Is it time to use activity-based urban transport models? - Springer Link

Ann Reg Sci 39:767–789 (2005) DOI 10.1007/s00168-005-0016-8

ORIGIN AL PAPER

Staffan Algers . Jonas Eliasson . Lars-Göran Mattsson

Is it time to use activity-based urban transport models? A discussion of planning needs and modelling possibilities

Received: 13 December 2002 / Accepted: 10 November 2004 / Published online: 19 November 2005 © Springer-Verlag 2005

Abstract For some decades now, transport researchers have put considerable efforts into developing what is called activity-based approaches for modelling urban travel demand. The basic idea is that travel demand is derived from people’s desires to take part in different activities. In particular, the interrelationships among different activities with respect to temporal and spatial constraints are in focus. It means that such models treat the activities and the travelling of the households with respect to where and when the activities can be carried out and how they may be scheduled, given characteristics of the households and potential opportunities, the transport networks and various institutional constraints. We discuss what demands we see on future travel demand models, with a focus on urban analysis. This discussion is somewhat biased towards what role activity-based models could play in meeting these demands. We then review in some detail three prominent and distinctly different representatives of operational activity-based models to give an indication of what new modelling possibilities they offer. Theoretical appeal, empirical validity, usefulness for planning, need for data and easiness of implementation are discussed. In the final section we draw some conclusions about the prospects of these models and of their descendants. JEL Classification C35 . R20 . R41 . R48

S. Algers . L.-G. Mattsson (*) Division of Transport and Location Analysis, Department of Transport and Economics, Royal Institute of Technology, 100 44 Stockholm, Sweden E-mail: [email protected] J. Eliasson Transek AB, Solna Torg 3, 171 45 Solna, Sweden

768

S. Algers et al.

1 Introduction For some decades now, transport researchers have put considerable efforts into developing what is called activity-based approaches for modelling urban travel demand (for overviews see Arentze and Timmermans 2000; Ettema and Timmermans 1997; Jovicic 2001; McNally 2000). The basic idea is that travel demand is derived from people’s desires to take part in different activities. In particular, the interrelationships among different activities with respect to temporal and spatial constraints are in focus. It means that an activity-based model typically, or ideally, attempts to derive the travel demand from the desire or need to perform activities, that it focuses on households and interactions among them rather than on individuals, that it handles interdependences among activities across the day, and that it observes the spatial and temporal constraints that the individuals of the households have to obey. Conventional travel demand models, on the other hand, usually take single trips (or possibly tours) of individuals as their starting point. The trips are then modelled with respect to their generation, distribution, modal split and assignment to the networks for road and public transport. This class of models has turned out to be very useful, and also reasonably easy to furnish with data and to implement. One of the major achievements has been the theoretical development of methods of calculating traffic equilibria so that congestion on the road network can be endogenously determined in response to transport policy changes. It has also been possible to extend their area of application as new policy issues have come to the forefront. Yet there is a growing awareness that these models may be inadequate for many important policy issues. They are sometimes criticised for being more suitable for construction of new infrastructure than for analysis of how an efficient management of existing demand can be accomplished (Gärling et al. 2002; Kitamura et al. 1997; Pendyala et al. 1997). Activity-based models with their explicit focus on how people organise their activities in time and space, seem to provide an appropriate perspective in evaluating such strategies. This may also be true for other urgent issues such as the implications of information and communications technology (ICT). This involves ICT for short-term traffic control, effects on short- and long-term demand for transport and on the location of activities that give rise to mobility and goods movement. However, conventional models rely on a long history of experience, while activity-based models are still close to the research frontier and have only been applied in policy studies to a limited extent. Activity-based models do not constitute a well-defined family of models. Some models are fairly close to conventional models in so far that they apply the same kind of probabilistic discrete choice framework based on random utility maximisation (RUM) as in conventional models (e.g., the Portland model by Bowman and BenAkiva 2001). Other activity-based models focus more explicitly on the scheduling process. They usually make a distinction between fixed (say work) and flexible (say shopping) activities. The open periods between the fixed activities can be filled up with different flexible activities. These activity choices can be modelled by logit models, as in PCATS by Kitamura and Fujii (1998), or by decision tables representing decision rules, as in Albatross by Arentze and Timmermans (2000, 2004a). AMOS (Activity–Mobility Simulator) by Pendyala et al. (1998) represents still another approach. Here the idea is to predict how observed activity patterns change in response to a specific policy change. An important distinction of activity-based models is that they are tour-based rather than trip-based. This means for the Portland model that all travel is viewed as

Is it time to use activity-based urban transport models?

