Improving Confidence in the Estimation of Values and Norms

Autonomous agents (AA) will increasingly be interacting with us in our daily lives. While we want the benefits attached to AAs, it is essential that their behavior aligned values and norms. Hence, an AA need estimate norms of humans interacts with, which not a straightforward task when solely observing agent's behavior. This paper analyses what extent able simulated human agent (SHA) based on its actions ultimatum game. We present two methods reduce ambiguity profiling SHAs: one search space exploration another counterfactual analysis. found both are increase confidence estimating norms, but differ applicability, latter being more efficient number interactions minimized. These insights useful improve alignment AAs