Choice Mobility in Transport Science

• Choice Mobility is a concept in human mobility science that encompasses decisions on destination, mode, route, and timing, influenced by cognitive and infrastructural factors.
• It employs behavioral models like random utility frameworks and agent-based simulations to capture how individuals form effective opportunity sets under dynamic conditions.
• Research integrates diverse data sources and computational methods to evaluate multimodal systems, mobility hubs, and even evolutionary game theory perspectives.

Choice Mobility is a term used primarily in human mobility science for the set of decisions individuals make about whether, where, when, and how to move, including destination choice, mode choice, route choice, and departure timing, under the joint influence of preferences, spatial cognition, perceptions of the environment, multimodal constraints, and algorithmic interventions (Pappalardo et al., 2024). In transport research, the term usually denotes behaviorally grounded modeling of these decisions across multimodal systems. In a distinct mathematical usage in evolutionary game theory, it denotes a uniform-in-dimension convergence property of finite-strategy approximations, with “choice paralysis” as its complement (Anderson, 10 Sep 2025).

1. Conceptual scope and domain

Within transport and mobility science, Choice Mobility is broader than mode choice alone. It covers the formation of feasible sets, the ranking of alternatives, and the way choices are altered by infrastructure, pricing, information, service availability, and cognition. The field therefore links individual decision mechanisms to system-level outcomes such as mode share, route diversity, accessibility, congestion, vehicle miles traveled, and consumer surplus (Pappalardo et al., 2024).

Empirical work often narrows the concept to a specific decision layer while retaining this broader framing. One line of work studies mode choice directly, as in joint revealed-preference and stated-preference models of immigrants’ travel behavior, where alternatives include automobile, bus, subway, train, walk, bike, and emerging e-mobility options (Alsaleh et al., 16 Apr 2026). Another line couples mobility with adjacent decisions, such as residential location, so that commuting mode and housing location are determined sequentially and iteratively through interacting scores over time, cost, diversity, and accessibility (Yurrita et al., 2021).

A central implication is that the “choice” in Choice Mobility is not merely selection among pre-existing modes. It includes the construction of the effective opportunity set. In multilayer, time-dependent systems, what is feasible at a given time depends on schedules, transfer opportunities, vehicle availability, first-/last-mile access, and platform-mediated coordination (Pappalardo et al., 2024). This suggests that Choice Mobility is as much about endogenous choice-set formation as about utility ranking within a fixed set.

2. Cognitive, perceptual, and habitual foundations

The cognitive program on Choice Mobility emphasizes that travelers do not act on complete metric representations of space. Spatial cognition proceeds from egocentric representations to allocentric, map-like structures; memory recruits landmarks, edges, nodes, and paths; and the hippocampal system represents locations, boundaries, and distances through place cells, border cells, and grid cells, supported by head- and goal-direction cells (Pappalardo et al., 2024). Route and mode decisions are therefore mediated by limited and distorted internal representations rather than by full-information optimization.

This cognitive view is reinforced by evidence on heuristics and bounded rationality. Pedestrians and drivers often avoid objectively shortest paths, favoring anchor-based routing, salient corridors, fewer turns, greener streets, or lower-noise segments. Habit formation, recency effects, and familiarity bias further structure routine mobility, while navigation aids can impair spatial memory and alter subsequent exploration and route selection (Pappalardo et al., 2024). A plausible implication is that stable observed flows may encode learning and perceptual artifacts rather than stable structural preferences alone.

Survey-based work on commuting perceptions operationalizes these mechanisms through six criteria—ecology, comfort, financial accessibility, practicality, safety, and speed—and shows that users systematically rate their own usual mode more favorably than non-users do, with strong user/non-user gaps on practicality and time (Adam, 2024). The same study interprets observed persistence in car use through halo bias, choice-supportive bias, and reactance, and finds that policies such as free public transport, safer cycling, or “15-minute city” conditions shift users mainly among soft modes rather than strongly away from cars (Adam, 2024).