769

round trips starting and ending at home and that the temporal and spatial connections between the activities within one tour are cared for (Bowman and Ben-Akiva 2001). A daily activity pattern, which consists of a set of tours, forms the basis for the individual’s activity and travel demand. The tours are modelled with respect to destination, mode and time of day. It should be noted, however, that to be tour-based is not a quality of activity-based models only. Today many conventional models are also tour-based, following the implementation of the idea in the Dutch National Model System in the 1980’s (Gunn 1994). Albatross is a prototype of a rule-based system that focuses on the scheduling of the activities (Arentze and Timmermans 2000). The intention has been to exploit previous research on activity-based approaches to develop a tool that could be more useful than conventional models for assessing changes in various constraints like flexible working hours or longer opening hours of shops. The AMOS model, developed by Pendyala, Kitamura and others, has a different purpose. It is meant as a system for direct analysis of behavioural adjustment, e.g. in response to proposed travel demand management policies like parking tax, improved bicycle/pedestrian facilities and congestion pricing. The idea is to start with an observed daily activity pattern for an individual; generate a modified pattern as an adaptation to some change in the external conditions; evaluate whether this new pattern is good enough with respect to some satisfacing rule; accept it or continue the search process until an acceptable pattern has been found. The decision criterion belongs to the class of bounded rationality rather than utility maximisation. The activity-based approach is not new. The intellectual roots go back to Hägerstrand (1970) and his time-geography and to Chapin (1974) and his emphasis on individuals’ desires and the personal characteristics behind their engagement in different activities. Whereas Hägerstrand stressed various forms of constraints, Chapin’s interest was more on opportunities and choice. Hägerstrand (1970) introduced three kinds of constraints on the activities an individual can undertake: “capability constraints” are biological constraints related to, e.g. the need of eating and sleeping; “coupling constraints” reflect that certain activities like a meeting require people to be at the same place at the same time; and “authority constraints” are external institutional constraints set by various kinds of regulations such as opening hours of shops and working hours of employees. This in combination with the location of the opportunities where different activities can be performed, and considering with which speed the individual can move given available means of transport, determine the prisms in the time-space, inside which the individual has to act. Activity-based models are very much about to make this simple but brilliant idea operational. In the rest of this article we first discuss what demands we see on future modelling tools for (primarily urban) transport analysis. This discussion is somewhat biased towards what role activity-based models could play in meeting these demands. We then review in some detail three prominent and distinctly different representatives of operational activity-based models to give an indication of what new modelling possibilities they offer. In the final section we draw some conclusions about the prospects of these models and of their descendants.

770

S. Algers et al.

2 Demands on future planning tools and modelling possibilities As a starting point for our reviewing of some prominent activity-based models we first discuss what demands we see on future transport modelling tools, with a focus on urban transport analysis. The need for transport models originally sprang from the need to forecast how traffic volumes would change due to infrastructure investments and exogenous changes such as economic growth and demographic changes. Soon, models were also used for short-term traffic management. Such questions probably still make up a majority of model applications. But the last decade or so, transport models have been used to analyse a broader range of questions and policies. Two major reasons for this change of focal point are the awareness that investments will not be sufficient to curb growing urban congestion, and the need to reduce the adverse environmental impacts from the transport sector. This has (among other things) increased the need to be able to evaluate demand management policies. Many such policies affect trip scheduling (trip timing and chaining), while others affect realtime travel decisions. Another type of policy measures tries to reduce congestion and traffic emissions by measures that affect land use and urban development. The increasing congestion levels have also made it even more important to take congestion-related phenomena into account, such as peak spread and travel time unreliability. With this background, we will briefly discuss 11 improvements of current modelling practice that we feel are the most needed. Some of them can be characterised as technical improvements of current standard models, while some describe new application areas. The list starts with the former group, and gradually switches over to the second group of improvements – there are improvements that fit both descriptions: 1. Temporally detailed traffic supply data, 2. Incorporating panel and time series data, 3. More realistic and/or flexible decision rules, 4. Accounting for population heterogeneity, 5. Effects of unreliable travel times, 6. Choice of departure time, trip chaining and trip interdependence, 7. Accounting for household interactions, 8. Responsiveness to changes in institutional constraints, 9. Modelling real-time traffic management and information, 10. Analysing effects of ICT on travel demand, 11. Interactions between the transport system and urban structure. These improvements could call for development of conventional as well as of activity-based models. Given our purpose, we will deal with the improvements that are particularly relevant to activity-based models in some more detail. 2.1 Temporally detailed traffic supply data As congestion increases, travellers try to avoid congestion by shifting their departure time. Many travel demand policies (such as time-varying congestion


