Page images
PDF
EPUB
[merged small][merged small][ocr errors][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][ocr errors][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small][merged small]

FIGURE 10:

10: Problem behavior Graph (PBG) for Fig. 9

(adapted from Newell, 1967, Fig. 10).

17
B97
R:9

[ocr errors]

search goes in this problem space. It is our fashion to write these programs as production systems, meaning an ordered set of condition-action expressions with a control structure that continually executes the first expression whose condition is

true.

(The actual expression being executed changes as the action parts continually

modify the immediate memory, which is what the conditions test.) Again, there is no space to describe a complete production system for our subject. A typical

production, available in almost all subjects can be paraphrased, "If a new equality

expression has just occurred, then find a column that contains the letter involved

and process that column." In symbols:

<letter>

<digit> new + Find-column(<letter>); Process-column (column).

The left side is written as a BNF grammar so that any expression in the knowledge

state of the right form will trigger the production (e.g., 'T = 0 new').

Given a production system (the one for the subject of Figure 9 and 14 productions), one can attempt an accounting of the nodes of the problem behavior graph that are described successfully by a production. Figure 11 shows this accounting

for the total protocol, in which the productions are ordered in terms of decreasing marginal utility. There are two kinds of errors. First, a certain fraction of the nodes in the problem behavior graph are not covered (errors of omission); this

gradually drops to 11% with the total set of productions. Second, the wrong production gets evoked at a node in the graph, due to the ordering (errors of commission); these gradually rise as the set of productions increases to about 9% or 14% (two levels of error were counted in the original work).

This brief description is not meant to do more than indicate a scientific style, namely, one where programs (the production systems) are constructed to deal

with specific instances of human data, with quantitative assessment of the

comparison. Not revealed in this brief account is the concern for understanding

how the programs of different subjects compare, so that a general theory of human

problem solving emerges. The example should reinforce the story told by the earlier ones -- that if one is serious about the implications of work in artificial intelligence for psychology, then one must come to terms with data on human

behavior.

[blocks in formation]

FIGURE 11: Summary of performance or production system for

Donald and Gerald (from Newell, 1967, Fig. 16).

7. FINAL POINT: ON PSYCHOLOGY'S PREFERENCES

The main points are now made, in response to the letter from my psychologist friend. First, the frame of reference must be expanded from artificial intelli.

gence to information processing systems on symbolic structures. Then, it is found that a very substantial penetration has occurred of information processing theories into psychology. Second, there is beginning a shift in experimental psychology from a concern exclusively with immediate memory to a concern with the whole of the

immediate processor. This testifies to the use of information processing theories at a detailed level, and not just as experiment-guiding metaphor. Third, the

coupling has become intimate enough that artificial intelligencers must take

experimental data seriously if they want to claim any direct relevance of their

work to psychology.

With these points made, I should rest content. But there is one place where I have finessed my letter-writing friend, and I must give him his due. For he did mean artificial intelligence and not information processing psychology in

general (at least, so I believe). Now, whereas I would insist on the latter view,

there is one respect in which his concern remains justified. The higher mental

processes in American psychology have always held a secondary place. There are many reasons for this: the impact of behaviorism that simply denied the relevance of the mental; the feeling that the important thing was the elements (i.e., the basic act of learning), out of which complexity would grow automatically; the sorewhat fanatical adherence to the scientific canon of simplicity first; the actual messiness of the area, as revealed by the relatively few, but continual, forays that have occurred. The reasons make little difference. The psychology of thinking and problem solving makes up a rather sad chapter in the history of psychology. It is one of my fond hopes that the situation has finally changed

that we

have the techniques to deal with integrated goal oriented behavior on a par with

any other part of psychology. But I am afraid that my correspondent is correct:

that psychologists will not shift in mass to the study of thinking and problem solving. These areas will remain relatively lightly populated and this was in

part what he was decrying, for he expected it to be different some ten years after

the initial work.

I think he is correct in his depression, since, as we have seen, there is large

contact between information processing theories and psychology at the level of immediate memory (and psycholinguists too, though I didn't stress it as much).

These are extremely exciting areas at the moment, as noted earlier. They can be

attacked with relatively small amounts of theory and large amounts of sophisticated experimental techniques. They fit psychology's image of the proper thing to study: basic structure and not too much complexity. Thus, as experimental psychologists

move towards assimilating information processing theories, they will gravitate

towards the study of the immediate processor and basic language structure, and not

toward thinking and complex problem solving. It is not without significance that only one out of five of my examples came from the heartland of artificial intelli

gence.

REFERENCES

1.

Attneave, F., Applications of Information Theory to Psychology, Holt-Drydın,

1959.

2. Boring, E., "Mind and mechanism," American J. Psychology, 59, 2, April,

1946.

3. Bourne, L.E., Jr., Human Conceptual Behavior, Allyn and Bacon, 1966.

4.

Broadbent, D., Perception and Communication, Pergamon, 1958.

5. Bruner, J.S., Goodnow, J.J. and Austin, G.A.; A Study of Thinking, Wiley, 1956. 6. Bugelski, B.R., Kidd, E. and Segmen, J., "Image as a mediator in one-trial

paired-associate learning," J. Experimental Psychology, 76, 69-73, 1968.

7. Chomsky, N., Aspects of Syntax, MIT Press, 1965.

8.

Churchrian, W., "The role of Weltanschauung in problem solving and inquiry,"

These Proceedings, 1970.

9. Colby, K.M. and Gilbert, J.P., "Programming a computer model of neurosis,"

J Math. Psychology, 1, 2, 1964.

10. Collins, A.M. and Quillian, M. R., "Retrieval time from semantic memory,"

j. Verbal Learning and verbal Behavior, 8, 204-247, 1969.

[ocr errors]

Dansereau, D., An Information Processing Model of Mental Multiplication,

Unpublished Ph.D. dissertation, Carnegie-Mellon University, 1969.

« PreviousContinue »