Constraint-directed improvisation for everyday activities

Existing approaches to planning in Artificial Intelligence (such as Universal and Classical Planning) are designed for very specific types of activities, and are largely inapplicable to areas outside their narrow ranges. In particular, everyday activities that are simple for humans, such as making a meal or getting from place to place, require long-term goal-directed and timely responses that are far beyond the bounds of these traditional approaches. This dissertation examines the nature of the everyday activities and develops a computational architecture for an agent able to participate in such activities. An analysis of everyday activities shows them to be difficult tasks made artificially simple through extensive activity-specific knowledge possessed by the agent performing them. I argue that existing approaches are unsuitable to everyday activity because they rely too heavily on compiled knowledge and fail to adequately apply the background knowledge from which these compilations were originally made. To address everyday activities, I present a theory of improvisation, a new approach that views the problem as satisficing intelligent control: providing resource-bounded responses to the environment in light of the agent's previous experience and its current and future intentions for activity. This process is based on the use of both heavily compiled routines the agent is accustomed to following, and an extensive collection of background knowledge used to apply those routines flexibly. The agent can rely on its routines in normative situations or when time is too scarce to spend examining the reasons behind its routines, and can conversely rely more heavily on background knowledge as situations become less normative. This allows the agent to take advantage of regularities in its environment and respond flexibly in less familiar situations. I then present an architecture embodying the improvisational approach based on the use of constraint-directed reasoning. This methodology provides a flexible control mechanism that allows the agent to respond as dynamically as necessary for the circumstances in which it finds itself. Implemented examples of improvised behaviour are also shown, using a simulation tool developed in conjunction with this research.