771

charges and public transport fees) try to reduce congestion by enhancing this effect, pushing travellers towards less crowded hours. To model this, it is necessary to relate both demand data (OD-matrices) and supply data (travel time matrices) to smaller time periods, e.g. 10 min or even smaller intervals. Detailed traffic supply data is an important, and sometimes overlooked, prerequisite to reap some of the main rewards of activity-based models, namely the (intended) ability to capture how departure times and activity schedules are affected by e.g. varying congestion levels over the day. The large-scale network equilibrium models (such as EMME/2) that are still almost universally used are inherently unable to handle such small time periods. Instead, approaches that model the dynamic properties of loading and unloading congested networks are needed. Attempts have been made to develop dynamic versions of the network equilibrium flow models (see e.g. Ran and Boyce 1996). For small- or medium-sized networks, micro- or mesoscopic models (such as AIMSUN and CONTRAM) are rapidly becoming more in use. Traffic simulation is, however, outside the scope of this survey – but it is important to note that without accurate supply data, a great deal of the intended advantages of activity-based modelling disappears. 2.2 Incorporating panel and time series data Most current models rely on cross-sectional data. The underlying assumption is that the influence of factors such as income, car ownership etc. is independent of the current situation. It is, however, not evident that one can extrapolate the differences in travel pattern between, for example, car owners and non-car owners or high income and low-income households. It is conceivable that differences in travel behaviour between car owners and non-car owners will decrease or increase when a larger share of the households owns cars. This may occur if there is selfselection with regard to car ownership due to unobserved factors – for example, if those who just happen to like car driving are those who are most likely to own a car. Similar problems arise when using cross-sectional data to analyse how travel behaviour is affected by e.g. income, age or dwelling type. One way out of this dilemma is to use panel or time series data combined with the usual type of data. Even if there remains work to be done regarding the econometrics of combining data sources from different points in time (especially from a practical point of view), the main problems are data supply and funding. This issue lies outside the scope of this article, however. 2.3 More realistic and/or flexible decision rules Current models have been criticised for lacking behavioural realism. This is partly a criticism of the RUM approach. Alternative decision rules do not appear to be more solidly motivated, but are nevertheless alternatives. The striving to include more realism and detail in the modelling of activity and trip schedules is entwined with the striving for more realistic modelling of the decision process. Two candidates are rule-based algorithms (e.g. Albatross) and micro-simulated learning processes (TRANSIMS) (Bush 2000). We believe, however, that the striving for

