This study presents a High-Performance Computing and Approximate Bayesian Computation-Sequential Monte Carlo (HPC-ABC-SMC) framework for calibrating complex agent-based models. Applied to a global food trade model, the framework calibrates trade partner and trade volume objectives separately, using precomputed result mapping and task-level parallelisation to achieve up to a 42.1-fold computational efficiency improvement over traditional methods. Unlike conventional optimisation approaches, our method provides posterior distributions rather than single-point estimates, enabling richer interpretation of parameter uncertainty. The results reveal systematic differences between partner-based and volume-based calibrations: partner calibration yields smaller parameter magnitudes due to higher sensitivity, while volume calibration requires larger weights under looser constraints. Comparative analysis shows consistency with genetic algorithms but demonstrates superior interpretability, while sensitivity analysis highlights the dominant role of GDP per capita weight in shaping international trade patterns. Overall, the framework offers a scalable and uncertainty-aware solution for calibrating empirical agent-based models with high-dimensional parameter spaces and alternative evaluation criteria.
This study investigates political party competition within a two-dimensional policy space. It departs from existing computational models that typically assume full turnout. Specifically, it extends Laver and Sergenti's "baseline" (chapter 5) framework by examining the impact of voting costs on six party system outcomes: voter turnout & citizen representation, mean party & winner eccentricity (i.e., polarization), and voter & abstainer profiles. Using agent-based simulations that vary the number of parties, voting costs, and party strategies (all-hunter, all-aggregator, and all-sticker party systems), the study finds that costly voting increases party eccentricity in small all-hunter systems, while reducing it in large all-hunter systems. It also finds that the ability of all-aggregator party systems to maximize citizen representation erodes with increases in voting costs. Across all types of party systems, it finds that abstainers arise at the extremes of the policy space, and grow more moderate as costs rise. Theoretically and methodologically, the research uses experiments in a computational setting, and applies fractional polynomial regression to analyze the effect sizes of the inputs on the outcomes. The models, which integrate classic spatial models of party competition with the "calculus of voting model," yield results which reveal complex interdependence between voting behavior, party behavior, and system-level properties.
We present an agent based model (ABM), describing the origins of inequalities and social stratification in certain class of affluent hunter-gatherer societies. These societies, defined in the anthropological literature as transegalitarian, are characterised by the availability of surplus resources and the emergence of social inequalities, predating more complex structures found in sedentary agricultural communities. The relative simplicity of the social structure allows us to create a model, which focuses on a few key drivers of the process. These are: variability of key individual characteristics (skills and talents, luck, greediness), effects of the tendency for assortative matching, genetic heritability of certain advantages and disadvantages, and the presence of teaching/learning. Our results provide insights into the relative importance of these individual and societal conditions in the appearance of stable stratification in initially egalitarian societies. Depending on the choice of which individual characteristics are more prized by the society (skills and contributions to the community versus accumulated and used surplus wealth), the resulting structure may be more meritocratic or oligarchic.
Large-scale agent-based models (ABM) of complex adaptive systems often have a high number of input parameters and complicated agent interactions affecting the performance of the system. Executing traditional design of experiments to select points from the parameter space could require considerable simulation runs to effectively analyze the impact of input parameters on the system output. Sequential sampling is a promising technique for reducing simulation runs by training a machine learning metamodel to emulate the relationship between input parameters and ABM outputs while leveraging the metamodel to strategically select points to explore next in the parameter space. While previous studies have applied sequential sampling to few small ABMs, this research extends the approach to large-scale ABM and compares mixed adaptive sampling and pure sequential sampling algorithms. These algorithms were used to train a random forest metamodel emulating the Segregation model and a larger stylized ABM of a crowd logistics platform. Results indicate that the pure sequential sampling algorithm can lead to excessive clustering of design points in highly nonlinear regions and insufficient exploration of the parameter space, while the mixed adaptive sampling algorithm ensures broader coverage of the parameter space while still refining complex regions. The mixed adaptive sampling algorithm was then used to generate feature importance and partial dependence plots for the crowd logistics ABM, thereby demonstrating the usefulness of this approach not only in reducing computational requirements for ABM experimentation but also in enhancing ABM output interpretation.
