We consider the realistic case of a GOLOG agent that only possesses incomplete knowledge about the state of its environment and has to resort to sensing in order to gather additional information at runtime, and where the agent is controlled by a knowledge based program in which test conditions explicitly refer to the agent's knowledge (or lack thereof). In this paper, we propose a formalization of knowledge-based agents that extends earlier proposals by a form of non-monotonic reasoning that includes Reiter-style defaults. We present a reasoning mechanism that enables us to reduce projection queries about future states of the agent's knowledge (including nesting of epistemic modalities) to classical Default Logic, and provide a corresponding Representation Theorem. We thus obtain the theoretical foundation for an implementation where reasoning subtasks can be handed to an embedded off-the-shelf reasoner for Default Logic, and that supports a (in some respects) more expressive epistemic action language than previous solutions.
展开▼