772

S. Algers et al.

more realistic decision rules and the striving to move from trip-based to activitybased modelling are only connected at a superficial level and partly by coincidence. Other types of decision rules can be applied to single trips as well as to activity schedules; activity-based modelling does not necessarily mean that more realistic types of decision rules are being employed. When discussing the relative merits of decision rules, it is useful to make a distinction between models aiming at understanding behaviour and models aiming at forecasting it. The former type operates at the individual level, and mostly tries to mirror the actual decision process. Sometimes, a model of this type is barely quantitative at all, but is more of a structured way of thinking about the decision process. The latter type is more interested in the aggregate outcome of hundreds or thousands of individual choices (even if modern models typically have a disaggregate structure). The structure of the model does not in general aim at mirroring the decision process, even if it is often casually explained in such terms. Having said that these two purposes (understanding and forecasting) are distinctly different, it is important to note that the two traditions have much to learn from each other. Better understanding of real decision processes are likely to improve econometrics aiming at forecasting, while recent advances within discrete choice econometrics (see, e.g. Train 2002) should allow us to model and possibly also to understand the nature of decision behaviour in a better way. For example, an extension of the random utility approach is described in Ben-Akiva et al. (1999), where psychological factors affecting the decision-making process are taken into account. In this approach, indicators related to attitudes, perceptions, motivation and decision protocols aid in estimating the model. Although some prototypical examples exist, more research is needed to find out how such indicators should be constructed and linked to stated or revealed choices. The approach provides an opportunity to enrich modelling of travel behaviour. The methods of deriving decision rules from observed data (as applied in Albatross) are also being more refined (Arentze and Timmermans 2004b). Parameterised action decision trees, feature selection methods and Bayesian Belief Networks are some promising alternatives. 2.4 Accounting for population heterogeneity Current models are often disaggregated with respect to socio-economic factors, trip purposes and (naturally) geography. This is important both from a modelling point of view and for evaluating distributional consequences. But when the number of population segments and choice dimensions increases, difficulties arise for conventional, probability-based models that will need to keep track of huge multidimensional matrices of choice probabilities. One way out of this “curse of dimensionality” is to employ stochastic micro-simulation techniques, which simulates one specific sequence of choices for each individual. Such methods are now emerging, and they are probably not only necessary in order to account for population heterogeneity, but also to be able to handle the very complex and multi-dimensional choice sets that activity-based models tend to generate. (See Vovsha et al. (2005) for discussions and applications of these issues.) Apart from the fact that there are preference differences across socio-economic groups, there are also preference differences within the same, seemingly


773

homogeneous group. Effects of various measures depend on the distribution of preferences (such as the willingness to pay for time savings), not only on the mean. This means that we need to take such differences into account, both to evaluate aggregate effects more accurately, but also to evaluate impacts for different groups. In recent years, a number of different approaches have been proposed. Of these approaches, the mixed logit approach has gained much interest, since it allows for a more general distribution of the error terms and provides a possibility to improve the model specification by specifying parameter-specific distributions instead of using a fixed parameter. This means that the variation in preferences is explicitly modelled and not just included in the error term. The mixed logit approach has not yet been used in connection with large-scale model systems, but a number of research applications have been carried out. Many of these are related to stated choice experiments, as the mixed logit approach also facilitates accounting for intra-person dependencies. 2.5 Effects of unreliable travel times When transport systems operate close to the capacity limit, they generally become very sensitive to demand and supply variations. Travel time variation and the risk for unexpected delays have a noticeable impact on travel behaviour and this effect is likely to increase as congestion increases. This means that it would be desirable to incorporate measures of travel time variation and delay risks into the models, just as it is customary to include travel time, travel cost etc. If measures of travel time reliability are to be incorporated in travel demand models, the traffic network models need to be able to produce such measures based on traffic volumes and network data. This can (probably) be carried out with extensions of current traffic models, although work in this area has been scarce so far. (See Bates et al. 2004) for elements of such an approach.) Besides affecting travel behaviour ex ante, there are rescheduling effects when people have to skip a later, planned activity due to traffic delays. These system effects can be very complex. While incorporating travel time reliability in people’s ex ante choices can (in principle) be done in a standard trip-based model, modelling real-time adjustments and rescheduling effects calls for activity-based modelling. To our knowledge, there have been very few attempts at such real-time modelling. One example may be TRANSIMS, although we are not in a position to judge whether this project’s attempts at modelling within-trip rescheduling have been successful. 2.6 Choice of departure time, trip chaining and trip interdependence Most of today’s operational models do not model the choice of departure time or trip chaining in an adequate way. In order to reliably predict consequences of e.g. increasing congestion, real-time traffic information or demand management policies such as congestion charges, models need to be able to reflect choices of departure time and trip chaining. Some of these choices can in principle be incorporated in a “traditional” (nested logit) model relatively easily, for example the choice of departure time. However, changes in e.g. morning departure time will

