RESCHEDULED: Planning and Explanatory Dialogue with Temporal Logic Norms
Even while performing such mundane activities as grocery shopping, humans follow a wide variety of moral and social norms. While learning and obeying these norms comes naturally to us, it remains a challenge to imbue artificial agents with such norms. This challenge is enhanced by the need for such systems to explain or account for their behaviors in terms of these norms, and to accept correction from the humans with whom they interact. In this dissertation, we develop an approach whereby artificial agents can plan to maximally satisfy a set of norms according to a mixed weighted-lexicographic preference structure, even when these norms conflict. These norms are specified in an object-oriented temporal logic, supporting a broad variety of interpretable norms. We provide an approach for our planning agent to explain its behaviors by answering "why" queries (also specified in temporal logic), where the resulting explanations reference the agent's norms. Leveraging the natural language pipeline of the DIARC architecture scheme, we build on this explanation module to allow an agent to engage in normative explanatory dialogue with a human, in which a human may ask the agent questions in natural language and receive natural language responses, as well as enabling correction by direct modification of the agent's norms. Noting the lack of scalability of our approach to complex environments, we develop ProperShopper, a multi-agent grocery store simulation with a large state space deliberately designed to challenge our approach. Finally, we develop an agent which builds off our planning approach and is capable of satisfying norms in ProperShopper. This approach builds on the DIARC architecture scheme, and plans to obey norms at two levels: an abstract level leveraging using numeric planning, and a more concrete level using MCTS to reactively avoid norm violations. Though the agent is not yet fast enough to coexist with human participants in ProperShopper, it indicates a path toward developing agents which can learn norms through interaction with humans, obey those norms in complex environments, and honestly explain their behavior with respect to these norms.
Please join meeting via Zoom: https://tufts.zoom.us/j/99072545984
Password: see colloquium email
Dial-in option not available for this event.