The current state of activity-based travel demand modelling and some possible next steps

ABSTRACT

Despite the clear theoretical advantages of activity-based models of travel behaviour relative to trip-based models, adoption of such models in planning practice has been slow. This editorial discusses some reasons underlying this fact, including “locking into” outmoded model structures and software and challenges in translating research advances into practice. It argues for more widespread adoption of an activity-scheduling approach to the problem and identifies a number of key areas requiring new research in order to improve the operational capabilities of these models.

KEYWORDS:

• Activity-based
• travel demand modelling
• activity-scheduling
• research needs

The activity-based nature of travel behaviour is axiomatic among travel behaviour researchers and has been foundational to extensive research in the field for over forty years. This has led to the (very) slow adoption of “activity-based” travel demand models in operational planning practice, notably in North America. Nevertheless, the classic trip-based (four-step) approach remains standard practice in much of the world. Many reasons for this slow translation of research into practice exist, including the conservatism of many transportation planning agencies, as well as, often, lack of technical and financial resources to support the adoption of “advanced” modelling methods. These, in turn, often reflect a lack of interest among governments in improved policy analysis capabilities, despite protestations concerning the need for “evidence-based decision-making”. But many technical and research-related issues exist that also contribute to the current state.

The first of these is that the current state of operational practice is largely locked into a semi-standardized model structure which, while routinely labelled “activity-based”, is actually “tour-based” in design. That is, the model directly builds daily tour patterns, usually using a nested logit formulation which is quite restrictive in terms of the tours that can be constructed with respect to the number of trips per tour, constraints on feasible trip mode combinations and limitations on treatment of trip timing throughout the day. These models derive in large part from the seminal work of Bowman and Ben-Akiva in the mid-1990s (Bowman & Ben-Akiva, 1997), which, while ground-breaking at the time, is now 25+ years old. The approach is systematically described in Castiglione et al. (2015).

Second, this “locking-into” the tour-based paradigm is compounded by the typical implementation of these models in proprietary commercial software which is also rigidly designed around the nested logit, tour-based framework and so is very difficult to use as a starting point for the testing of innovative alternative modelling approaches. One exception to the proprietary software norm is ActivitySim, which is open source, but which, unfortunately, has chosen to implement an outmoded tour-based framework, and so continues to lock the field into a dead end for future model development. These software systems are also generally overly complicated in their design (reflecting, in part, the models’ poor theoretical foundations) and computationally extremely inefficient.

As a result of this state of affairs, “activity-based” models often have a “bad rap” among practitioners as being overly complicated, difficult to develop, maintain and use, and computationally inefficient. As sketched above, this may well be a fair assessment of the current state of tour-based modelling practice, but, as argued below, it need not be for the activity-based approach per se.

The alternative to tour-based models is what can be labelled activity scheduling models. These models take the activity-based paradigm more seriously in that they model activity participation explicitly, with trips emerging as the derived demand that they, in fact, are, resulting from the need to travel from one out-of-home activity location to another. Activity scheduling models focus in the first instance on generating activity episodes by type, start time, duration and location. These episodes are scheduled within a daily activity pattern, with the trips needed to move from one location to another organically forming tours of arbitrary complexity and composition over the course of continuous time within the simulated day. While inevitably somewhat complex in their software implementation, the conceptual framework of these models is much more theoretically sound, transparent and amenable to extensions and modifications than current tour-based models. They provide a sound foundation for extending current implementations to deal with enhanced behavioural theory, in-home/out-of-home trade-offs, post-pandemic shifts in activity/travel behaviour, introduction of new technologies and services, extension beyond single-day modelling, etc.

A number of advanced activity-scheduling models have been developed over the past twenty years, including ADAPTS (Auld & Mohammadian, 2012), ALBATROSS (Arentze & Timmermans, 2004), CEMDAP (Bhat et al., 2004), FAMOS (Pendyala et al., 2005), FEATHERS (Bellemans et al., 2010), MATSim (Balmer et al., 2006) and TASHA (Miller & Roorda, 2003), among others. With only a few notable exceptions, these models have not been implemented in mainstream operational planning practice. Several reasons for this relative lack of transfer into practice can be posited, including:

• As with their tour-based cousins, many activity-scheduling model implementations are not computationally efficient and complicated in programme design.

• Unlike the tour-based models, a standard model system design does not exist, with each system being bespoke and based on a variety of approaches. They also usually have been developed within a single urban region, leaving the transferability of the model to other urban regions in question. This leaves practitioners uncertain concerning the strengths and weaknesses of the alternative approaches and wary of what approach to adopt and why.

• These models have all been developed within academic settings, with the research groups often lacking the resources (and, at least sometimes, the interest) to take their models into practice. With governments generally not in the business of promoting software development, it is left to consultants/software companies to pick up the ideas floating around in academia to implement. While some of this occasionally occurs, consultants, like their public sector clients, are often conservative with respect to innovation and hesitant to invest in new software development, especially given their existing investments in tour-based models.

• The bespoke nature of these model systems may also reflect the pressures within academia to publish “unique” research that “breaks new ground” rather than contributes to the steady improvement of existing methods. While investigation of a diversity of approaches is essential to the scientific method, if there is not at some point convergence towards “best” theories and methods it is not clear what the long-run scientific contribution actually is.