774

S. Algers et al.

likely affect the entire activity schedule of the day. The ability to reflect such trip interdependences is the core of activity-based modelling. A major problem, however, is the lack of reliable temporarily detailed travel time data. This applies to both trip-based and activity-based models. 2.7 Accounting for household interactions Many travel demand management policies affect how individuals and households organise their daily life. Questions that are solved at a household level is for example who picks up the children, who does the shopping, who gets to use the car (if there is only one) etc. Examples of policies that affect these decisions are incentives for car pooling and allowing high-occupancy vehicles to use bus lanes. Since most of today’s models are trip-based, they are likely to produce unreliable forecasts of these effects. There might also be a general concern that models might overestimate the effects of policies that stimulate switching to other modes than car, if household interactions such as “kiss-and-ride” are neglected. This is an area where activity-based modelling should have a potential strength. However, it is also possible to include household interactions in “traditional” approaches as has been done, for example, in the SIMS model for Stockholm (Algers et al. 1995). In fact, there are very few applied activity-based models that incorporate this – see Vovsha et al. (2005) for a nice example and Albatross for elements of household interactions. One of the difficulties is to get data on what coupling and other constraints a family faces and the nature of them – whether the constraints can be “broken at a price”, for example. Although the problems might be less severe than some advocates of activitybased modelling claim, the issue will remain unresolved until dedicated investigations of this have been carried out. 2.8 Responsiveness to changes in institutional constraints Few if any conventional models take into account institutional constraints such as opening hours of shops, rules and distribution of working hours (e.g. the share of workers having different types of rules for flexible working hours) and school or day-care hours. This does not only mean that changes in such constraints cannot be modelled, but also decreases the validity and realism of the model. Several activitybased models point out this area as important and attempt to solve it. We are not aware of any dedicated studies trying to estimate how serious this problem is for conventional models. It is likely to be more important for models incorporating schedules and/or departure time choices. 2.9 Modelling real-time traffic management and information An interesting development is the increasing possibility to manage traffic in real time. A trip-maker that has on-line access to information about the current traffic situation and supply of options may wish to reschedule his activity programme, or


775

combine the activities into trip chains in a new way, as the situation changes, uncertainties dissolve or new options appear. The main, and fundamentally different, modelling consequence is that not only pre-trip information is available, but also within-trip information, which poses the need to model both the traffic process and the traveller’s decision process in a more elaborate (dynamic) way. An activity-based model that can cope with such activity rescheduling could be of great interest for evaluating the benefits of information systems in a proper way. As mentioned above, there seems to be very few attempts at real-time modelling, except for en route path choice modelling. 2.10 Analysing effects of ICT on travel demand As information technology advances, it will be possible for individuals to substitute trips for other types of contacts, predominantly through the use of Internet. This concerns not only work or business trips, but also all kinds of trips (activities) that do not require a physical presence. Thus, the ranges of substitution possibilities will in reality be extended to include also non-travel alternatives. Some attempts have been made to survey communication in a broader sense (including travel as well as non-travel contacts) (Zumkeller 2002; SIKA 1998). Our current models are less well equipped for analysing these impacts. A natural idea would be to include ICT as an additional “travel mode” in a more or less traditional mode choice model. In doing so, it will be necessary to consider the differences between activities carried out physically and by means of ICT (Salomon 1998; Mokhtarian and Salomon 2002). At the heart of this lies the fact that most current models do not analyse what the underlying “need” that causes the trip is. This means that it is difficult to analyse to what extent ICT can be a substitute for travel, and to what extent it will be a complement. Further, the time made available by teleshopping (instead of normal shopping) will perhaps be used for other trip-generating activities. This rebound effect is not captured in standard models, since trips (or at least tours) are treated as independent, and activity duration is not considered. For these kind of questions, activity-based approaches seem promising, with their explicit consideration of temporal and spatial constraints that ensures that the duration of and coupling among the activities are coped with in a consistent way. We are, however, not aware of any attempts at including ICT activities in a full-scale activity-based model. 2.11 Interactions between the transport system and urban structure There is no doubt that land use affects traffic and thus congestion. Denser cities are easier to provide with good public transport (due to the larger potential market for public transport lines), for example. Reductions of urban sprawl and attempts at integrated transport/land use planning have potentially very large impacts on the future traffic situation. The interactions between the infrastructure systems in a wide sense, the resulting travel patterns and the land-use and location structures (including land prices) and the effects on the environment are very complex. A few attempts at analysing and modelling this interaction have proven themselves operational and useful. Some of the problems that activity-based models are

