S MAN RUSHES to build his replacements, he communication. In the meantime at least, Though such abstraction is difficult, we already have computers must be able to, but cannot, understand the given some of our machines limited ability to read printing writing and talking of men. We are protected from technological in certain type faces [1], [2]. But reading scratchpad unemployment so long as we are buffered by handwriting or transcribing conversational speech punched cards, magnetic tapes, and on-line or off-line by machine is far beyond our ken. Also, it seems clear printers. But the day will come!

Reprinted from Information Theory, Fourth London Symposium published by Butterworths, 88 Kingsway, London, W.C.2. MARVIN MINSKY and OLIVER G. SELFRIDGE Lincoln Laboratory*, Massachusetts Institute of Technology INTRODUCTION THE general nature of the problem is that an organism must learn to make the'right', or appropriate, response to its inputs. Typically, the inputs are large amounts of data, so that the machine must learn to recognize the similarities between different inputs which call for the same response, contrasted with the distinctions that call for different responses. The particular machines we are concerned with are random nets. A random net is a large set of similar and simply-acting elements whose attributes and interactive connections may be randomly established. The extent to which randomness is a part of setting up or maintaining a net varies in the literature, and more recent accounts tend to minimize the use of randomness. Some of the units are usually designated input, and some output units. The units themselves are termed neurons or cells. The underlying reason for the interest in random nets is the belief that if'right' responses are rewarded by some'reinforcement', perhaps of the contributing connections, and'wrong' ones discouraged, then the net as a whole will organize itself so as to tend to make only right responses, even when they are very complicated and abstruse.

Summary If we have a machine with 1000 knobs, how can we set them so as to minimize some output S of the machine, which represents, say, its error or departure from the behavior we want from it? We describe units, driven by S, which will each work a knob so that the whole system will tend toward an optimum. The units can be substantially identical, regardless of the actual structure of the machine. We exhibit three versions of these units, each with virtues and faults, and discuss -their behavior, with concrete and synthetic illustrative experiments. There are divers aspects of their joint behaviors shown, and some caveats about their use, especially in large numbers. We have not finished testing these units in large assemblies, but it is probably unimportant that the component parts of each unit work accurately or reliably.

By this we mean the extraction of E CONSIDER the process we call Pattern clearly one of the most primitive sources of patterns.

Papers read at a Symposium on'Information Theory' held at the Royal Institution, London, September 12th to 16th 1955 Published by BUTTERWORTHS SCIENTIFIC PUBLICATIONS 88 KINGSWAY, LONDON, W.C.2 MANY psychologists studying learning have assumed that the subject--rat, dog, or graduate student--invariably knows what the stimulus is. They have not concerned themselves with how a dog knows that it is the bell ringing which is the stimulus to jump over a fence. A bell ringing never gives the same set of nervous impulses into the brain twice (of course the argument would still apply even if it did); why then should the dog classify all cases of bell ringing into one category--'stimulus'? There is then the further question of how this category is more or less quickly'associated' with a response: the point is that the stimulus is not a priori considered a significant entity by the subject. In designing programmes for computers to imitate conditioned reflexes, for instance, we have found that the real problem was to identify the stimulus.

At the time of the Dartmouth Well, when Digital built the PDP-1, you and we had studied philosophy. Not only conference, there were certain mathematical sat at the console and you wrote your program that, but we also knew McCulloch, who games called Post tag systems.

In either view, typical examples are closer where (only) positive examples ought to exceptional cases, case memory, analogy, to the prototype, and atypical examples and be.

"... software technology has promulgated a series of A vital problem is the measurement of overall performance, failed promises....

Although we too are discussing futuristic behavior, we believe This is another emergent property.

The effects of the parameters 19 8.3 When the periods are the same 22 8.4 Using the RNG when the periods are the same 24 8.5 Trajectories with numerically related periods 24 9.0 Adaptive AP Strategies or AAP Strategies 26 10.0 Discussion and Generalizations 26 10.1 Another problem: the adapting of integer variables 27 10.2