Robots that Learn Games by Asking People Questions
The study goal is to develop robots that can learn simple games by interviewing humans. The robot interview strategies are represented as Markov Decision Process policies, and are developed offline through robot interactions with simulated interviewees. The investigator can set the simulated interviewee to respond to the robot's queries with different amounts of information. The games include board games like Connect 4, Quarto, and Checkers. The communication between the robot and simulated interviewee is in a formal language like first order logic. After a simulated dialogue, the robot will have more or less knowledge of the game, depending on the conditions of the simulated interviewee. Human subjects will interact with the robots in two ways. First, to test the utility of the new knowledge, human subjects will participate in experiments where they play the game with virtual robots that have learned the game under different conditions. Second, to investigate the kinds of misunderstandings of questions and answers that might arise if the robot could interview a human, we will present subjects with extracts from the formal language dialogues that we have mapped to English in different ways, and ask subjects to imagine how they would express the same meaning in English, or answer questions in different ways, or what sorts of misunderstandings they might imagine in a real dialogue.
Ability to communicate in written and spoken English
Able to give informed consent
Unable to communicate in written and spoken English
Unable to give informed consent