Principled Reasoning about Dynamical Systems

点击翻译按钮获取中文摘要

Dynamic phenomena are found in a broad range of contexts from discrete event control in technical systems, to business processes, to atoms changing bonds in chemical reactions. Despite their diversity, these dynamic phenomena often have common conceptualization either in terms of discrete transitions affected by actions, or in terms of hybrid systems where there are continuous processes and flows that can be initiated or terminated by discrete actions or events.  What is important these phenomena can be described using common underlying principles. These principles should be formally represented in a mathematical language that facilitates their analysis. This helps to design general solutions that can be subsequently deployed in a variety of applications. The proposed research program contributes to developing general principled representations for actions and their effects, and to demonstrating how these representations can be used to perform computationally tractable reasoning about the direct and indirect effects of actions. There are several conceptual and computational challenges that prevent the existing principled representations and reasoning mechanisms from making practical contributions to solving real-world problems. The proposed research will address some of these remaining challenges. The objective of the proposed research program is  advance our knowledge about specialized reasoning mechanisms that can lead to development of efficient domain independent techniques for solving problems in dynamical systems. This include identifying the kinds of dynamical systems that frequently occur in practical applications and investigating the use of principled logical representations to model these systems. The proposed research will focus on application of specialized reasoning mechanisms to solve large scale planning problems. In particular, there is a need to explore how lifted representations can be used to solve deterministic planning problems when initial knowledge is incomplete. Moreover, I will explore the question how lifted heuristic planning can be done when actions have indirect effects. This is important since in some realistic domains preconditions of actions can be defined using nested abbreviations whose truth values change indirectly when actions are executed. The proposed research will concentrate on the cases where computationally tractable, domain independent techniques can be developed.  Also, I will explore how can we correctly and efficiently determine the actual causes of an observed effect. This task includes finding  not only the primary actual cause, but also the whole causal chain including the root cause of an effect. The outcome of this research will be advancement of our knowledge and the development of techniques with potentially broad applicability.