mobile icon

Perception of agents' skill

WorkgroupPerception and Action Lab
FundingIWM budget resources
Project description

The human environment is populated by other agents. Other humans make up an important part of this, but animals and increasingly autonomously acting machines, i.e., artificial agents, also act without explicit prompting. As a result, their actions are sometimes incomprehensible. This hinders joint work, which is meant to be facilitated by automation. An important aspect of any interaction with other agents is understanding the other. This includes assessing the capabilities of the other person. In a cooperative situation, it is important to be able to assess how much you can rely on your partner. In a competitive situation, you need to assess in which aspects you are superior.

In this project, we will look at the dynamics of an observer and an agent in different experiments. We think that the ability of an observer to estimate the skill of an agent in a specific task is constrained by the skill of the observer in that task. For example, a chess master will find it easier to estimate the level of any student than a novice. This statement has implications not only for interpersonal interactions. It also affects interactions with artificial agents. One problem many people and scientists have with artificial agents is the inexplicability of their decisions. However, with human agents, it is naturally accepted that decisions are made by specialists through decision-making processes that a layperson cannot comprehend.

The experiments in this project address the interaction of humans with artificial agents. In one series of experiments, a cooperative game was used to test the ability of humans to predict the behavior of an artificial partner as a function of the task understanding of the human and his agent partner. Another series of studies addresses the previous step. To anticipate behavior from others, one must first evaluate the agent's task understanding. For this purpose, in another experiment, the task comprehension of experimental participants was manipulated, and their ability to evaluate the task comprehension of an agent was tested.


Prof. Dr. Holger Brandt, University of Tübingen

Prof. Dr. Augustin Kelava, University of Tübingen