Agent-based formulations make these mechanisms explicit. In one model, each agent evaluates car, bus, bicycle, and walking using perceived attributes

$\hat{v}_{i}(m,c,t) = f_i(m,c,t)\cdot v(m,c),$

and a weighted score

$$\text{score}_i(m) = \sum_c \hat{v}_{i}(m,c,t)\cdot \text{prio}_i(c).$$

Habits are updated by a rolling-frequency rule,

$$H_m(t+1) = \left(1-\frac{1}{k}\right)H_m(t) + \frac{1}{k}\mathbf{1}[m\ \text{chosen at}\ t],$$

so that mode reuse and perception bias reinforce one another and generate lock-in (Adam et al., 2024). This model reports slow transitions under gradual infrastructure change and rapid reallocations when habits are reset, implying that policy timing and disruption effects can be as important as steady-state level-of-service changes (Adam et al., 2024).

3. Formal modeling frameworks

The canonical formalization of Choice Mobility is the random utility model. For individual $n$ and alternative $i$,

$U_{ni} = V_{ni} + \varepsilon_{ni},$

with multinomial logit probabilities

$P_{ni} = \frac{\exp(V_{ni})}{\sum_j \exp(V_{nj})}.$

Nested logit introduces correlated alternatives through inclusive values, while mixed logit integrates over random coefficients to represent taste heterogeneity, habit, and stochasticity (Pappalardo et al., 2024). This framework remains the dominant basis for empirical mode, destination, and route choice.

Mode-choice applications illustrate the range of elaborations built on this core. Mixed logit models estimated on GPS panel data allow in-vehicle time and cost sensitivities to vary across individuals and groups, while joint RP-SP formulations admit alternatives not yet observed in revealed behavior (Alsaleh et al., 16 Apr 2026). In one such study, immigrants’ value of travel time is approximately 66% lower than that of Canadian-born respondents, and a one standard deviation increase in a composite integration index decreases the probability of choosing public transit by approximately five percentage points (Alsaleh et al., 16 Apr 2026). These findings place heterogeneity, adaptation, and social integration inside the formal structure of Choice Mobility rather than treating them as exogenous descriptors.

Spatial interaction models extend the concept to destination choice and flow generation. The gravity family and radiation/intervening-opportunities models map aggregate flows from individual decisions, but the “Destination Choice Game” introduces explicit congestion and destination crowding into the utility,

$$U_{ij} = \alpha \ln A_j - \beta \ln d_{ij} - \gamma \ln D_j - \ln T_{ij},$$

subject to origin conservation (Yan et al., 2018). In this formulation, the gravity model emerges as the equilibrium of a degenerated case with no destination crowding. The paper argues that explicit interaction externalities improve predictions from intracity trips to intercity travel and migration (Yan et al., 2018).

Route-choice formulations adapted to Mobility-as-a-Service depart from standard separable cost models by treating capacities as endogenous to flows. In a MaaS network, link capacities are modeled as functions of flows across multiple links and modes, and latent binding capacities enter route utility through shadow prices estimated online (Xu et al., 2019). This converts service scarcity—such as bike or dock availability—into a real-time utility penalty and allows consumer surplus monitoring through the log-sum of route utilities (Xu et al., 2019).

Game-theoretic formulations push Choice Mobility toward strategic interaction. A multimodal “mobility game” models route, hub, and payment choice jointly, with linear congestion costs, hub waiting times, and a quadratic pricing mechanism; it is shown to admit a pure-strategy Nash equilibrium and a Price of Anarchy bound that approaches $2$ plus a term decaying with the number of travelers (Chremos et al., 2022). A different but related line develops a bilevel mobility-hub platform design model in which an upper-level platform sets subsidies while the lower level uses a link-based Perturbed Utility Route Choice assignment, yielding a strictly convex quadratic program and a single-level KKT reformulation (Yang et al., 18 Aug 2025).

4. Data, measurement, and inference

Choice Mobility research is methodologically plural because no single data source captures cognition, constraints, mode availability, and realized decisions simultaneously. Current work combines GPS traces, mobile phone CDR/XDR, smart-card data, surveys, smartphone-based travel diaries, and connected-vehicle streams, while emphasizing coverage biases, activity-dependent sampling, and the need for post-stratification and multi-source fusion (Pappalardo et al., 2024). Common measures include mode share, route diversity and entropy, generalized-cost accessibility, log-sum accessibility, and encounter-based measures of social mixing (Pappalardo et al., 2024).

