Agent-based Modeling for Policy-Making in Inequity Contexts

1 Research Problem andMotivation

In recent years, research in Artificial Intelligence (AI) has led to giant leaps forward in the field of agent-based simulations. These simulations have huge potential to be used as tools for policy-making, as their architecture allows to represent both bottom-up and top-down effects of policies in a non-invasive way. They have been applied in diverse social contexts where political action can, and should, drive meaningful change, such as in pandemics or housing crises.

Our research is centered around scenarios where inequity arises due to the presence (or absence) of legal and social norms. For agent-based models to be informative in these scenarios, they must accurately represent reality. This requires defining representative (i) agent profiles, (ii) agent decisionmaking, and (iii) agent environment. Our project aims to incorporate these three elements consistently to guide the evaluation of alternative policy approaches that aim to mitigate inequity, working towards the UN SDGs.

For the first element (i), agent profiles, population synthesis techniques, such as Bayesian networks, are employed. The second one (ii), agent decision-making, is defined as being driven by needs and values. We start with the needs-based model [Dignum, 2021] and then improve this by relying on a Markov decision process (MDP) framework with more complex motivators of behavior. For the third (iii), agent environment, we distinguish between the physical environment (e.g. discrete grids), and the regulatory environment (e.g. legal policies modelled with institutional syntax [Frantz and Siddiki, 2024]). Both elements constrain the agents' behavior in some way and, thus, the simulation's outcome.

2 State of the Art and Contributions

This research builds upon prior work in social simulations for policy evaluation, particularly [Montes et al., 2023]'s research proposal. Their work explores how an ABM could be used to assess the impact that aporophobia [Cortina, 2017], represented within the legal policies, has on poverty levels. It adapts the needs-based model from [Dignum, 2021] to the socioeconomic problem under consideration (poverty). Our research introduces several key contributions:

Proof-of-concept Model. In this proof-of-concept [Aguilera et al., 2024], (i) agent profiles are sampled from macro data, (ii) agent behavior is driven by needs, and (iii) the physical and regulatory environments are modeled using a simplified grid structure and a set of norms represented through preconditions and postconditions. This study provides innovative insights into how agent-based models can incorporate normative frameworks to evaluate the consequences of discrimination on vulnerable populations. However, technical and conceptual refinements are needed for this model to be informative in the real world.

Motivational Population Synthesis. In [Aguilera et al., 2025], we present a novel population synthesis method to create synthetic populations that account for cultural variation, using Bayesian networks. This approach integrates motivational attributes, such as human values and needs, with demographic ones, such as gender, age or employment status. In this way, we can use these motivators of behavior, aligned with the rest of the agent's profile, as input for the decisionmaking mechanisms in agent-based simulations.

CA-based Simulation Framework. Our latest research introduces a novel and enhanced agent-based simulation framework for policy-making. It is rooted in the Capability Approach (CA), a widely recognised framework in human development studies, used for analysing and assessing human well-being, development and social justice. Rather than looking only at resources or outcomes, it focuses on the real opportunities (capabilities) people have to lead the lives they value, such as being able to be adequately sheltered or being able to undertake employment. Policy interventions can then help restore these capabilities or even expand them.

In our framework, agent-based simulations are the tool to evaluate the impact of legal norms on people's capabilities, helping to identify the optimal ones in terms of mitigating inequity. We define and assess equity and inequity in terms of having (or lacking) central capabilities [Nussbaum, 2011], among other elements of value, in line with the CA. Beyond evaluation, we pay particular attention to the agents' decision-making. Our model defines behavior not only in terms of what people value or need to do but also in terms of what people can actually do given their circumstances. The CA emphasizes that our action space is constrained by resources and conversion factors, i.e. there are things we cannot do because we lack the means or the personal, social and environmental circumstances to do them. Capabilities provide this counterfactual information that is normally overlooked in decision-making architectures: they represent the set of an individual's impossible and possible actions. The choice of which possible action we realize in a particular instant of time, depends on the individual choice factors.

In Fig. 1, the key elements of the CA are mapped to the key elements of an ABM (agent profiles, decision-making, regulatory environment, etc.) and an MDP (states, actions, transition probabilities, rewards) to structure the simulation and define decision-making on a reinforcement learning basis. With this integration, we aim to enhance the modeling of human behavior and contribute to prior research into the computational representation of human values [Osman and d'Inverno, 2024]. The framework is currently being developed for a real-world case study on health inequity.