The core algorithm is an agent-based model, which simulates travel patterns on a network based on microscopic decision-making by each traveler.

SONG is a simulator designed for simulating the process of transportation network growth.

A model of innovation diffusion in a structured population with two groups who are averse to adopting a produce popular with the outgroup.

The Non-Deterministic model of affordable housing Negotiations (NoD-Neg) is designed for generating hypotheses about the possible outcomes of negotiating affordable housing obligations in new developments in England. By outcomes we mean, the probabilities of failing the negotiation and/or the different possibilities of agreement.
The model focuses on two negotiations which are key in the provision of affordable housing. The first is between a developer (DEV) who is submitting a planning application for approval and the relevant Local Planning Authority (LPA) who is responsible for reviewing the application and enforcing the affordable housing obligations. The second negotiation is between the developer and a Registered Social Landlord (RSL) who buys the affordable units from the developer and rents them out. They can negotiate the price of selling the affordable units to the RSL.
The model runs the two negotiations on the same development project several times to enable agents representing stakeholders to apply different negotiation tactics (different agendas and concession-making tactics), hence, explore the different possibilities of outcomes.
The model produces three types of outputs: (i) histograms showing the distribution of the negotiation outcomes in all the simulation runs and the probability of each outcome; (ii) a data file with the exact values shown in the histograms; and (iii) a conversation log detailing the exchange of messages between agents in each simulation run.

The “Descriptive Norm and Fraud Dynamics” model demonstrates how fraudulent behavior can either proliferate or be contained within non-hierarchical organizations, such as peer networks, through social influence taking the form of a descriptive norm. This model expands on the fraud triangle theory, which posits that an individual must concurrently possess a financial motive, perceive an opportunity, and hold a pro-fraud attitude to engage in fraudulent activities (red agent). In the absence of any of these elements, the individual will act honestly (green agent).

The model explores variations in a descriptive norm mechanism, ranging from local distorted knowledge to global perfect knowledge. In the case of local distorted knowledge, agents primarily rely on information from their first-degree colleagues. This knowledge is often distorted because agents are slow to update their empirical expectations, which are only partially revised after one-to-one interactions. On the other end of the spectrum, local perfect knowledge is achieved by incorporating a secondary source of information into the agents’ decision-making process. Here, accurate information provided by an observer is used to update empirical expectations.

The model shows that the same variation of the descriptive norm mechanism could lead to varying aggregate fraud levels across different fraud categories. Two empirically measured norm sensitivity distributions associated with different fraud categories can be selected into the model to see the different aggregate outcomes.

MASTOC-LLM extends the classic Multi-Agent System Tragedy of the Commons (MASTOC) model by replacing hard-coded behavioral rules with autonomous decision-making powered by large language models (LLMs). Three heterogeneous agents manage herds of cows on a shared grassland commons. Each tick, an agent receives a structured prompt describing current resource levels, its own herd size, peer behavior, and — optionally — a rolling memory of recent rounds and messages from neighboring agents. The LLM returns a stocking decision (add, remove, or hold cows) together with a natural-language rationale and, when communication is enabled, a short message to broadcast to peers.

The model is designed to test whether LLM agents spontaneously develop Ostrom-style common-pool resource governance (mutual monitoring, graduated sanctions, graduated rule revision) or instead fall into identifiable failure modes. Preliminary experiments with Claude Haiku 4.5, GPT-5.4-mini, and DeepSeek R1:32b have revealed four recurring collapse patterns — Cooperative Paralysis, Defection Cascade, Overshoot-Panic, and Hybrid Architecture Failure — whose onset timing is sensitive to memory length, inter-agent communication, and the post-training alignment approach of the underlying model.

MASTOC-LLM is intended as a laboratory for generative agent-based modelling (GABM) methodology: it provides a clean, well-understood commons baseline against which LLM behavioral hypotheses can be systematically tested and compared across models, parameter sweeps, and alignment regimes.

Is the mass shooter a maniac or a relatively normal person in a state of great stress? According to the FBI report (Silver, J., Simons, A., & Craun, S. (2018). A Study of the Pre-Attack Behaviors of Active Shooters in the United States Between 2000 – 2013. Federal Bureau of Investigation, U.S. Department of Justice,Washington, D.C. 20535.), only 25% of the active shooters were known to have been diagnosed by a mental health professional with a mental illness of any kind prior to the offense.
The main objects of the model are the humans and the guns. The main factors influencing behavior are the population size, the number of people with mental disabilities (“psycho” in the model terminology) per 100,000 population, the total number of weapons (“guns”) in the population, the availability of guns for humans, the intensity of stressors affecting humans and the threshold level of stress, upon reaching which a person commits an act of mass shooting.
The key difference (in the model) between a normal person and a psycho is that a psycho accumulates stressors and, upon reaching a threshold level, commits an act of mass shooting. A normal person is exposed to stressors, but reaching the threshold level for killing occurs only when the simultaneous effect of stressors on him exceeds this level.
The population dynamics are determined by the following factors: average (normally distributed) life expectancy (“life_span” attribute of humans) and population growth with the percentage of newborns set by the value of the TickReprRatio% slider of the current population volume from 16 to 45 years old.Thus, one step of model time corresponds to a year.

The agent-based model WEEM (Woodlot Establishment and Expansion Model) as described in the journal article, has been designed to make use of household socio-demographics (household status, birth, and death events of households), to better understand the temporal dynamics of woodlot in the buffer zones of Budongo protected forest reserve, Masindi district, Uganda. The results contribute to a mechanistic understanding of what determines the current gap between intention and actual behavior in forest land restoration at farm level.

This is an agent-based model, simulating wolf (Canis Lupus) reappearance in the Netherlands. The model’s purpose is to allow researchers to investigate the reappearance of wolves in the Netherlands and the possible effect of human interference. Wolf behaviour is modelled according to the literature. The suitability of the Dutch landscape for wolf settlement has been determined by Lelieveld (2012) [1] and is transformed into a colour-coded map of the Netherlands. The colour-coding is the main determinant of wolf settlement. Human involvement is modelled through the public opinion, which varies according to the size, composition and behaviour of the wolf population.

[1] Lelieveld, G.: Room for wolf comeback in the Netherlands, (2012).

This model simulates diffusion curves and it allows to test how social influence, network structure and consumer heterogeneity affect their spreads and their speeds.