Towards Situated Activity Management: Representation, Inference and Decision Making

MSR-TR-2015-64 |

Published by Microsoft

We introduce and discuss the problem of situated activity management, and present an approach for guiding and coordinating an agent’s speech and activities amidst the dynamics of evolving settings. Specifically, we couple a hierarchical representation of activities with a novel state tracking approach based on conditional Markov Networks. We show how the approach can enable an agent to reason jointly about parallel coordinated actions and the changing situational context. The hierarchical structure and joint inference allows for a modular authoring of rules used for making in-stream decisions. We have implemented a proof-of-concept for this approach and illustrate its functionality with a simulated dialog trace.