Mode Choice#
The mode choice model in POLARIS is a mixed-logit model. It incorporates eleven modes, including walking, biking, car-driver (named SOV), car-passenger in household vehicle (named HOV_HH), car passenger in non-household vehicle (named HOV), taxis (including TNCs), and four public transit modes: transit (walk access and egress), rail (walk access and egress), park-and-ride (transit with car access or egress), park-and-rail (rail with car access or egress), and TNC-and-Ride (transit with any access/egress/transfer leg by TNC).
The model uses both error components, i.e., mode-specific random perception factors, and random parameters on time variables (in-vehicle time and wait time where applicable). The distribution of the error components is a multivariate normal, which can capture correlations between random mode-specific preferences. These correlations can capture detailed substitution patterns. The time-related parameters are independent truncated normal distributions on the negative axis and are therefore guaranteed to be negative. To capture agent-specific heterogeneity, the model is applied as an instantaneous panel, meaning all random parameters are drawn once per agent and then used for all mode choice decisions throughout the simulation.
Availability of modes for a specific choice situation are based on various conditions, such as household vehicle availability, walking distance to transit stations, etc. The agent-based nature of the simulation means constraints like availability of cars on tours, or availability of household members to offer rides, can be represented explicitly.
By default, level-of service attributs like mode-specific travel times (including in-vehicle time, wait time, transfer time, etc.) and distances are extracted from the skims for each mode. Optionally, these can instead be generated by the router, providing point-to-point level-of-service at second-level temporal resolution with agent-specific routing weights, but at the cost of significant runtime increase.
The utilities for each mode are calculated based on trip properties, land-use characteristics, and demographic information of the trip-makers. Demographic information includes gender, age, education, vehicle ownership, household income, and other factors. Land use characteristics utilized are activity density and network density. The value of travel time (VOTT) is calculated for each agent based on the travel mode, activity type (mandatory/non-mandatory) and household income and available for use in other models of the simulation.
Derived Modes#
Some mode types, like first-mile-last-mile (TNC-and-Ride), are difficult to estimate from survey data owing to small shares. In such cases, mode preferences are derived from the combination of preferences of the constituent modes (TNC and PT for TNC-and-Ride) and ASCs are calibrated to observed mode shares.