A recent GPS-panel study illustrates this measurement stack at high resolution. Using a custom-built mobile application, semi-passive GPS tracking, and a joint RP-SP framework, it analyzes more than 80,000 raw GPS trip traces from 100 participants in Toronto and Montreal, cleaned to 14,502 revealed-preference trips for one model family, plus 622 in-app stated-preference scenarios (Alsaleh et al., 16 Apr 2026). Estimation is carried out in Apollo with 500 Halton draws for mixed logit, and five-fold cross-validation reports mean prediction accuracy between 80% and 82% across specifications (Alsaleh et al., 16 Apr 2026). This supports a view of Choice Mobility as a panel process shaped by repeated decisions rather than a single-shot discrete choice.

Inference also increasingly incorporates privacy and governance constraints. A blockchain-based Smart Mobility Data-market proposes local computation of objective contributions, encrypted peer-to-peer exchange, and distributed simulated annealing for parameter estimation in mobility choice models (Lopez et al., 2019). In the case study, 246 observations are split across four worker nodes; the distributed and centralized estimates coincide numerically, runtime is approximately 2 hours, and communication volume is approximately 200,000 messages per worker, or about 80 MB transmitted per worker (Lopez et al., 2019). This indicates that privacy-preserving Choice Mobility estimation is feasible, though still communication-intensive.

Causal inference remains a major difficulty. Because mobility platforms, routing apps, and new modes are selectively deployed and often proprietary, the literature recommends instrumental variables, difference-in-differences, natural experiments, sampling corrections, and privacy-preserving multi-source fusion to identify behavioral effects rather than merely correlational patterns (Pappalardo et al., 2024). This suggests that the inferential problem in Choice Mobility is not only estimating utilities, but isolating how utilities, feasible sets, and behaviorally salient information are jointly altered by interventions.

5. Multimodal systems, mobility hubs, and algorithmic agents

A major applied strand uses Choice Mobility models to evaluate multimodal system design. In mobility-hub research, one approach integrates observed hub usage into a large-scale mode choice model estimated on synthetic trips, adds a nested multimodal “hub” branch to the baseline choice set, and calibrates hub-specific parameters using on-site survey responses and ground-truth counts (Ren et al., 9 Oct 2025). In the Capital District, NY case, the implemented UAlbany Downtown Campus hub is projected to generate 8.83 multimodal trips per day, reduce annual VMT by 20.37 thousand miles, and increase daily consumer surplus by $4,000; the Downtown Cohoes hub is projected to generate 6.17 multimodal trips per day, reduce annual VMT by 13.16 thousand miles, and increase daily consumer surplus by$1,742 (Ren et al., 9 Oct 2025). The same study evaluates 1,100 candidate bus-stop clusters regionwide and finds that hubs along intercity corridors and at urban peripheries with park-and-ride access yield the strongest behavioral effects (Ren et al., 9 Oct 2025).

Mobility-on-Demand design provides another prominent use case. A multimodal framework for Manhattan couples an inner fixed-point iteration of mode shares and realized level of service with an outer Bayesian optimization over fleet sizes and discount factors (Liu et al., 2018). In a policy experiment imposing a $2 per-ride tax on ride-hailing, the optimized system shifts from ride-hailing toward ridepooling, reduces VMT from 33,835.7 to 30,288.3, raises PMT/VMT from 1.11 to 1.17, and reduces operator profit from 145,015 to 100,748 while increasing transit revenue (Liu et al., 2018). This literature treats Choice Mobility as endogenous demand that must be co-designed with supply, rather than forecast from exogenous service attributes.

Transit-centric AMoD design extends this logic to network design and frequency setting. In a Chicago case, the operator jointly sets bus frequencies, AMoD fleet allocations, and an AMoD discount factor to minimize total passenger disutility under discrete mode-and-route choice (Guo et al., 2024). The experiments show that moderate substitution of buses by AMoD can reduce average disutility when enough AMoD capacity is available, but complete replacement of buses raises excess waiting and degrades performance, particularly for downtown-oriented demand (Guo et al., 2024). A plausible implication is that Choice Mobility in dense networks is fundamentally complementarity-driven: flexible modes expand access, but mass transit remains essential for high-throughput peaks.

Micromobility research sharpens the heterogeneity dimension. A latent class choice model for electric micro-mobility in Brisbane identifies three classes among car users and two among public-transport users, with substantial variation in adoption propensity, weather sensitivity, and time and cost elasticities (Wu et al., 1 Apr 2025). Among car users, the “Multimodal trip supporters” class comprises 36.70% of the sample and has predicted shares of PMM 24%, PTSMM 47%, and Car 29%; the “Micromobility resistant” class comprises 41.68% and remains 90% car; the “Personal micromobility lovers” class comprises 21.62% and chooses PMM with predicted probability 67% (Wu et al., 1 Apr 2025). The study concludes that weather conditions, travel time, and cost are primary drivers of adoption, while resistance is strongly associated with age, gender, income, and lack of EMM experience (Wu et al., 1 Apr 2025).

