Mediators: Participatory Collective Intelligence for Multi-Stakeholder Urban Decision-Making

Abstract

Cities are dynamic and evolving organisms shaped through the check-and-balance of interest exchange. As cities gain complexity and more stakeholders become involved in decision-making, reaching consensus becomes the core challenge and the essence of the urbanism process. This thesis introduces a computational framework for AI-augmented collective decision-making in urban settings. Based on real-world case studies, the core decision-making process is abstracted as a multiplayer board game modeling the check-and-balance dynamics among stakeholders with differing values. Players are encouraged to balance short-term interests and long-term resilience, and evaluate the risks and benefits of collaboration. The system is implemented as a physical interactive play-table with digital interfaces, enabling two use cases: simulating potential outcomes via AI self-play, and human–agent co-play via human-inthe-loop interactions. Technically, the framework integrates multi-agent reinforcement learning (MARL) for agent strategy training, multi-agent large language model (LLM) discussions to enable natural language negotiation, and retrieval-augmented generation (RAG) to ground decisions in contextual knowledge. Together, these components form a full-stack pipeline for simulating collective decision-making enriched by human participation. This research offers a novel participatory tool for planners, policymakers, architects, and the public to examine how differing values shape development trajectories. It also demonstrates an integrated approach to collective intelligence, combining numerical optimization, language-based reasoning, and human participation, to explore how AI–AI and AI–human collaboration can emerge within complex multi-stakeholder environments.