Debate over whether behavioural policy should focus on individual-level (i-frame) or system-level (s-frame) interventions has been rich, but mostly qualitative. What remains underdeveloped is a formal comparative account of how errors in policy assumptions change the relative performance of these two families of interventions. We address that gap with an agent-based model because the comparison depends on spillovers, network structure, threshold effects, heterogeneous response, and shocks that cannot be inferred from average treatment effects alone. The model compares i-frame interventions (such as nudges, information campaigns, and rotating micro-targeting) with s-frame interventions (uniform structural levers and targeted seeding) across three policy-relevant sources of misspecification: imperfect structural knowledge, heterogeneous response, and external shocks. For each experimental dimension we track final adoption, diffusion speed, cost-efficiency, and downside risk, including the share of runs ending below 10% adoption. Two findings stand out. First, when structural assumptions are approximately correct, targeted s-frame seeding can trigger self-reinforcing diffusion and is markedly more transformative and cost-efficient than i-frame benchmarks. Second, the same reliance on social reinforcement creates fragility: targeting errors and backsliding shocks reduce s-frame performance more sharply, whereas i-frame designs deliver smaller but steadier gains. The contribution is a mechanism-oriented framework for comparing intervention families under uncertainty and for making the transformation-robustness trade-off explicit for behavioural policy design.
Food environments such as supermarkets and school cafeterias are increasingly dominated by ultraprocessed foods (UPFs). Consumption of these foods is partly shaped by reinforcement learning, where repeated intake leads to expected rewards. In contrast, encountering foods without consuming them can weaken these expectations through extinction - a process missing in existing models simulating food reinforcement learning. As a result, current models may fail to predict changes in learning, particularly following interventions like dieting or changes in food availability. We integrated extinction into a model of food reinforcement learning, allowing agents to update reward expectations through both consumption and non-consumption. First, we examine the conditions under which extinction prevents agents from learning to favor UPFs over less processed foods (LPFs) in environments with varying UPF availability. Second, we test interventions after agents learn in food environments resembling American supermarkets by varying (1) dieting probability and UPF availability, and (2) UPF reward value and availability. In prevention scenarios, extinction led agents to favor UPFs by suppressing learning opportunities for LPFs when UPFs were abundant. Learning outcomes were most sensitive to UPF availability. In intervention scenarios, learning shifted toward LPFs only when UPF availability was substantially reduced (from 75.9% to at most 21.6%), with minimal dieting required to initiate extinction. Neither dieting nor reducing UPF reward value alone was sufficient to reverse learning. Overall, both preventing and reversing learning for UPFs requires large reductions in their availability relative to current food environments.
Digital spaces increasingly shape the dynamics of contemporary activism in authoritarian contexts, often complementing or even substituting physical protests. Social media have been particularly empowering by providing relatively safer means of expressing and amplifying grievances in networked environments. However, they have also enabled regimes and their supporters to engage in digital repression. Existing research highlights the need for exploring online protest-repression dynamics, which we address through an agentbased modeling approach. In this study, we integrate theoretical and empirical insights into online protest and repression in authoritarian settings to model social media interactions that can result in enduring campaigns or suppressed dissent. Our findings suggest that active protest expression on social media can positively contribute to ordinary users' latent propensity to protest, even in highly authoritarian contexts and despite prevailing silence and repression. Conversely, when online protest expression is less visible, digital repression becomes more effective, especially in low-authoritarian settings, as it heightens perceptions of successful suppression among ordinary users. We also find that repressors mobilize a small yet stable subset of ordinary users, which helps them reinforce silence across most scenarios, while the impact of voiced dissent quietly accumulates within those who are potentially ready to protest. Our model implementation provides a baseline for studying online protest and repression dynamics, contributing to the theoretical understanding of digital activism in authoritarian settings and offers insights for policymakers, activists, and social media stakeholders seeking to navigate, counter, and prevent digital repression.
Conventional network generation models exhibit limitations in capturing complex structural patterns influenced by nodal and link attributes. While recent models have partially addressed this issue, significant gaps persist. This work introduces an adaptation of the Configuration Model that incorporates nodal attributes as crucial factors in determining network structure. We implemented a modification in the generation process to increase the clustering coefficient and propose two distinct algorithms: one for scenarios where the complete network topology is known and another for situations of incomplete network information. We demonstrate that our methodology can replicate network topology with edge losses below 12%, results in up to a tenfold increase in the clustering coefficient, and faithfully reproduces link distributions among different nodal categories.
Agent-based models are flexible tools that allow modellers to capture heterogeneity in agent attributes, characteristics, and behaviours. In this paper, heterogeneity is defined as agent granularity, referring to the level of detail used to describe agent attributes, behaviours, interaction processes, and decision-making rules. However, the increased complexity associated with greater levels of heterogeneity, and hence more parameters, can make the already challenging process of model calibration even more difficult. While modellers recognise the importance of calibration, the issue of uniquely determining model input based on a given output, known as parameter identification, is often overlooked. A central point of this study is that identifiability crucially depends on the outcomes or summary statistics chosen for calibration: even a well-specified model may become empirically uninformative if the selected statistics are not sufficiently sensitive to parameter variation. This paper argues that one significant impact of increasing heterogeneity in an agent-based model is the parameter identification problem, where the effects of model inputs cannot be uniquely distinguished in model outputs. To address this issue, the paper presents a comparative study of homogeneous and heterogeneous scenarios in agent-based models. Using a simple contagion case study model and approximate Bayesian computation for calibration, the study demonstrates that introducing heterogeneity reduces the accuracy of parameter calibration compared to the homogeneous case. This decline in accuracy is attributed to the difficulty in isolating the effects of the additional parameters introduced by heterogeneity. Rather than proposing computational fixes, the paper situates these findings within the broader methodological debate between KISS ("Keep It Simple, Stupid") and KIDS ("Keep It Descriptive, Stupid") strategies, highlighting how the trade-off between descriptive realism and tractability directly shapes the reliability of inference from ABMs.
