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.
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.
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.
The COVID-19 pandemic prompted a surge in computational models to simulate disease dynamics and guide interventions. Agent-based models (ABMs) are well-suited to capture population and environmental heterogeneity, but their rapid deployment raised questions about utility for health policy. We systematically reviewed 536 COVID-19 ABM studies published from January 2020 to December 2023, retrieved from Web of Science, PubMed, and Wiley on January 30, 2024. Studies were included if they used ABMs to simulate COVID-19 transmission, where reviews were excluded. Studies were assessed against nine criteria of model usefulness, including transparency and re-use, interdisciplinary collaboration and stakeholder engagement, and evaluation practices. Publications peaked in late 2021 and were concentrated in a few countries. Most models explored behavioral or policy interventions (n = 294, 54.85
Online social networks are often seen as a breeding ground of political polarization. This study introduces an additional explanation why, attributing polarization to success-driven user activity. Using an agentbased model, we demonstrate that polarization intensifies when users become more active after experiencing rewarding interactions on the platform. We compare a basic version of Axelrod's cultural dissemination model, which lacks success-driven activity, with an extended version that does include it. Our analyses replicate key findings from the literature and show that success-driven activity consistently enhances polarization, even in scenarios like complete networks or minimal network clustering, where Axelrod's model typically predicts uniformity to be unavoidable. Success-driven activity triggers a self-reinforcing "rich-get-richer" dynamic, where success leads to more activity and vice versa, resulting in a highly skewed success distribution. A few highly successful users dominate discussions, causing local convergence and fragmenting the population into distinct groups. This polarization arises in a model without biased media, polarized elites, algorithmic echo chambers, or users intentionally distancing themselves from others. We discuss the implications for designing online social networks that do not exacerbate polarization and for creating digital twins of online platforms for regulatory and analytical purposes.
How can a participatory simulation model contribute to understanding social-ecological dynamics and fostering alternative strategies for the sustainable management of trees, considering their role in enhancing soil fertility and contributing to farmers' food security? In Senegal's groundnut-growing region (Peanut Basin), the Sahelian ecosystems have undergone significant degradation, characterized by a reduction in tree cover, as a consequence of the droughts in the 1960s and the 1990s. The Peanut Basin stands out for its positive interrelationships between trees, crops, and pastoralism. However, the regeneration of the Faidherbia albida tree population, essential to these interrelationships, has declined since the major droughts. Through collaborative efforts with agro-pastoral farmers living in the area, we developed a simulation model - The SAFIRe model: Simulation of Agents for Fertility, Integrated Energy, Food security, and Reforestation - that aims to identify, in collaboration with local communities, ways to reduce the pressure on tree populations and to restore conditions for medium-term tree regeneration. By exploring the results of the model co-designed with local stakeholders, we identified potentially more effective management strategies. More importantly, we have collectively questioned the conditions for improving tree cover and the viability of the socio-ecosystem, particularly in relation to the demand for firewood and local cereal for sustenance. This has prompted the stakeholders to engage in community-wide discussions and transform agro-pastoralists into leaf carers.
Agent-based modeling proliferates across applications and scientific disciplines. The downsides of this success are the plurality of code implementations and redundant solutions to recurring modeling tasks. It is especially critical for simulations concerned with modeling human behavior and social institutions. Reusable building blocks (RBBs) are seen as a solution due to their potential to foster standardization grounded in best practices, integration of domain knowledge (including qualitative social sciences) in code, and efficient model design. RBBs are compact code components representing mechanisms or processes useful across models and applications. RBBs have been extensively discussed in the agent-based community, with little progress in implementation. Here, we present an open-access online community platform-AGENTBLOCKS-designed to facilitate the sharing, comparison, review, reuse, and improvement of RBBs. As an international community effort, AGENTBLOCKS leverages lessons from past RBBs discussions and principles from other modeling communities that successfully apply modular, reusable code practices. The paper introduces the interface and structure of this repository, presents templates for RBBs documentation, provides tips to support aspiring users, and first examples. We highlight the need for alternative RBB implementations that share the same generic description. We also acknowledge that RBBs might represent different levels of interactions, starting from decisions concerning a single agent to interactions between multiple agents or agents and their environment. While initially designed to assist agent-based community, the platform can be utilized by other modelers (e.g. system dynamics, integrated assessment, equilibrium) who seek to improve the representation of human behavior, micro-level processes, heterogeneity, interactions, learning, and other complex dynamics. Naturally, the platform is only one element in the chain towards a successful adoption of best software development practices like RBBs. Future work should focus on populating the repository, refining review processes, and systematizing the variety of RBBs' implementations including engagement with domain experts. Following this initial phase, we hope to further support technical improvements of the platform and widen its impact in and beyond the agent-based community.
To mitigate the spread of contagious diseases, there is an ongoing discussion surrounding interventions that strategically target individuals who, due to their social network position, are responsible for more infections than others. However, the practical identification of these individuals using conventional network metrics is considerably challenging due to the lack of required data. A potential remedy to this quandary is the development of easily observable proxy metrics for measuring influential spreading. This study aims to assess the viability of such an approach using the example of contact tracing. Utilizing an empirically calibrated agent-based model, the study investigates the extent to which the efficacy of contact tracing can be enhanced by prioritizing influential spreaders, identified using age and household size as proxy measures. The results reveal that the effectiveness of contact tracing is significantly influenced by whose contacts are traced. When the contacts of those causing the most infections are traced, it can substantially enhance the efficacy of contact tracing, even when they are identified solely based on proxy metrics such as age and household size. For the examined case of the German state of Baden-W & uuml;rttemberg, it appears that middle-aged individuals residing in larger households are responsible for most infections. Therefore, prioritizing contact tracing for this specific demographic group seems to be a robust strategy to improve contact tracing. Overall, the results support the potential of the proposed approach to reduce the overall societal costs of non-pharmaceutical interventions while increasing their impact. Further empirical testing of the approach appears worthwhile.
