CAREER: Compute Intensive Methods for Artificial Intelligence

点击翻译按钮获取中文摘要

The objectives of this research and educational program are to develop new formalisms and methods for artificial intelligence by combining a sound theoretical approach with a principled experimental component, to provide a practical evaluation of the proposed models and relate them to real-world applications and challenges, and to educate the next generation of computer science students about the ambitious and challenging goals of artificial intelligence. To develop successful applications, general search and reasoning has been avoided in recent years by explicitly incorporating large amounts of domain-specific knowledge. Such a knowledge-intensive approach has been successful in, for example, expert systems; but in areas such as general planning or reasoning, progress has been disappointing, and specific knowledge acquisition is often difficult and expensive. However, recent advances in general methods combined with faster hardware and better implementations provide strong evidence that a compute-intensive approach is not only suitable for dealing with the combinatorial nature of many AI formalisms, but may also be required to supplement domain-specific knowledge. This research studies fast general reasoning and search methods, with an emphasis on stochastic procedures, which are a promising recent development for solving computationally hard problems. It also investigates the various sources of complexity in hard problems, using both theoretical and experimental methods, exploring interesting connections between computer science, artificial intelligence, and statistical physics. In addition, it is studying issues in problem representation, including robustness of encodings, abstraction, compilation, and approximation methods. The result is expected to be new general approaches to artificial intelligence that are applicable to a wide variety of problems in such areas as planning, operations research, and computational biology. http://simon.cs.cornell.edu/home/selman/