Private households account for 25.8% of the EU's final energy consumption and have a rooftop photovoltaic (PV) potential of 680 TWh/year, making them key players in climate change mitigation. Unlike institutional actors, household decisions are influenced by non-rational factors such as opinion dynamics, societal and peer pressure, perceptions, preferences, advertisement and numerous cognitive factors. Understanding their decision-process and their influence on each other and the system overall is paramount for policy making and successful product launch planning, making it informative for strategies of policy makers and companies. To model sustainable product diffusion in a flexible manner, we developed the agent-based innovation diffusion framework IRPact and implemented the granular rooftop PV diffusion model PVact as a case study for the diffusion of PV systems in private households. Most models focus on explanation or prediction; however, this requires low uncertainty and a detailed understanding of the phenomena under study. In contrast, explorative modeling is suited for studying the dynamics of systems of high uncertainty. As the influence and relative strength of different decision factors on the adoption decision of private households in municipal context is not well understood, explorative modeling suggests to be a promising approach to their investigation. Through evaluating extensive simulations with a focus on the interplay of monetary evaluation, normative pressure, agent attitudes, opinion dynamics and social network parameters, we found a high sensitivity of the system to normative pressure and opinion dynamics. Variation of normative pressure showed strong phase transitions, where opinion dynamics and financial evaluation exhibited more gradual behavior. Contrary to most existing literature, the system showed no sensitivity to the variation of network parameters.
We investigate how sentiment towards certain companies spreads in social media. We focus on Reddit's financial communities, particularly those dealing with Google, Apple, and Amazon. We reconstruct users' interaction networks and we track their sentiment overtime using a few Deep Learning-based Zero-Shot approaches inspired by recent works. We then introduce an epidemiological-like agent-based model (ABM) to describe sentiment propagation within these networks. Specifically, we employ a Susceptible-InfectiousSusceptible (SIS) framework, where users expressing negative sentiment are modeled as the "infectious" agents. We also introduce trust dynamics which affects the transition probabilities between susceptible and infectious states. We calibrate the model against empirical Reddit data to align with observed sentiment trends. Finally, we validate the model using short-term out-of-sample data, demonstrating its robustness and predictive capability in forecasting sentiment.
Recent research has been conducted on "oil spill polarization", whereby preferences for topics that are not inherently politically nuanced become polarized based on political identity. Previous studies that explored the mechanisms of oil spill polarization through simulations demonstrated that this phenomenon occurs among intense partisans. However, empirical studies have shown that oil spill polarization occurs even among the general public and individuals with relatively moderate partisan identities, indicating a divergence between the simulation results and the phenomena observed in reality. We hypothesize that this discrepancy might be because of network homophily, a phenomenon often observed in social media communication. To address this, we conducted experiments using a multi-agent simulation model that implements network construction based on homophily. Our findings indicate that, when network construction is based on homophily, oil spill polarization arises among moderate partisans and diverse partisans.
Real-world social policy problems invariably involve complex dynamics with cross-domain, multifactor interactions, posing significant challenges to ex-ante policy simulation and impact evaluation. Traditional statistical modeling approaches often fail to adequately capture individual behavioral heterogeneity and system-level dynamic propagation effects, while conventional agent-based modeling (ABM) relies heavily on hand-crafted behavioral rules, which constrains its adaptability to realistic decision-making processes and diverse policy contexts. To address this challenge, we propose GPLab (Generative Policy Laboratory), a generalpurpose framework for policy simulation and evaluation that integrates generative Large Language Models (LLMs). By leveraging LLM-based agents with bounded rationality, GPLab simulates the cognition and behavior of social individuals, overcoming critical limitations of rule-based ABM in policy semantic understanding and scenario adaptation. Furthermore, its modular architecture explicitly represents interconnected social subsystems, enabling heterogeneous agents to interact dynamically across multiple domains and produce contextually grounded behavioral responses. In two representative policy evaluation cases, we successfully captured policy intensity-dependent effects, group behavioral heterogeneity, and emotional opinion evolution patterns. Formal consistency metrics further confirmed the stability of agent behavior and coherence of individual traits throughout simulations. Extended experiments demonstrated that our framework achieved high rationality scores across five policy scenarios, validating its cross-domain transferability. This study presents a scenarioagnostic policy simulation platform that significantly advances computational social science applications in public decision support. The code is available at: https://github.com/SmartLegislation/GPLab
The opioid epidemic remains a critical public health issue in the United States, with North Carolina experiencing particularly high opioid-related mortality rates. This study introduces OPOSim, an agent-based model designed to assess the long-term impacts of opioid policy interventions. By incorporating social networks and simulating opioid use dynamics, we use OPOSim to evaluate the effects of prevention strategies, specifically targeting the reduction of transitions from prescribed opioids and non-opioid substance use to heroin/synthetic opioid use disorder within North Carolina. The results demonstrate a significant delay between intervention implementation and observable mortality reductions, emphasizing the need for long-term planning. Even partial reductions in transition rates can notably decrease opioid-related deaths over time. OPOSim offers a valuable tool for understanding the opioid crisis and informing policy decisions, providing insights into the effectiveness of various interventions to mitigate the epidemic in North Carolina and similar settings.
