Asking Questions and Developing Trust

No. 1: Agents that Learn from Human Teachers

Volume

Papers from the 2009 AAAI Spring Symposium

Contents

In a number of domains, researchers instrument an interface or the environment with software and hardware sensors to collect observational data, but it is often quite hard to label that data accurately. We are interested in how an agent can ask many humans at once for help to label its data and we present two studies towards this goal. First, we investigate how a computer can automatically elicit labels from users as they interact with different technologies. Then, we present a study comparing different algorithms for how an agent decides which users to trust when a lot of people answer the agent's questions and the answers are conflicting. We discuss the implications of the results in each of these studies and present some ideas for future work towards agents asking humans questions.