776

S. Algers et al.

designed to solve also have implications for the problem of modelling land use and transport interactions. Most activity-based models have been short-term models, making them less suitable for this type of modelling. See, however, Waddell et al. (2002) for an attempt to combine an activity-based transport model with a land use model. 3 The Portland model and beyond – a RUM-based discrete choice approach The Portland model (Bowman et al. 1998; see also Bowman and Ben-Akiva 2001) represents a class of activity-based models that is based on the conventional RUM-based discrete choice framework. With the Portland model a more consistent way of modelling a person’s daily activity pattern within this framework emerged. An important reason for this modelling effort was the need to address time differentiated policy measures like congestion charging. The ability of the RUM-based discrete framework to incorporate improvements in modelling travel behaviour has allowed subsequent extensions of the Portland model concept to be established (Vovsha et al. 2005), in some cases by changing the nature of implementation rather than the model structure. Different types of further model enhancements have been introduced in recent years in several model systems in the USA (San Francisco, New York City, Columbus and Atlanta). In this section we first describe the original Portland model and then some of the later extensions. 3.1 Description of the Portland model For Portland Metro (Oregon, USA), a model system has been developed and implemented, containing an advanced example of how the choice of a combination of activities during a day can be modelled together with other choice dimensions (Bowman et al. 1998; see also Bowman and Ben-Akiva 2001). The model system is of the nested logit type, containing five major parts (see Fig. 1). The highest level consists of the Activity Pattern model, which handles the activity choices for an individual throughout a day. The model is explained in more detail below. The next level is a time of day choice model, which determines – for each home-based tour – the departure time from home and the departure time from the primary activity on the tour. This means that for each tour, a combination of two time periods forms the time of day choice, which then also determines the duration of the activity (including activities at intermediate stops). It does not, however, include a full activity scheduling facility. The subsequent levels contain mode and destination choices, which are not described here in further detail. It is important to note, however, that the accessibility to activities, described in terms of logsum variables from the mode and destination levels, influences the choices at upper levels. Necessary Level of Service data is produced by time period using a network assignment system, facilitating a consistent treatment of congestion effects and travel choices by iterative methods. As the focus is on activity modelling, a more detailed description of the Activity Pattern model follows here.


777

INPUT: households, zonal data, network data

Activity Pattern Pattern (and associated tour) probabilities

Home-based tours: time-of-day

Tour time-of-day probabilities

Home-based tours: mode and destination

Tour mode and destination probabilities

Work-based subtours

Expected tour time-of-day utilities Expected tour mode and destination utilities

Expected subtour and intermediate stop utilities (not in current implementation)

Intermediate stop locations for car driver tours OUTPUT: OD trip matrices by mode, purpose, timeof-day and income class

Fig. 1 The Portland activity schedule model system; source: Bowman et al. (1998)

In this model system, trip frequency for different trip purposes is modelled as a choice of a combination of activities. The Day Activity Pattern Model, as it is called, contains 114 alternatives, differing with respect to activities involved and the order in which the activities are performed. The choice set is described in the following way. First there are the primary activities: – – – – – –

Subsistence (work or school) on tour, Subsistence (work or school) at home, Maintenance (shopping, personal business etc.) on tour, Maintenance at home, Discretionary (social, recreation, entertainment etc.) on tour, and Discretionary (social, recreation, entertainment etc.) at home.

If the primary activity is on tour, the Day Activity Pattern Model also determines the trip chain type for that tour. There are eight possible types for work/ school tours and four possible types for maintenance and discretionary tours. The number and sequence of the stops on the tour define the trip chain type. The alternatives that apply to all trip purposes are: – Simple tour, – One or more intermediate activities on the way from home to the primary destination, – One or more intermediate activities on the way from the primary destination to home, and – Intermediate activities in both directions.

