Do You Trust the Process?: Modeling Institutional Trust for Community Adoption of Reinforcement Learning Policies

TL;DR

The paper tackles how to embed institutional trust into reinforcement learning for policy-like resource allocation in communities. It proposes three models—trust-unaware, trust-aware, and learned-trust—implemented via a DDPG framework and evaluated on simulated social networks with a budget constraint $\rho$ and a resource-tradeoff parameter $c$. Key findings show that learned-trust with Bayesian trust estimates often improves fairness and average trust, especially when resource-saving pressures are high, albeit with reduced organizational success; trust-aware approaches can boost outcomes in some settings but may worsen fairness under certain parameters. To safeguard communities, the authors examine a quota intervention by external entities, which can improve fairness and trust in some cases while reducing organizational utility, highlighting a tension between public welfare and organizational objectives. The work underscores the importance of incorporating trust dynamics into algorithmic governance and provides a framework for predictive, fair, and ethically aware RL policies in humanitarian contexts.

Abstract

Many governmental bodies are adopting AI policies for decision-making. In particular, Reinforcement Learning has been used to design policies that citizens would be expected to follow if implemented. Much RL work assumes that citizens follow these policies, and evaluate them with this in mind. However, we know from prior work that without institutional trust, citizens will not follow policies put in place by governments. In this work, we develop a trust-aware RL algorithm for resource allocation in communities. We consider the case of humanitarian engineering, where the organization is aiming to distribute some technology or resource to community members. We use a Deep Deterministic Policy Gradient approach to learn a resource allocation that fits the needs of the organization. Then, we simulate resource allocation according to the learned policy, and model the changes in institutional trust of community members. We investigate how this incorporation of institutional trust affects outcomes, and ask how effectively an organization can learn policies if trust values are private. We find that incorporating trust into RL algorithms can lead to more successful policies, specifically when the organization's goals are less certain. We find more conservative trust estimates lead to increased fairness and average community trust, though organization success suffers. Finally, we explore a strategy to prevent unfair outcomes to communities. We implement a quota system by an external entity which decreases the organization's utility when it does not serve enough community members. We find this intervention can improve fairness and trust among communities in some cases, while decreasing the success of the organization. This work underscores the importance of institutional trust in algorithm design and implementation, and identifies a tension between organization success and community well-being.

Figures (5)

• Figure 1: Organization success (or utility $\mathcal{U}$) (a & d), global fairness of resource allocation (b & e), and average individual trust (c & f) for our trust-aware and unaware RL algorithms. The trends for each algorithm are quite similar---high $c$ values, or high desire to save resources, lead to success for the RL agent, but lost trust. For low $c$ values, higher initial average trust allows the RL agent to learn a more successful policy that positively impacts the community, maximizing fairness as well as trust. Notice that the scale for average trust is different than the other metrics (see color bar).
• Figure 2: Raw difference for each metric between trust-aware and trust-unaware RL algorithms. Green cells represent the trust-aware algorithm outperforming trust-unaware, pink is the opposite, and white is neutral. Our trust-aware algorithm most outperforms unaware in fairness and average trust when $c = 0.5$ and $\overline{\tau} \geq 0.67$. However, this boost in fairness and trust comes at the cost of organization success. For other $c$ values, we find that the additional certainty does not help performance, or has inconsistent effects.
• Figure 3: We show social networks for select iterations of the trust-aware algorithm after services have been provided. We fix $\overline{\tau} = 0.8$ for ease of comparison. The color of node $v$ indicates the updated $\tau_v$ value after services have been provided; bluer nodes have higher values. The size indicates the accumulated utility $u_v$; larger nodes have accepted more services. For high values of $c$, community members do not receive any services; as $c$ decreases, service values increase (a). Isolated nodes tend to be more trusting, as they always perceive perfect fairness in their neighborhoods.
• Figure 4: Raw difference for each metric between learned-trust and trust-unaware RL algorithms. Green cells represent the learned-trust algorithm outperforming trust-unaware, pink is the opposite, and white is neutral. Notably, the learned-trust model does not produce a sharp dip in fairness (and therefore trust) at $c=0.5$; this is due to the more equitable resource distributions learned by the agent. In most cases, the learned-trust model dominates trust-unaware for fairness and trust, but organization success suffers.
• Figure 5: Raw difference for each metric between our intervention and trust-aware RL algorithm. Green cells represent the interventions outperforming trust-aware, pink is the opposite, and white is neutral. The organization's utility $\mathcal{U}$ suffers for high $c$ values, as does fairness and average community member trust. For mid to low $c$ values, fairness and average trust of agents increase.