Summary
Cities worldwide face the dual challenge of increasing density while developing sustainably. These pressures strain transportation systems, necessitating optimal performance. In the Netherlands, transport policy focuses on three main objectives: accessibility, liveability, and safety. This thesis addresses accessibility and liveability.
Liveability concerns the fit between residents and their living environment. This fit is an ecological relationship involving interactions with the built, natural, and social environments. The thesis explores two approaches to operationalising this ecological fit: static and dynamic. The static approach, exemplified by tools like the Leefbaarometer, views liveability as a measurable outcome. It considers the valuation of the environment as a proxy for fit, obtained through surveys or hedonic pricing methods. This approach is widely used in current transport policy evaluation but is limited in its ability to do ex-ante analysis.
On the other hand, the dynamic approach views liveability as an ongoing process. It considers fit the difference between what a resident expects to observe and experiences. This approach uses generative modelling, which involves causal probabilistic relationships. The manner in which these generative models are updated is unique, attempting to capture the dynamics of the reciprocal relationship between the resident and the living environment. The resident embodies the generative model, containing both internal and external states. The generative model is definitive of a resident's niche. Construction of the niche is afforded by accessibility.

Objective
The thesis aims to work towards future operationalisation of the dynamic approach to liveability. While full implementation is far beyond the scope, the objective is to understand the role of representations in defining and operationalising liveability. In the context of transport systems, representations can be understood as simplified and compressed digital twins of complicated urban environments. These representations can be used for decision-making in infrastructure planning and policy formulation.
The dynamic approach to liveability relies on the action-perception loop as its generative model. This loop means that perception is anticipated, an action in itself. Selecting optimal actions is a matter of maximising epistemic and pragmatic value. Writing a thesis is an action; hence, we follow the formal decomposition of action selection:
• Epistemic component: develop a theoretical framework, exploring the action-perception loop in liveability and the role of representations within this loop. Bridging cognitive science and transport modelling enables novel machine learning applications.

• Pragmatic component: automate the operationalisation of liveability by applying urban representation learning techniques to the province of South Holland. Our novel contribution is the developed mathematical model based on spatial convolutions. The study uses various data sources relevant to transport planning, including road networks, public transport schedules, aerial and street view images, and points of interest.

Modelling Study
Addressing the pragmatic component, we develop a novel urban representation learning technique using the H3 hierarchical hexagonal discrete global grid system (DGGS). Neural networks are used to extract representations such that each discrete hexagon is assigned a coordinate in the metric representation space. Features from different data sources are extracted using individual encoder networks. These features are then combined using a late-fusion network to obtain the final representations. We develop Ring Aggregation, a mathematical model to fuse multiple features while accounting for spatial context through sampling heuristics and spatial convolutions. The methodology considers insights from the dynamic approach to liveability, like landscapes of affordances, analogous to location-based accessibility. No operationalisation of the dynamic approach is developed in this study. Instead, this study applies the static approach and evaluates urban representations using multiple univariate linear regressions with Leefbaarometer scores as targets.

Key Epistemic Observations
• The action-perception loop is central to the dynamic approach to liveability but absent in the static approach.
• Representations play an instrumental role in the action-perception loop, informing action selection.

Key Pragmatic Observations
• Regarding sampling heuristics used to train the late-fusion network, Euclidean distance and location-based accessibility perform similarly, better preserving urban area integrity.
• The configuration of Ring Aggregation significantly impacts urban representation quality, showing heterogeneity between Leefbaarometer scores. Socially oriented scores can perform just as well with smaller receptive fields and steeper weighted average functions like exponential.
• Different data sources excel in predicting various aspects of liveability, suggesting the importance of integrated data approaches in transport policy.
• Compared to urban2vec and M3G studies, Ring Aggregation outperforms them in predicting Leefbaarometer scores across the board.

Looking Forward
By reframing the roles of indicators, perception and preferences, this thesis concludes that static and dynamic approaches to liveability are complementary. Complementarity implies that the static approach to liveability can be used to bootstrap the development of the dynamic approach. While 'living' digital twins operationalising the dynamic approach do not yet exist, this study provides building blocks towards their development. Future directions for research and application in transport policy centre on the hypothesis that hierarchical active inference models are already transport models, just waiting to be used.

This paper investigates cycling route preferences, with a focus on the cycling environment. To represent the cycling environment street-level images were used. A recently proposed model incorporates computer vision into a traditional discrete choice model to accommodate choice tasks involving numerical attributes and images. This computer vision-enriched discrete choice model (cv-dcm) was applied using a stated choice experiment, where respondents had to choose between two cycling routes. Each route was defined by three attributes, including commute time, number of traffic lights, and the cycling environment, the latter visualised using street-level images. While the cv-dcm relies on a neural network, making interpretability challenging, this study addressed this by collecting detailed cycling environment attributes. Results showed that the cycling environment was the most influential factor, with cyclists preferring green areas and separated cycling lanes. On average, cyclists were willing to take a 1.5-minute detour for a cycle trip of 11 minutes to use a separated cycling lane instead of a mixed-traffic road. These insights offer valuable insights for policymakers aiming to design cycling environments that align with cyclists’ preferences.