778

S. Algers et al.

For work/school tours, four additional types are defined as above with the addition of a work-based subtour. Simultaneously with primary activity and primary tour type, the Day Activity Pattern Model predicts the number and purposes of secondary tours. There are six alternatives: – – – – – –

No secondary tours, One secondary tour for work or maintenance, Two or more secondary tours for work or maintenance, One secondary tour for discretionary purpose, Two or more secondary tours for discretionary purpose, and Two or more secondary tours: at least one for work or maintenance and at least one for discretionary purpose.

Since not all of the trip chain types apply to all of the primary activity types, there are 19 possible combinations of primary activity/trip chain types. Each of the six secondary tour alternatives are possible for all primary activity/trip chain types, so the model has a total of 19×6=114 alternatives. The Portland model concept is currently being extended, and a useful overview is found in Vovsha et al. (2005). Drawing on this paper, the character of the extensions is briefly summarised in the next section. 3.2 Major extensions of the Portland model concept Modelling choices of activity combinations introduces an additional multialternative level in the choice structure compared to standard models. This leads to a significantly heavier computational burden, which has proven unpractical in large applications (such as New York). Introducing stochastic micro-simulation – instead of probability calculations for a number of categories or a sample enumeration approach – implies that subsequent choice levels can be conditioned on one choice outcome rather than on a distribution of probabilities across the full choice set. This simplifies computations and makes modelling of more complex choices possible, but requires a large enough sample to produce stable results (which may vary between the type of result wanted, giving some flexibility to use such a system in a sketch planning phase or for detailed forecasts). Full population stochastic simulations have been used in Portland, San Francisco, New York, Columbus and Atlanta. Capturing intra-household effects becomes more complex when activity patterns are concerned as compared to tours. In the Columbus system however, intra-household effects are captured in three separate ways. One way is to condition activity patterns of one household member on activities of other household members. A second way is to generate home based tours made by more than one household member at the household level rather than at the individual level. A third way is to generate maintenance activities at the household level and then allocate them to individuals. Concerning the last two cases, similar approaches were used in the SIMS model for Stockholm, but there in a tour-based setting (Algers et al. 1995).


779

The level of geographical detail is important for modelling short trips as well as for modelling public transport access and egress. The degree of detail has been depending on zone sizes, but can in principle be extended to the level of GIS system resolution. In the second version of the Portland model system, individual trips were defined at a 9400 link face level rather than at the 1250 zone level used in the METRO zone system. In the Atlanta model, land use is being treated at the 200 m2 grid cell level. The level of temporal detail is quite important for a number of applications. The Portland model applied a five-period breakdown of the day. However, many important temporal substitutions will take place within a single period using this period breakdown. In the Columbus model, a greater temporal detail was allowed, and the day was broken down into 1-h time periods. Generation and scheduling of tours are made in a consistent way, by first generating and scheduling work and school tours and then using the rest of the time to generate remaining nonmandatory tours. Each time a tour is scheduled, the hours of the day that that tour uses is made unavailable for subsequent tours. 3.3 Conclusion The RUM-based discrete choice framework has proven to be a viable framework for incorporating a number of different improvements to describe travel behaviour. There is still a way to go until all aspects listed in Section 2 are incorporated, but there is reason to believe that many of these also can be incorporated in this approach. One of the more difficult aspects would be to make the model dynamic, which would be required if the model were to be used for real-time traffic management. 4 Albatross – a rule-based activity approach Albatross is a rule-based activity model for predicting travel demand. The ability to handle institutional constraints and travel demand management strategies is in focus. The development of the prototype model is documented in Arentze and Timmermans (2000, 2004a), on which this review is based. Recently, a number of improvements and suggestions for further elaboration have been documented (Arentze and Timmermans 2004b). Albatross is to our knowledge the only fully operational computational process model. 4.1 The structure of Albatross An activity pattern is the result of a complex interaction between the needs of the individual or the household and various constraints. One obvious such constraint is the physical environment that describes where different activities can be conducted and the attributes of these locations. Activities are also constrained by institutional context. It could be rules or regulations concerning the number of working hours and flextime, opening hours of shops or requirements on individuals. All these constraints are summarised in the land-use pattern.

