The tragedy of the commons in public resource governance is essentially the result of repeated interaction and adaptive learning among heterogeneous agents under dynamic resource constraints. Existing studies have generated rich insights into common-pool resource governance, institutional constraints, and cooperation, but they still rely mainly on theoretical deduction, static games, or econometric identification, making it difficult to jointly characterize resource dynamics, agent heterogeneity, behavioral learning, and policy scenarios. From the perspective of computational economics, this paper develops a multi-agent reinforcement learning simulation model for the governance of the tragedy of the commons. Specifically, the fish-pond resource system is formulated as a Markov decision process in discrete states, and adaptive decision-making under pure economic incentives, sustainability penalties, behavioral heterogeneity, and cooperation is characterized through the resource dynamics equation, harvesting equation, and differentiated reward functions. The model is further examined through sensitivity analysis, parameter calibration, and theoretical validation, and then used for policy simulation. The results show that pure economic incentives quickly induce resource collapse, sustainability penalties significantly reduce harvesting intensity and maintain a low but sustainable steady state, heterogeneous behavior parameters generate clear strategic divergence, and cooperation internalizes group harvesting constraints into individual payoffs and yields the strongest resource recovery and behavioral convergence.

The aim of this model is to study the dynamic propagation of individual climate adaptive behaviours in different scenarios within the analytical framework of conservation motivation theory, focusing on the impact of social and experiential learning on the adoption of climate adaptive behaviours by coastal farmers.
Model for paper “Promoting climate resilience through learning-based behavioural change: Insights from an agent-based model of a coastal farming community in Guangxi, China” in Environmental Science & Policy, Volume 179, May 2026, 104375, https://doi.org/10.1016/j.envsci.2026.104375

This model visualizes gradient descent optimization - the fundamental algorithm used to train neural networks and other machine learning models. Agents represent different optimization algorithms searching for the minimum of a loss landscape (the “error surface” that ML models try to minimize during training).

The model demonstrates how different optimizer types (SGD, Momentum with different parameters) behave on various loss landscapes, from simple bowls to the notoriously difficult Rosenbrock “banana valley” function. This helps build intuition about why certain optimization algorithms work better than others for different problem geometries.

HOW IT WORKS

…

A dynamic identity model for misinformation in social networks, an agent-based model of social identity and misinformation dynamics.

I developed this model as a part of my master’s thesis, “Does social identity drive belief and persistence in online misinformation? An agent-based modelling approach” at University College Dublin, Ireland (2024-2025).

The purpose of this model is to further understand the dynamics of misinformation sharing as an expression of social identity. I introduce a framework to understand the influence of self-categorisation on misinformation persistence in social network. It integrates a social learning model with the Dynamic Identity Model for Agents (DIMA) using simple logic to simulate the social trade-offs driving misinformation and observe the effects on misinformation spread.

This Agent-Based Model is designed to simulate how similarity-based partner selection (homophily) shapes the formation of co-offending networks and the diffusion of skills within those networks. Its purpose is to isolate and test the effects of offenders’ preference for similar partners on network structure and information flow, under controlled conditions.

In the model, offenders are represented as agents with an individual attribute and a set of skills. At each time step, agents attempt to select partners based on similarity preference. When two agents mutually select each other, they commit a co-offense, forming a tie and exchanging a skill. The model tracks the evolution of network properties (e.g., density, clustering, and tie strength) as well as the spread of skills over time.

This simple and theoretical model does not aim to produce precise empirical predictions but rather to generate insights and test hypotheses about the trade-off between partnership stability and information diffusion. It provides a flexible framework for exploring how changes in partner selection preferences may lead to differences in criminal network dynamics. Although the model was developed to simulate offenders’ interactions, in principle, it could be applied to other social processes involving social learning and skills exchange.

