Reasoning and acting in time.

摘要

This dissertation investigates aspects of the design of an embodied cognitive agent that interleaves reasoning, acting, and interacting with other agents while maintaining a record of what has happened and is happening in its environment. Among other things, such knowledge of its own history allows the agent to reason more effectively when faced with an emergency situation such as the failure of one of its acts. Crucial to distinguishing what has happened from what is happening is a notion of the present time, or “now”. There has been much research in artificial intelligence on issues of time. However, it is typically the case that, although an agent reasons about time, it is not itself situated in time. Once the agent has a concept of “now”, one that continuously changes to reflect the progression of time, a different kind of temporal reasoning problem emerges. For example, given that an agent could be told anything, how are formal representations of present states to be distinguished from those of past states, where the former, but not the latter, are pertinent to the agent's actions? And, given that the present must be distinguished, how is “now” to be modeled so that the mere passage of time does not result in computationally costly knowledge-base-editing routines? How can the agent reason about “now”, when the very process of reasoning results in “now” changing? How can the agent be endowed with a feel for how much time has passed, which seems crucial for reasoning about persistence as time passes while the agent acts?; In this dissertation, the above issues are investigated in detail. A theory of subjective time is presented, accounting for a cognitive agent's vague concept of “now”, its sense of temporal progression, and its feel for how much time has passed. An investigation of the impact of embodiment and time perception on issues of reasoning about persistence as the agent acts comes out as a natural by-product of the theory. The theory of subjective time is wrapped around a core logic of objective time that axiomatizes various aspectual phenomena needed for reasoning, acting, and natural language interaction. Unlike most theories of linguistic aspect, the logic of aspect presented here accounts for the agent's knowledge of states, events, and processes, not only from static natural language inputs, but also from the dynamic accumulation of perceptual and proprioceptual information as events unfold in time. Based on the logical analysis of the notion of telicity (the analysis goes beyond the standard telic/atelic distinction), a theory of how cognitive agents may employ reasoning to control the execution of sequences of acts is presented. The theory proposes a principled way by which agents may decide when it is time to move on to the next step in a sequence; an issue that has not been given much attention in the literature. Finally, the dissertation establishes a framework for interrupt-handling and error recovery based on a system of context-sensitive priorities among acts.