Trustworthy Decision Making via Commitments

Existing approaches to calculate trust between agents rely on the strength of their relationships based on their degrees of friendship, the frequencies of their interactions, the sentiments extracted from their interactions, and so on. These approaches rely heavily on numerical measures and disregard the deep structure underlying trust. By contrast, we establish the idea of trust among agents based on their normative relationships (norms in short). Norms provide the standard of correctness of the interactions between agents and provide a basis for accountability for their actions. We focus on commitments to capture relationships between agents. Based on commitments, we provide a framework using which we can calculate trust probabilistically between agents. We evaluate our approach in two parts. First, we specified interactions between agents and asked subjects to estimate the level of trust between each pair of agents. Second, we selected over 5,000 email sentences from the well-known Enron dataset and asked subjects to estimate the level of trust between each pair of participants. We learned subjects’ trust parameters from their estimation. We recalculated the trust values between participants in datasets using these parameters. We found that the mean absolute error (MAE) between trust values calculated from learned parameters and estimates provided by subjects is minimum when compared against trust values calculated using any fixed parameter.