WHAT IS CYBERNETICS? — RIacKAY 403 



reflexes" found necessary in the course of past interaction with its 

 field of activity. The scope of "artificial intelligence" along these 

 lines is virtually unlimited in principle. It is thus hardly surprising 

 that cybernetics in popular thought is so closely identified today with 

 the theory of computers. 



The earliest and most characteristic mathematical developments in 

 cybernetics, however (in a paper by J. Clerk Maxwell in 1868), were 

 concerned with a more central problem — that of warding off instabil- 

 ity. In any closed cycle of control, if the response to control is too 

 sluggish or too violent, it is fatally easy for the system to become un- 

 stable, overcorrecting itself in a series of wild swings in opposite direc- 

 tions, called "hunting." Although the mathematics of unstable be- 

 havior has mostly been developed for "linear" systems (those whose 

 responses change in strict proportion to changes in input), there are 

 a number of general principles and rules of thumb which invite appli- 

 cation in a wide range of fields at present plagued by instability. 

 These include, for example, keeping the number of stages in a chain as 

 small as possible; reducing sensitivity to the minimum acceptable; 

 and combating sluggishness by taking rates of change of indication 

 as a guide to action. 



The second great area of mathematical development specific to cy- 

 bernetics has become known as Information Theory. To control a 

 task of a given complexity to a given accuracy, in an environment 

 with given statistical features, how much information does a cyber- 

 netic governor need? The theory of information sets out to make 

 such questions precise, and to give mathematical answers to them. 

 Thanks largely to the work of C. E. Shannon, it has been generalized 

 to enable communication engineers to evaluate and compare the chan- 

 nel capacities of different encoding or transmitting systems, and to 

 take precise account of the effects of random disturbance, or "noise." 

 Even more important, it has shown how statistical correlation between 

 different elements of a signal ("redundancy") can be used to enable 

 random errors in transmission to be detected and corrected, so that 

 a "noisy" channel can — in principle, and given a long enough run 

 for statistical purposes — transmit up to a definite rate with arbitrarily 

 little error. 



Once again it should be added, however, that in the bulk of cyber- 

 netic investigations to date it is the qualitative notions of information 

 theory — information, encoding, noise, redundancy, channel capacity, 

 error-correction — rather than its mathematical apparatus, which have 

 so far found illuminating uses. No one interested in the cybernetic 

 approach should be frightened off, or unduly impressed, by sprinkled 

 references to unfamiliar mathematics in the somewhat uneven litera- 

 ture of the field. With few exceptions to date, their function will turn 

 out to be decorative rather than pivotal. 



