Pragmatic Approaches to Reasoning Under Uncertainty (Computer and Information Science)

点击翻译按钮获取中文摘要

This research investigates pragmatic approaches to computer-based probabilistic reasoning systems. In the past, artificial intelligence researchers have often avoided probability theory for reasoning with uncertainty because of a common perception that the application of probability theory is invariably associated with a commitment to intractable algorithms and inordinate knowledge-acquisition time. The development of efficient algorithmic and assessment techniques would allow investigators to apply a theoretically justified model of belief entailment to complex problems. As a result, the goals in this research are (1) to develop techniques for using knowledge about problem-solving tradeoffs to optimize dynamically the value of computer performance to the user, (2) to construct efficient algorithms for probabilistic reasoning, and (3) to investigate pragmatic techniques for the elicitation of knowledge from experts. Two existing diagnostic systems are testbeds for this research, which is directly applicable to many areas of science and engineering. Although the techniques of artificial intelligence are being used to provide decision support tools ("expert systems") in a variety of settings within science, technology, and industry, several fundamental research problems remain. One significant concern is the lack of well-developed methods for dealing with the uncertainty that characterizes many problem-solving tasks. For example, in medicine or geological exploration, experts must infer the state of the process they are investigating (e.g., a patient's disease or a possible mineral deposit) by interpreting indirect measurements that generally provide only partial support for hypotheses. Most available tools deal with such uncertainty by using ad hoc methods whose applicability and limitations are poorly understood. By developing practical methods for using formal probability theory in such systems, this research replaces ad hoc techniques with approaches for which the underlying assumptions and limitations can be demonstrated analytically and considered during system construction and use.