…

The wisdom of the crowd refers to the phenomenon in which a group of individuals, each making independent decisions, can collectively arrive at highly accurate solutions—often more accurate than any individual within the group. This principle relies heavily on independence: if individual opinions are unbiased and uncorrelated, their errors tend to cancel out when averaged, reducing overall bias. However, in real-world social networks, individuals are often influenced by their neighbors, introducing correlations between decisions. Such social influence can amplify biases, disrupting the benefits of independent voting. This trade-off between independence and interdependence has striking parallels to ensemble learning methods in machine learning. Bagging (bootstrap aggregating) improves classification performance by combining independently trained weak learners, reducing bias. Boosting, on the other hand, explicitly introduces sequential dependence among learners, where each learner focuses on correcting the errors of its predecessors. This process can reinforce biases present in the data even if it reduces variance. Here, we introduce a new meta-algorithm, casting, which captures this biological and computational trade-off. Casting forms partially connected groups (“castes”) of weak learners that are internally linked through boosting, while the castes themselves remain independent and are aggregated using bagging. This creates a continuum between full independence (i.e., bagging) and full dependence (i.e., boosting). This method allows for the testing of model capabilities across values of the hyperparameter which controls connectedness. We specifically investigate classification tasks, but the method can be used for regression tasks as well. Ultimately, casting can provide insights for how real systems contend with classification problems.

Negotiation plays a fundamental role in shaping human societies, underpinning conflict resolution, institutional design, and economic coordination. This article introduces E³-MAN, a novel multi-agent model for negotiation that integrates individual utility maximization with fairness and institutional legitimacy. Unlike classical approaches grounded solely in game theory, our model incorporates Bayesian opponent modeling, transfer learning from past negotiation domains, and fallback institutional rules to resolve deadlocks. Agents interact in dynamic environments characterized by strategic heterogeneity and asymmetric information, negotiating over multidimensional issues under time constraints. Through extensive simulation experiments, we compare E³-MAN against the Nash bargaining solution and equal-split baselines using key performance metrics: utilitarian efficiency, Nash social welfare, Jain fairness index, Gini coefficient, and institutional compliance. Results show that E³-MAN achieves near-optimal efficiency while significantly improving distributive equity and agreement stability. A legal application simulating multilateral labor arbitration demonstrates that institutional default rules foster more balanced outcomes and increase negotiation success rates from 58% to 98%. By combining computational intelligence with normative constraints, this work contributes to the growing field of socially aware autonomous agents. It offers a virtual laboratory for exploring how simple institutional interventions can enhance justice, cooperation, and robustness in complex socio-legal systems.

This base model uses an agent-based approach to represent heterogeneous farmers’ trading partners selection among multiple recipients (other farmers, village collectives, and firms). Each period, a potential transfer-out farmer decides whether to transfer based on a net-return versus transaction-cost trade-off; if transferring, the farmer selects the counterparty with the highest expected profit. Meanwhile, social learning—operationalized as logistic accumulation of neighborhood experience—continuously updates uncertainty, which in turn shapes transaction costs and subsequent decisions.

This agent-based model simulates the lifecycle, movement, and satisfaction of teachers within an urban educational system composed of multiple universities and schools. Each teacher agent transitions through several possible roles: newcomer, university student, unemployed graduate, and employed teacher. Teachers’ pathways are shaped by spatial configuration, institutional capacities, individual characteristics, and dynamic interactions with schools and universities. Universities are assigned spatial locations with a controllable level of centralization and are characterized by academic ratings, capacity, and alumni records. Schools are distributed throughout the city, each with a limited number of vacancies, hiring requirements, and offered salaries. Teachers apply to universities based on the alignment of their personal academic profiles with institutional ratings, pursue studies, and upon graduation become candidates for employment at schools.
The employment process is driven by a decentralized matching of teacher expectations and school offers, taking into account factors such as salary, proximity, and peer similarity. Teachers’ satisfaction evolves over time, reflecting both institutional characteristics and the composition of their colleagues; low satisfaction may prompt teachers to transfer between schools within their mobility radius. Mortality and teacher attrition further shape workforce dynamics, leading to continuous recruitment of newcomers to maintain a stable population. The model tracks university reputation through the academic performance and number of alumni, and visualizes key metrics including teacher status distribution, school staffing, university alumni counts, and overall satisfaction. This structure enables the exploration of policy interventions, hiring and training strategies, and the impact of spatial and institutional design on the allocation, retention, and happiness of urban educational staff.

ABM model studying impact of social cohesion on wellbeing of a society. Ibn Khaldun’s cyclical theory of history is being used as the theoretical lens along with some other theories. Social cohesion is measured as TSC = (TVE + 2 * (TPI * TPL - TNI * TNL))/((TPI+TNI))
Where
TSC total-social-cohesion ; Variable for social cohesion
TPI total-positive-interactions ; Count of positive interactions
TNI total-negative-interactions ; Count of negative interactions
TPL total-positive-learning ; Count of positive learning outcomes
…