An Agent-Based Journey in Social Systems Research: Motivations & Applications

[eng] This thesis highlights the uses of agent-based models (ABMs) in social systems, ranging from simplified testing grounds for bottom-up hypotheses to trustworthy, policy-ready tools. Throughout the course of the thesis the different models will exemplify how multi-scale heterogeneity can be woven into ABMs and stress-tested “in silico” before considering real-world implementation of the modelled phenomena. The thesis is structured as a “journey” across four social settings, progressively increasing in realism and scale. First, an evolutionary-game model explores human-AI interactions in one-shot Prisoner's Dilemmas. The setting is treated via different analytical and computational approaches. The model finds that the effectiveness of the AI behaviour to procure cooperation is highly context dependent, with purely cooperative AI -called "Samaritan AI"- finding its usefulness when imitation pressure is low. For high selective societies, only "Discriminatory AIs" based on reward denial to defectors manages to create niche phases of cooperation. Allowing humans to recognise their non-human partners further boosts cooperation when humans ignore AI performance, offering design cues for socially beneficial artificial agents. Second, a minimalist yet powerful coupling between daily mobility and contagion uncovers a “critical mobility” threshold separating stale and outbreak phases. Tying restrictions to incidence rates without accounting for this criticality drives systems into self-sustained epidemic oscillations. This chapter argues for the inclusion of the critical nature of mobility in epidemic intervention policy design, suggesting future models which minimise the stress and public fatigue associated with intermittent lockdowns and buying time for pharmaceutical interventions to arrive. Third, an ABM of digital contact tracing calibrated to the French population shows that even modest smartphone uptake profoundly shapes epidemic curves. With 50% detection of clinical cases, a mere 20% app adoption already trims the epidemic peak by roughly a third at the initial transmissibility measured for COVID-19. The model finds that with a lower transmissibility and a coverage among working-age adults above 30%, the epidemic can be driven to controllable levels. This study highlights the importance and role of heterogeneity in contact patterns with setting, with digital contact tracing serving as the primary means of detecting spurious contacts and ensuring effective isolation of cases otherwise difficult to detect. Finally, the thesis culminates in MiMoS, a high-resolution micro-mobility digital twin with the aim of reconstructing an entire city's daily transportation habits and testing low-emission-zone scenarios. Though still under active development, MiMoS illustrates how the same ABM philosophy can scale to millions of agents and richly multi-layered infrastructures, extending the thesis vision into sustainable urban planning and practical policy testing. All together, these studies exemplify the central goal of the thesis: a single modelling paradigm —agent-based simulation— can illuminate phenomena from individual behaviour to city-wide dynamics, delivering actionable insight for public health, AI-for-good implementations and urban policy.