Why the domain fits
Urban planning is built on extensive, noisy, incomplete sensing (traffic counts, energy and water flows, land-use patterns, social indicators) that maps directly onto active inference's observations, which agents use to infer latent states such as demand patterns, infrastructure condition, and social need. It is also fundamentally about policy selection and control — zoning, infrastructure investment, operational rules, governance arrangements — which active inference treats as actions chosen to minimize expected free energy across epistemic (uncertainty-reducing) and pragmatic (outcome-realizing) terms. Cities are additionally non-stationary (demographic shifts, new technology, policy change) and multi-scale/multi-agent (intersections to districts to whole-city strategy), both of which hierarchical, continuously-updating generative models are designed to handle.
State of the literature
Foundational theory comes from Friston's work on attention, uncertainty, and free energy (recognition dynamics and variational inference) and on planning/navigation as active inference, plus a narrative-as-active-inference account (Frontiers in Psychology) of how collective sensemaking functions as group-level inference. Applied, peer-reviewed work exists in two domains: urban water governance, where Karpouzoglou and colleagues (Ecology and Society) analyze state-reinforced knowledge infrastructures as supporting group-based active inference in adaptive governance, and the built environment, where BEACON (Frontiers in the Built Environment) embeds an Active Inference Simulation ontology for representing tacit architectural design knowledge. Traffic control work — a SUMO-based active inference controller and an arXiv preprint on adaptive signal control in noisy, non-stationary IoT environments — and Millidge's active inference tree search in large POMDPs remain at the simulation/preprint stage, alongside adjacent variational Bayesian reinforcement learning applied to urban infrastructure and sustainable mobility.
Open problems
The report flags scaling as the most pressing challenge: real cities have far larger state spaces and more complex dynamics than the simulations and demonstrators tested so far, requiring more tractable hierarchical inference and, eventually, large-scale simulations and operational pilot deployments with longitudinal performance data. It also identifies unresolved work on integrating human behavior and narrative into generative models, on distinguishing descriptive uses of active inference (explaining how institutions already behave, as in the water-governance paper) from normative uses (designing new controllers, as in BEACON, GEO-INFER, and EcoNet), on establishing shared benchmarks and evaluation standards across domains, and on the ethical, legal, and political questions raised by priors and objectives encoded into planning-relevant generative models.
Reference Backbone
Karl J. Friston (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience. DOI: 10.1038/nrn2787. Maxwell J. D. Ramstead, Karl J. Friston, Axel Constant, Lancelot Da Costa, Casper Hesp, Beren Millidge, Alexander Tschantz (2023). On Bayesian Mechanics: A Physics of and by Beliefs. arXiv. Giovanni Pezzulo, Francesco Rigoli, Karl J. Friston (2017). Active Inference, homeostatic regulation and adaptive behavioural control. Progress in Neurobiology. DOI: 10.1016/j.pneurobio.2017.08.001. Pablo Lanillos, Cristian Meo, Corrado Pezzato, et al. (2021). Active Inference in Robotics and Artificial Agents: Survey and Challenges. arXiv. DOI: 10.48550/arXiv.2112.01871.