This methodological paper is part of a companion modeling approach conducted in the Senegalese groundnut basin. It aims to address two key aspects: (1) the process of co-constructing the model as it was carried out with local stakeholders, and (2) the use of calibration to account for abstract (non-measurable) elements that stakeholders needed to describe and evaluate their system effectively. Thematically, this study explores the interplay of social solidarity and agro-pastoral systems in the Senegalese village of Diohine, where communal fallow lands persist as a critical marker of socio-ecological resilience. Using a co-constructed agent-based model, "Me Re Diem", we investigate the dynamics of soil fertility, population, and agricultural yields over 25 years, integrating local knowledge and ecological data. Our findings highlight the central role of social solidarity mechanisms, such as parcel lending and food sharing, in stabilizing agro-pastoral systems under resource scarcity. Calibration of the model demonstrates its ability to reproduce historical dynamics and test alternative practices. Without solidarity mechanisms, the system collapses, revealing their indispensability for long-term sustainability. This work underscores the necessity of collaborative modeling to bridge empirical data and local practices, fostering actionable insights for addressing soil degradation and demographic pressures.
Societies can become a conspiratorial society where there is a majority of humans that believe, and therefore spread, conspiracy theories. Artificial intelligence gave rise to social media bots that can spread conspiracies in an automated fashion. Currently, organizations combat the spread of conspiracies through manual fact-checking processes and the dissemination of counter-narratives. However, the effects of harnessing the same automation to create useful bots are not well explored. To address this, we create BotSim, an Agent-Based Model of a society in which useful bots are introduced into a small world network. These useful bots are: Info-Correction Bots, which correct bad information into good, and Good Bots, which put out good messaging. The simulated agents interact through generating, consuming and propagating information. Our results show that, left unchecked, Bad Bots can create a conspiratorial society, and this can be mitigated by either Info-Correction Bots or Good Bots; however, Good Bots are more efficient and sustainable than Info-Correction Bots . Proactive good messaging is more resource-effective than reactive information correction. With our observations, we expand the concept of bots as a malicious social media agent towards automated social media agent that can be used for both good and bad purposes. These results have implications for designing communication strategies to maintain a healthy social cyber ecosystem.
Understanding infectious disease transmission in institutional settings requires modelling approaches that can represent how contacts arise from structured routines, roles, and spatial constraints. In aged care facilities, interactions are shaped by care delivery processes, staff scheduling, and resident mobility, producing contact patterns that differ fundamentally from those assumed in population-level models. However, standard contact matrices used in epidemiological modelling are typically derived from general population surveys and do not capture these institutional mechanisms. This study develops an agent-based modelling framework to generate high-resolution contact matrices for aged care facilities by simulating task-driven behaviour, staff workflows, and movement through shared spaces. Rather than prescribing contact structure, interactions emerge endogenously from scheduled activities and proximity during task execution. The model is parameterised using collected activity-diary data from aged care workers and is implemented with behavioural logic decoupled from the physical layout, allowing adaptation to alternative facility designs without modifying core mechanisms. Simulation results show pronounced heterogeneity in contact patterns across resident care levels and staff shifts. Low and medium care residents exhibited substantially higher contact frequencies than high care residents, while staff working day and afternoon shifts accounted for the majority of resident-staff interactions. Temporal analyses revealed clustering of contacts around structured daily routines, including meals and communal activities. Integrating a proximity-based airborne transmission component parameterised for SARS-CoV-2 demonstrated that transmission risk was concentrated during high-contact shifts and among more mobile resident groups. Vaccination scenarios reduced predicted transmission substantially, with the greatest reductions observed when both staff and residents were vaccinated. By explicitly linking organisational processes to emergent contact structure, this framework provides a reproducible and transferable approach to contact matrix generation for institutional environments. The model supports more realistic transmission modelling and offers a basis for evaluating targeted infection control strategies in high-risk care settings.
