Activity Based Models

?

Improving realism

Efforts have been made to base transport models on a deeper understanding of the reasons for travelling. Today, these efforts see life as a series of activities undertaken at various places, and to engage in these activities, we need to travel. The trips so generated are interconnected by the sequence of activities over time and they should be the central focus of the model. This note is a broad introduction to such a modelling approach.

Travel is considered derived demand, as we rarely embark on a journey solely for the purpose of travelling itself. Rather, we travel to fulfil specific needs and perform an activity at particular locations. Looking at trips independently overlooks the intricate behavioural patterns that arise when linking activities across different locations and timeframes. Some activities can be rescheduled within certain constraints, such as rescheduling a trip to the gym based on its opening and my personal working hours. However, certain activities like work or school attendance are more challenging to shift in time. Furthermore, certain activities can be rescheduled and assigned to different individuals within a household and to different days of the week, like undertaking a major grocery shopping trip.

The development of models that attempt to capture these aspects of travel behaviour requires a change of focus and unit of analysis. The new focus is on activities and their linkage rather than simple trips and the new unit of analysis is the individual within a household performing activities over time. This group of Activity Based Models (AcBMs) requires a more granular approach to focus on individual members of households; these models are usually implemented as Agent Based Models (AgBMs) in essentially a microsimulation framework. To achieve this, a synthetic population must be generated to represent every resident and household in a specific study area. I have discussed both of these approaches in earlier posts.

What activities

Activity is a very general term that includes those performed outside the home, for example, face-to-face work, and those undertaken at home, for example having a meal or sleeping. Therefore, there is a wide variety of activities but some of them are more interesting for modelling. These would be those that normally require travel including the possible substitution with remote access: presential and remote work and education, physical and internet shopping, social, and “maintenance” activities like eating and sleeping. We perform these activities within some constraints: we only have 24 hours per day and we normally spend 1/3 of them sleeping and resting.

From a modelling perspective, we need to identify a few unambiguous activities performed by a sample of household members by means of an instrument like an enhanced Household Travel Survey. This survey should identify the main activities performed outside the home, their possible substitution by remote access, their expected duration and potential flexibility (note that for many jobs worktime is now more flexible than 10 years ago). Some activities, like sleep and mealtimes at home, can be treated as constraints rather than modelled explicitly.

As the focus is on the total activities of interest performed by individuals, AcBMs are all-day models; peak and off-peak periods can then be extracted by recording flows and performance only at certain times during the day. Because of this all-day approach, consistencies on issues like the mode of travel are easier to maintain; one can only return home by car if one used it in the “out” journey.

The key sub-models

The first step in the classic approach (Trip generation) is now replaced with a new model: Activity Generation, which is the sequence and duration of the activities individuals would perform during the modelled day. These activities may be initially fixed and identified during the survey. However, during execution of the model, the activities should become more flexible in time and space responding, in essence, to changes in travel costs and policy measures (for example restriction to vehicles that do not meet environmental protection criteria). This new sub-model, the re-scheduling of activities, is a key component of any AcBM. The activity-travel scheduling sub-model estimates for each individual the sequence of activities, tours, stops and trips during an average day and whether these would change when costs change in the future.

Tours and trips will connect these activities if their duration enables completing them within the constraints. If some trip times or costs change, the individual may wish to change the sequence and timing of their activities (their schedule) and potentially their location and mode of transport to reach them. Classic mode choice models can then be applied to tours and some trips, for example, Internal to External movements. Activities like escorting a child to school or shopping for food can be re-allocated to different members of a household and in some cases even be shifted to other days of the week; these are negotiations internal to the household. Many assumptions are needed to make this sub-model manageable but still realistic.

The re-scheduling of activities is the key element, in my view, of any of Activity Based Model; different implementations sometimes use different approaches in terms of the range of activities that are modelled, the flexibilities allowed in their scheduling and how the re-scheduling choices are actually modelled (rules or choice model). I consider that many so-called Activity Based Models do not allow such re-scheduling and focus only on how tours would adapt to changes in travel costs.

Three issues to note from this description: (1) In order to allow the logical re-scheduling of activities in the model, data must be collected on their flexibility and constraints (opening hours, minimum gap between meals, need to work for a minimum number of hours per day, etc.); therefore, the survey instrument must collect this information and then the population synthesiser must attach these characteristics to the right individuals. (2) If no re-ordering of activities is modelled, the Activity Based Model becomes a simpler Agent Based Model based on tours as they become fixed, and (3) The re-allocation of activities among household members is a reasonable response to changes in travel costs, for example after the introduction of Road User Charges or peak tolling. Modelling them requires a deeper understanding of how the decisions to share activities are dealt with in the household and the modelling of a week rather than just a day. The negotiations involved are complex with travel costs probably less important than evolving attitudes to gender roles and effective contributions to family life. It is rare to incorporate this feature into AcBMs.

Agent-based approach

It is easy to see, from the description above, that most Activity Based Models are implemented using an agent-based framework based on synthetic populations. This micro-simulation approach lends itself to treating individuals and their activities and their interactions with other members of the household. The models would not exactly clone the activities and tours of each individual in the survey sample. Monte Carlo methods are used to model some of the variability in the activities, their timings and locations as well as some of the personal attributes like gender, income, and driving license holding. Monte Carlo methods are often used to assign choices (of activities, tours, modes) to each individual during the day.

The resulting models are reasonably simple to explain but quite complex to implement and often take long to run. The use of random numbers in Monte Carlo requires multiple runs to ensure results are not influenced by the choice of the random number seed. Convergence is not guaranteed although it would appear that often stable results can be obtained with a reasonable number of runs.

To AcBM or not to AcBM

There are many costs associated with the enhanced realism of Agent and Activity Based models and therefore careful consideration is needed before adopting them as the main modelling tool. Among the costs are the need for training and the adoption of new modelling structures, the additional computer power needed to run models in a reasonable time and new data collection techniques if one wants to have a realistic activity re-scheduling model. Developing such a model from scratch seems to take up to 4 years. This timeline can be shortened if a model from another area is used as a “donor” and only a few parameters are re-estimated for the new region. However, the two areas must be similar and the assumptions adopted when developing the donor model must be acceptable for the new one.

In addition to these costs, I would suggest the following considerations should be borne in mind before deciding to adopt an Agent and an Activity Based Model:

• What are the scope and objectives of the model; is it for long-term planning, developing a Business Case for a scheme or testing new policies?
• What type of intervention will be tested in the model: new infrastructure, RUC, dealing with new technologies, addressing uncertainty, etc.
• What are the main behavioural responses needed in the model to estimate the impact of these interventions in terms of costs and benefits, equity and the environment?
• What are the data requirements, their timing and costs?
• The need for a reasonable balance between the limitations of the data (population synthesised from a 2% sample, etc.) and the granularity of the model in terms of the reliability of any forecast; in other words, does it make sense to have a very refined model based on limited data given our imperfect understanding of future uncertainties?
• Ultimately, would a better segmentation of a classic model be able to answer the urgent questions above at lower cost?

I suspect the answer to these questions will not always be that only an Agent and/or Activity Based Model will do. Nevertheless, as the role of Demand Responsive Transit grows, the need to adopt better policies to nudge them to societal objectives increasingly requires the granularity of agent-based models.