Data-scarce urban simulation has recently imported LLMs into Choice Mobility. A “Preference Chain” architecture builds a weighted Belief–Desire–Intention graph, retrieves similar persons, computes path-based prior choice probabilities, and then lets an LLM remodel them using contextual text such as weather, time, and place (Hu et al., 22 Aug 2025). On the Replica dataset, the method outperforms LLM-only simulation in KLD and MAE when the reference set is below roughly 100 samples, and in a 24-hour Cambridge simulation with 1,000 agents it reduces traffic-flow KLD from 0.814 to 0.621 relative to a purely LLM-based agent (Hu et al., 22 Aug 2025). This indicates that algorithmic Choice Mobility can now be simulated through hybrid retrieval-and-reasoning systems when conventional local training data are sparse.

6. Polysemy, limitations, and open questions

The term has a formally distinct meaning in evolutionary game theory. There, for a collection of finite-dimensional trajectories $$\text{score}_i(m) = \sum_c \hat{v}_{i}(m,c,t)\cdot \text{prio}_i(c).$$0 converging to equilibria $$\text{score}_i(m) = \sum_c \hat{v}_{i}(m,c,t)\cdot \text{prio}_i(c).$$1, choice mobility is defined by

$$\text{score}_i(m) = \sum_c \hat{v}_{i}(m,c,t)\cdot \text{prio}_i(c).$$2

and choice paralysis is its negation (Anderson, 10 Sep 2025). The concept is used to determine whether finite-strategy approximations reliably reproduce the long-run behavior of an infinite-strategy game. An anticoordination example on $$\text{score}_i(m) = \sum_c \hat{v}_{i}(m,c,t)\cdot \text{prio}_i(c).$$3 shows that finite replicator dynamics can converge to uniform while the infinite game remains stuck at its initial absolutely continuous state, precisely because convergence rates collapse with dimension (Anderson, 10 Sep 2025). This usage is mathematically rigorous but semantically separate from transport-choice research.

In transport contexts, a common misconception is to equate Choice Mobility with fully rational utility maximization over stable alternatives. The literature instead portrays a layered process shaped by cognitive maps, heuristics, habits, route saliency, dynamic availability, and algorithmic anchoring (Pappalardo et al., 2024). A second misconception is to assume that introducing a new mode or hub automatically enlarges effective opportunity sets. In practice, realized accessibility depends on schedules, transfer friction, vehicle availability, weather, first-/last-mile feasibility, and segment-specific unobserved factors, many of which are only partially captured by current specifications (Ren et al., 9 Oct 2025).

Methodological limitations remain substantial. Mobility-hub calibration can rest on small survey samples and synthetic-trip baselines; one study explicitly notes a hub survey sample of $$\text{score}_i(m) = \sum_c \hat{v}_{i}(m,c,t)\cdot \text{prio}_i(c).$$4, 2019 baseline estimation for 2023–24 operations, and missing treatment of multi-hub competition, parking pricing, real-time wait, amenities, and temporal disaggregation (Ren et al., 9 Oct 2025). LLM-based simulation is subject to slow inference and residual hallucination, even when grounded by Graph RAG (Hu et al., 22 Aug 2025). Across the field, ownership bias, platform-induced nudges, proprietary algorithms, and inequitable service exposure complicate both estimation and evaluation (Pappalardo et al., 2024).

Current research directions therefore converge on several themes. One is richer multimodal, time-dependent network representation with explicit schedules, availability, and transfer reliability (Pappalardo et al., 2024). Another is better calibration through passive sensing, larger surveys, and dynamic assignment of transfers in hub systems (Ren et al., 9 Oct 2025). A third is hybridization: combining interpretable mechanisms such as gravity, EPR, nested logit, or PURC with learned components, LLM-based priors, or mobility-specific explainable AI (Pappalardo et al., 2024). Taken together, these directions suggest that Choice Mobility is evolving from a narrow mode-choice label into a general research program on how human movement decisions are formed, represented, inferred, and steered across cognitive, infrastructural, and algorithmic layers.