The eventual goal of AI research is human level AI. As some expected and others did not, understanding intelligence well enough to write programs has proved difficult. For example, no existing program can learn any branch of science from textbooks on the Web. Consequently, most present AI research is aimed at much more limited goals and is committed to formalisms incapable of expressing many facts that even young children know. Basic research in AI needs to be evaluated according to whether it advances the goal of human level AI.
There are two main lines of approach to AI. One attempts to imitate, to varying extents, human neurophysiology. The other looks at the problems the world, including the world of other actors, presents to an intelligence trying to achieve goals or just to survive. This lecture concerns the second approach.
The key problem is representing common sense knowledge and reasoning. My approach since 1958 is to represent knowledges in suitable languages of mathematical logic and to use logical reasoning to infer strategies that achieve goals.
A major concept of this research is the common sense informatic situation in which the actor has limited information about the objects in its world and limited knowledge of the effects of actions and other events. Among the phenomena that must be represented are partially known physical objects and the beliefs and goals of other actors. This requires extensions to the standard languages of mathematical logic. The reasoning needed for human level intelligence includes reflexion principles (known to logicians) and non-monotonic reasoning, which logicians and philosophers are only beginning to study.
This lecture discusses a path to human level logical artificial intelligence and the conceptual obstacles to following it. It's like crossing the desert in an automated vehicle but much harder.
About the speaker:
John McCarthy is Professor Emeritus of Computer Science at Stanford
University. His research has included the Lisp programming language, the
concept of time-sharing computers, formalisms for proving that computer
programs meet their specifications, and logical AI. McCarthy published
the first article on logical AI and its application to common sense
in 1959. He proposed the situation calculus in 1963 and 1969 (with a
major extension in 2002), the circumscription method of non-monotonic
reasoning in 1977, 1980 and 1986 and formalisms for representing
individual concepts and propositions, and formalisms for the formal
representation of contexts.
McCarthy's articles, some originally published in obscure places, are all to be found on his web page www-formal.stanford.edu/jmc.
Computer Science Department
Vox: (650) 857-0672