers, rather than having competing factions claim it for themselves. That means understanding where the discipline comes from; using the term ‘digital twin’ is a signal that you are not acquainted with the rich legacy of urban models.
Stop Calling Integrated Models ‘Digital Twins’: A Brief Reading List:
Lee (1973)
Lee (1994)
Ben-Akiva et al. (1997)
McFadden & Reid (1975)
McFadden (1978)
Harris (1966)
Harris (1960)
Harris (1985)
Harris (1994)
Anas (1978)
Anas (1980)
Anas (1983)
Anas (1984)
Anas & Arnott (1993)
Anas & Arnott (1994)
Spiekermann & Wegener (2018)
Wegener (1998)
Wegener (1994)
Wegener (2004)
Wegener (2021)
Hanley & Hopkins (2007)
Hopkins (1974)
Hopkins (1984)
Hopkins (2001b)
Hopkins & Zapata (2007)
Hopkins & Knaap (2019)
Avin & Goodspeed (2020)
Knaap et al. (2020)
References
Footnotes
in grad school I worked with MSTM, SILO, and used UrbanSim for my dissertation (Rolf taught research design and Fred was on my committee.), and the distinction between British modeling and American modeling has always baffled me. Now they’re (almost) doing integrated modeling, claiming it’s new, and offering advice, all without acknowledging the field’s history.🙄↩︎
I only met the guy once, but personally, I find it odd that Batty (2018) entertains the idea of ‘digital twins’–though I really appreciate the skepticism of Fotheringham (2023)↩︎
I’m also excruciatingly sick of geographers competing to define “urban” but that’s for another time… (but, really… lets have seven geographers define ‘urban informatics’ (Goodchild et al., 2024)? GIScience is not urban science, folks↩︎
Personally, I think of land markets as the defining feature of urban systems (and any model that doesn’t account for them, explicitly, as misguided), so I much prefer the former, but they have different strengths.↩︎
Ten years ago, UrbanSim was “successfully used in an applied study concerning the Grand Paris project, which can be considered as the most important investment plan in the transportation system of the region since the construction of Paris subway network at the beginning of the 20th century” (de Palma et al., 2014). That is the very definition of “infrastructure”.↩︎
This is also one of the weird and rather unfortunate things about the American urban planning aparatus… the only regional planning power any metropolitan jurisdiction really has is the ability to dictate transportation funds. So land-use models always take the back seat.↩︎
for example by hooking them up to long-developed models like LEAM↩︎
back when NSCG was a modeling center↩︎