Transcript document
History of AI •1st generation: pre 1956 •2nd generation: mid 1950's to 1970 •3rd generation: 1970 on •Fourth generation: ??: st 1 generation: pre 1956 Alan Turing - Wrote famous paper in 1937 in which he invented the concept of the Turing Machine to explain why some problems are uncomputable. Devised Turing Test to serve as an objective tool for deciding if a program is "intelligent". 1st generation: pre 1956 Warren McCulloch (U of Illinois psychiatrist) and Walter Pitts (18-year-old mathematician) - In 1943, came up with a mathematical model of a binary neuron which could be emulated on a computer and which could learn by adjusting the strength of its connection with its inputs. Gave rise to the study of perceptrons, which were developed by Frank Rosenblatt in the 50's. 1st generation: pre 1956 (Perceptrons were later criticized by Marvin Minsky and Seymour Papert at MIT (1969) and research along these lines was generally abandoned until Bernard Widrow revitalized the field in the late 80's with more powerful backpropagation nets.) 1st generation: pre 1956 John von Neumann (Princeton mathematician) - helped Mauchley and Eckert develop the ENIAC (1946). Brilliant computer science theoretician. Built EDVAC, which embodied the stored program concept. Thought a lot about AI issues. Used McCulloch-Pitts neuron to describe his theories. 1st generation: pre 1956 Claude Shannon - mathematician who worked in the area of communication theory. Showed how the binary switches in a computer could actually store information. Described a chess playing algorithm (1950) that anticipated much of the more recent work in game playing. 1st generation: pre 1956 H. Ross Ashby - U of Illinois. Wrote Design for a Brain (1952), an influential analysis of what would be necessary to emulate the functions of an intelligent system. 2nd generation: mid 1950's to 1970 1956 Dartmouth conference (organized by McCarthy and Minsky) defined the field. Attended by all the big names in AI. AI researchers at this time came up with the physical symbol system hypothesis; humans think by manipulating symbols, so, instead of trying to have computers emulate the hardware (neurons), AI research should concentrate on symbol manipulation (software). 2nd generation: mid 1950's to 1970 • John McCarthy - inventor of LISP, used for most previous AI work, especially in Natural Language Processing. Originated term "Artificial Intelligence". • Marvin Minsky - Worked in many areas of AI. In knowledge representation, developed the concept of "frames". • Arthur Samuel - developed a magnificent checkers playing program that learned from playing its opponents. 2nd generation: mid 1950's to 1970 Allen Newell, Herb Simon, J. C. Shaw developed Logic Theorist, which proved 38 of the first 52 theorems from Russell and Whitehead's Principia Mathematica (one proof was shorter and more elegant than the original!). This program was later expanded into the General Problem Solver. 2nd generation: mid 1950's to 1970 • Ed Feigenbaum - developed DENDRAL, one of the first expert systems. • Seymour Papert - studied under Jean Piaget & worked with Minsky. Developed LOGO, worked on computer-assisted instruction. 3rd generation: 1970 -2000 • Terry Winograd - wrote SHRDLU, Blocks World. • Bertram Rafael - worked in robotics, developed SHAKEY to respond to human instructions. • Nils Nilsson & Richard Fikes - developed STRIPS to achieve goals by the use of plans and a sequence of operators. • Daniel Bobrow - wrote STUDENT, which could solve algebraic word problems. 3rd generation: 1970 -2000 • David Slate & Larry Atkin - developed Chess 4.5, the first world-class chess playing program. • Raj Reddy - HEARSAY - understood human speech with an accuracy of 90% or better. • Roger Schank and Richard Ableson developed the idea of scripts to provide a framework for representing actions. Models common-sense knowledge of stereotypical situations. 3rd generation: 1970 -2000 • David Waltz - used a constraint satisfaction approach for understanding visual scenes. Can handle cracks, shadows, etc. • Edward Shortliffe - developed first real expert system, MYCIN, which can now diagnose blood diseases more accurately than most human physicians. 3rd generation: 1970 -2000 • Richard Duda - developed PROSPECTOR, a geological analysis expert system which found a commercially valuable deposit of molybdenum. First expert system to incorporate Bayes' rule. • Doug Lenat - working on CYC, a huge database (1 million items) of common-sense knowledge. 4th generation: 2000 - ?? • John Hopfield - rediscovered neural networks. • Many current AI researchers (e.g., Bernard Widrow, John Holland, David Goldberg, John Koza) have recently turned from symbolic processing to AI approaches which emulate natural systems, such as evolution or biological neural networks. Summary • First period of development of the field of AI laid the foundations. • Second period defined the field and explored many approaches to AI, but achieved success primarily with "toy" problems - expert systems is the only significant exception that has had a wide impact in the real world. Also trained the 3rd generation of AI researchers. Summary • Third period focused on building intelligent systems in restricted domains, instead of trying to construct a general intelligent agent. Recognized the importance of a knowledgebased approach. • Fourth period may be beginning now, with more of a focus on "soft" computing based on biological and physical metaphors: genetic algorithms, neural networks, simulated annealing, cellular automata, artificial life, etc.