780

S. Algers et al.

An activity agenda is a set of activities that the individuals in a household need or wish to carry out within some time period. It is expressed at the household level to indicate that some activities like shopping or leaving or picking up children may be allocated to different members of the household. An important classification is between in-home and out-of-home activities and between mandatory and discretionary activities. From the activity agenda activities are drawn that are planned to be conducted on a specific day. The activity schedule is then the precise sequence of these activities in time and space during the day. An individual may also be involved in unplanned activities. The activities that are actually performed may differ from those planned. The realised activities are referred to as the activity pattern. Albatross predicts for each adult individual in a household and each day the probability of each combination of activity, location, travel party and mode of transport, dependent on activity agenda, available modes and land use pattern. 4.2 The scheduling process The household is assumed to be the fundamental decision unit and the scheduling is restricted to activities for the adults only. The problem is now to schedule the activities for a particular (adult) individual in a particular household during a particular day of the week. The outcome of the process, the activity pattern, will be a schedule in form of a list of activities, ordered sequentially, and an activity profile for each activity including the information about activity type, travel party, duration, start time, location, transport mode and travel time. The different activities that can be scheduled are exogenously separated into fixed and flexible activities. This reflects the distinction between mandatory and discretionary activities. The fixed activities are those that have to be included in the schedule and for which the location, start time and duration are treated as given. Work would be a typical example. Flexible activities are those that may or may not be included in the schedule. The set of fixed activities constitutes the schedule skeleton. The scheduling process will complement this skeleton by adding flexible activities to available time slots. The scheduling process is carried out in a number of steps. These steps are run through for each individual and day. The term choose is used to indicate when the decision rules of the model are applied: Step 0 Initialise the current schedule with the given set of fixed activities. Step 1 Choose transport mode for each primary work activity among the fixed activities. Step 2 For each flexible activity, initialise an episode by specifying time constraints and choose whether it should be added to the current schedule or not. If so, choose travel party and duration, and add it in a possible position given time constraints. Decide whether a new episode of the same activity should be added. Step 3 For each episode of a flexible activity, choose the time of the day of the episode, and place it in an appropriate position in the schedule.


781

Step 4 For each episode of a flexible activity, choose whether it should be part of a trip chain with other activities or form a separate trip. Step 5 For each trip chain or separate trip, choose transport mode. If a work trip is part of the trip chain, the transport mode will be that of the work trip. Step 6 For each flexible out-of-home activity, choose the location and determine the travel time. As indicated in step 1, the mode choice for the primary work activity is considered a high level decision that determines how different members of the household will use the car, if there is exactly one, for different activities. When flexible activities are considered for inclusion in step 2, which is done according to a pre-defined priority list, the locations are unknown and time constraints are set according to the widest possible opening hours across available facilities. If an activity is added, the position is chosen so as to keep maximum flexibility for remaining activities. The position of the flexible activities is further determined in step 3. In step 4 out-of-home activities are combined into trip chains, unless they are separated by an in-home activity. New in-home activities may be added to divide a trip chain into smaller chains or separate trips. The choices are based on imprecise information about temporal constraints since location, travel time and duration are still not finally determined. As indicated in step 5, mode choice is made at the trip chain rather than the trip level, i.e., it is not possible to change mode within a trip chain. Finally, locations and travel times are determined in step 6. Different activity types are then assigned to locations according to a priority list. 4.3 Decision rules The choices that are made at various stages in the scheduling process are governed by decision rules, which is the reason to classify Albatross as rule-based. A decision rule is represented in form of a decision table, DT, which consists of a list of condition variables and a list of action variables. The condition variables are related to characteristics of the individual, the household, the activities, the physical environment, the transport system, the institutional context and other schedule information. The action variables represent available choice alternatives for each choice situation. For each combination of condition variables, the DT expresses what action is taken. This idea is simplest illustrated by an example, see Table 1. We have two condition variables: travel distance to a location and parking facilities at the location. The action variables are travel modes: bike, car and public transport. The DT expresses whether a particular travel mode will be chosen or not for each combination of condition variables. An interesting feature of Albatross is that these decision rules are derived in a formalised way from empirical data. The algorithm that was applied requires a sample of person-days for which we have observations of the condition states, with respect to a number of pre-defined condition variables, and the chosen actions, as exemplified in Table 1. Assume first that there is only one column in our potential

782

S. Algers et al.

Table 1 Example of a decision table Distance (m)

D