Activity-Based Models (ABMs) in practice mostly rely on traditional logit models to generate individual travel choices.  However, with the growing number of choice dimensions and realistic representation of constraints and interactions associated with individual travel behavior it becomes important to extend the toolbox.  While multiple promising directions have been identified in academic research (and some stand-alone examples in practice) the mainstream of ABM industry proved to be quite reluctant to adopt new methods.   

The paper overviews several extensions beyond the standard logit models with the corresponding motivation and obstacles for adoption:         

·        Optimization methods such as Liner Programming proved practical for coordinated scheduling of individual activities and travel in continuous time (essential for ABM integration with Dynamic Traffic Assignment) and resource allocation problems such as intrahousehold car allocation.  They, however, require special methods of calibration.    

·        Machine Learning (ML) / Artificial Intelligence (AI) represent a potential replacement for logit models, but this is hampered by complex elasticities and non-transparency.  Some of these methods proved to be very useful as a behavioral analysis tool.  Probably the most appealing practical direction was borrowing the ML backward propagation methods for travel model system calibration to match traffic/transit counts and/or ‘big data’. 

·        Multiple Discrete-Continuous Extreme Value (MDCEV) class of models allows for an integrated generation of individual activity participation under the time/budget constraints without a combinatorial explosion pertinent to discrete choice models.  Adoption of MDCEV in mainstream ABM until recently was hampered by a complex iterative application algorithm and somewhat incompatible ABM framework that operates with discrete activity/travel episodes rather than continuous time allocation. 

·        Mixed logit models with random coefficients. While very difficult in estimation with real-world data due to the statistical identification problems, some particular examples demonstrated an important advantage of these methods for accounting of distributed Value 0f Time (VOT) for toll road forecasting.  Replacing a distributed individual VOT with crude average values results in systematic biases in traffic and revenue forecasting.

·        Latent class discrete choice models.  While quite complex in model estimation these models offer unique insights into the hidden segmentation of individuals. The main stumbling block in practice proved to be policy sensitivity of the class membership component that in many cases resulted in a very constrained model response.  

·        Incorporating travel time reliability.  Very behaviorally appealing methods were suggested such as “schedule delay” concept.  However, their adoption in forecasting models proved to be problematic because of the required input such as preferred arrival time that is essentially unobserved.  

We discuss how to boost the adoption of these promising methods in mainstream ABMs in practice.     

The objective of this study is to analyze the activity/travel behavior of persons serving dependent members of a household. These interactions are important to incorporate in activity-based travel demand modeling because of their impact on activity scheduling and travel behavior. Activity generation and scheduling is one of the most critical and challenging components of activity-based travel models. The desire to fulfill activities generates travel demand. Activity scheduling is a function of activity type and intra-household interactions. Most of the activity/travel models focus on mandatory and discretionary activities. Mandatory activities include work and school activities. These activities are generally pre-planned with fixed locations. However, new emerging working models such as flexible work arrangements are also becoming popular. Discretionary activities include maintenance activities such as shopping and visiting a doctor; and leisure activities such as visiting friends and participating in sports. These activities have flexible schedules and change from day-to-day. Most activity/travel models focus on modeling mandatory and discretionary activities. A third category of activities which is often neglected in modeling is helping or serving a dependent household member. We use the term ‘serve dependent activity’ for these types of activities.

A dependent is a person who is unable to undertake certain activities unassisted and/or cannot be left unattended due to factors such as age and disability. In some cases, decisions concerning dependents’ activity participation are also taken by independent family members. For example, in the case of young children, parents decide their activity participation (join a play group, enroll in daycare, etc.). In this study, we mainly focus on the activity behavior of young children as a critical use case. These activities involve supervision, chauffeuring, and joint participation in activities (e.g., visiting a doctor). These activities are important to study because they largely impact the activity scheduling of independent household members and involve complex household interactions. Some serve dependent activities which are pre-planned can also be classified as mandatory activities. For example, activities such as daycare and evening school sports class have fixed schedules and locations. Supervision of serve dependents at home also impacts participation in other types of activities. Household members interact with each other to manage serve dependent activities. In this paper, we present activity behavior of serve dependents and its related household interactions. We also present a framework to model these activities and household interactions in an activity-based travel demand model i.e., Toronto Area Scheduling Model for Household Agents (TASHA) developed for the Toronto Area in Canada.

Activity-based models (ABMs) in practice involve random numbers to convert probabilistic outcomes of the logit choice models into discrete individual choices.  It is important to evaluate the model properties with respect to the microsimulation variability and its impact on the sensitivity due to changes in model inputs should be well understood.

Microsimulation variability can be better understood by evaluating aggregate and disaggregate model elasticities against several important properties:

• Repeatability of the results with the same inputs. 
• Continuity across the range of inputs from large differences to small differences.  At the aggregate level, continuity means that small changes in inputs result in small changes in aggregate shares.  At the disaggregate level, continuity means that small changes in inputs result in a small number of individual switches.
• Monotonicity across the range of scenario inputs.     
• Comparability that requires a stable difference between two scenarios across different random seeds.

The properties exhibited by various microsimulation approaches will differ in important ways. Here, we provide a systematic evaluation of the detailed scenario responses resulting from application of three different statistical options used in choice modeling:

• Monte-Carlo with Uncontrolled Random Numbers (MCR).
• Monte-Carlo with a fixed seed for each individual choice (MCF).
• Random utility simulation with fixed seed (RUF) for each individual and choice alternative that are generated in advance and kept fixed across compared scenarios.

The insights are first illustrated for a single choice dimension (one sub-model), and then for a sequence of interlinked sub-models. The results indicate that only RUF can ensure all desired properties (at both aggregate and disaggregate level) in the context of one sub-model while MCF can ensure only some of them (primarily at the aggregate level). Results indicate that MCR is the most volatile method that may violate any of the desired properties. 

Differences between the three microsimulation methods might appear minimal at the aggregate level when the scale of the simulation (number of affected individuals) and magnitude of differences between scenarios are large.  However, these differences become more prominent when considered against smaller samples on scenarios with minor changes in inputs. In the context of a sequence of interlinked sub-models, RUF choice method results in the least number of irrelevant changes in the model results due to a transportation policy and hence is the most stable out of the three choice methods.

All microsimulations for this presentation are performed using the Regional Person Travel Model for Edmonton modelling region.

This presentation introduces an extended data model designed to support disaggregate Activity-Based Models (ABMs), enabling enhanced analysis of individual travel behavior across diverse transportation modes. The model extensions facilitate detailed path tracing and user profiling, offering planners and researchers a powerful framework for multimodal mobility insights.

A generalized relational data structure connects daily activity plans—tours and trips—with spatial and modal attributes, supporting advanced select link analysis and multimodal tour representation. These enhancements enable person-level analytics, such as income and age distribution across auto, transit, and active modes.

Comparative applications using CT-RAMP, ActivitySim, and DaySim demonstrate the model’s flexibility in profiling transit rider characteristics, evaluating resource access, and assessing equity impacts. By integrating ABM outputs with network modeling tools like Visum, the framework supports robust scenario testing and policy evaluation.

This work highlights the potential of enriched ABM data structures to inform equitable, data-driven transportation planning and investment strategies.