Stahl678 bmp

202 WALTER R. STAHL

A. Turing’s Test and Artificial Intelligence

The concept of algorithmic modeling was probably initiated by an article of Turing (1956), entitled “Can a Machinę Think?” which pro-posed an objective method (Turing’s test) for comparing a human and a thinking machinę, namely, playing a ąuestion and answer gamę with them over a teletypewriter. Turing concluded that a living brain and Computer should be considered as being operationally indistinguishable to the extent that they gave the same set of answers in a certain well-defined test situation. It is elear that presently available artificial brains are not nearly as versatile as humans, but can be madę to act in a rather anthropomorphic manner in checkers, medical diagnosis, literaturę searching, translation, etc.

The operational viewpoint embodied in Turing’s test is in a sense ap-plicable to all modeling eomparisons. A model and prototype are never completely similar unless they are identical, but the degree of similarity or extent of “modeling” is a fundamentally arbitrary matter defined by a specific test criteria, which can (and should) be stated as an algorithm. In a hydrodynamic model, the test algorithm may consist of computing the Reynold’s number, whereas in a servomechanism model it might involve comparing the gain and Nyąuist stability criteria of the model and prototype.

Numerous works dealing with modeling of phenomenological or heuristic brain action are now available. Yolumes on the subject include collections of articles edited by Sayre and Crosson (1963), Feigenbaum and Feldman (1963), and Muses (1962). Newell and Simon (1961) re-view progress in Computer simulation of human Boolean logie decision trees in various situations, while Rosenblatt (1962) considers a quasi-analog multiple layer switching network (perceptron) as a generał model of the mammalian brain. There is a great deal of interest in brain modeling in the Soviet Union. Glushkov (1963c) provides a detailed discussion of how and why human thought algorithms can be simulated on digital computers, while GasuP (1962) discusses specifically “obtaining a model of a thought function.” All these works deal fundamentally with algorithmic models of human thought. They are in no way concemed with physical realizations of brain function and imply no physical or numer-ical similarity invariants.

A typical example of human thought algorithms intensively analyzed for purposes of Computer implementation are those connected with medical diagnosis, as discussed in Lusted (1965), Ledley and Lusted