The COVID-19 pandemic has further challenged all types of travel demand models, including activity-scheduling models, by exposing a number of “maintained hypotheses” within these models which may well not hold going forward into “the new normal”. Evidence exists concerning the relative stability of travel behaviour over an extended period of time pre-pandemic, at least for work and school commuting (Fox et al., 2014; Ozonder & Miller, 2021, among others). But it is clear worldwide that propensities for working from home (WfH), online shopping, and increased resistance to using public transit (among other possible impacts such as residential location) have changed dramatically and may or may not return to the pre-pandemic “norm”. No current travel demand forecasting system is capable of endogenously responding to these pandemic effects, leaving practitioners and academics alike scrambling to explore a very uncertain future through scenario analyses and/or assertion of new model parameters based on, at best, very sketchy empirical data.

Key areas of research needed to advance activity-scheduling models into a “next generation” that is better positioned to respond to future uncertainties includes improved understanding/models of:

• Activity episode generation. Whether trip-based or activity-based, our generation models are still largely “hard-wired” to historically observed participation rates. Why do we choose to work from home rather than go into to the office? Why do we shop online versus go to the store? I do not think that we can answer these questions at the moment; and, certainly, dozens of fixed constants in a model, typical of current practice, do not do so.

• Activity location choice. Whether trip-based or activity-based, the Achilles heel of all travel demand models is our weak ability to predict activity location choice robustly, especially for non-work/school activities.

• Sequencing of location/trip components choice. Location choice first or mode? ADAPTS has attempted to take this question head-on, but is this the “right” solution? How are activity episodes dynamically scheduled within the day/week? Do people “globally optimize” their daily/weekly activity pattern? (I would emphatically say not!) But if not, how do we model scheduling dynamics?

• In-home vs. out-of-home activities. We have barely begun to scratch the surface of in- vs. out-of-home activity scheduling, but this is central to understanding the “need to travel”.

• Utilising passive location tracking data. Researchers globally are exploring how to best make use of the ever-increasing available of large/huge quantities of passive location tracking data, such as cellphone traces, smartphone-based GPS traces, public transit smartcard transaction data, etc. While enormously attractive given both their large sample sizes and dynamic/time-series nature, the inevitably anonymized nature of the data (i.e. nothing is known about the trip-maker's socio-economic characteristics) and, typically, challenges with respect to the spatial–temporal precision of the data, continue to pose significant research challenges for their use in many modelling applications.

• Intra-household interactions. Any parent will tell you how much taking care of their children dominates their daily activity schedule (getting my son to hockey practice was far more “mandatory” and spatially/temporally fixed than when I needed to be at work most days). A household-based approach to activity/travel modelling is essential. But this is pragmatically challenged by many household travel surveys that do not collect detailed activity/travel data for all household members, let alone the trend towards trying to use passive location tracking data which, in addition to be anonymized, is person-based, not household-based.

And the list goes on. There is virtually no end to the number of potential research questions that can be usefully addressed. As Donella Meadows (2008) said: “We know a tremendous amount about how the world works, but not nearly enough. Our knowledge is amazing; our ignorance even more so”. But are there actions that we can take collectively that would increase the impact of individual, “curiosity-driven” research which, arguably, is the current norm in our field? Attractive options include:

• Sharing data. Access to standard datasets to test new hypotheses and model formulations would facilitate the rejection of weaker approaches and the coalescence around stronger ones. Data sharing is common in many fields, but much less so in travel behaviour research.

• Sharing software. Perhaps an even bigger barrier to the evolution of advanced, commonly supported and used models is the lack of common software platforms that are sufficiently flexible/extensible to allow alternative model formulations to be readily and consistently tested. The emergence of open-source languages such as R and Python, as well as travel-specific platforms such as MATSim and XTMF (TMG, 2023) have been welcome advances, but much more could be done to advance such platforms and have them more widely used. Such platforms could also create a potentially large user community that could collectively contribute to the evolution of the software (we see this to a certain extent with MATSim today), to the mutual benefit of all and the reduction in overall software development efforts that often duplicate rather than extend the current state of the art.

• Testbeds. In particular, bringing shared data and shared software together into an open-source testbed in which alternative models and model systems could be run head-to-head would challenge researchers to demonstrate that their contributions are truly advancing the state-of-the-art. It would also, hopefully, help break the “my model” syndrome, in which each modeller sees his/her model as “the best”, replacing this with researchers collectively investing in advancing the common state of the art.

Finally, in parallel to these efforts, our field would benefit from more meta-analyses of findings to date in various key elements of activity/travel demand modelling, including those listed above. To take but one example: there have been countless mode choice modelling papers published over the past several decades. What have we actually learned from this mass of paper and research effort that has truly improved our ability to model mode choice, especially in terms of transferability from one urban region to another? Given Transport Reviews’ mandate to provide “authoritative and up to date research-based reviews of transport related topics” this would seem to be a natural cause for the journal to champion.

Disclosure statement

No potential conflict of interest was reported by the author(s).