Earlier we made a distinction between factual and practical knowledge (knowing how and knowing that). Another distinction worth stressing is between domain-specific and domain-independent knowledge. The specific knowledge domain of a London tourist guide (the needs and interests of tourists to London) is made up of a number of sub-domains, such as the location of places of interest, the transport system in London, and the prices of various activities and goods. Having the knowledge clustered into identifiable domains makes the job of designing a system like our Tourist Guide considerably easier and is one reason for the commercial success of expert systems. Expert systems can reproduce the specific domain expertise of, for example, a geological mineral prospector (Duda, Gaschnig and Hart, 1979) or a specialist in blood infections (Shortliffe, 1976) and are often able to perform as well as, or in some cases even better than, their human counterparts.
Compare the following two paragraphs (from Hayes-Roth, 1983) summarizing the pulmonary function diagnosis of a particular patient at a San Francisco hospital. One is written by a human physician, and the other by an expert system called PUFF, using exactly the same details from the patient. Try and guess which is which.
Conclusions: The low diffusing capacity, in combination with obstruction and a high total lung capacity is consistent with a diagnosis of emphysema. Although bronchodilators were only slightly useful in this one case, prolonged use may prove to be beneficial to the patient. PULMONARY FUNCTION DIAGNOSIS: MODERATELY SEVERE OBSTRUCTIVE AIRWAYS DISEASE. EMPHYSEMATOUS TYPE.
Conclusion: Overinflation, fixed airway obstruction and low diffusing capacity would all indicate moderately severe obstructive airway disease of the emphysematous type. Although there is no response to bronchodilators on this one occasion, more prolonged use may prove to be more helpful. PULMONARY FUNCTION DIAGNOSIS: OBSTRUCTIVE AIRWAYS DISEASE, MODERATELY SEVERE, EMPHYSEMATOUS TYPE.
The first was the computer-generated diagnosis, but you might be forgiven if you got it wrong.
Currently, an expert system capable of exhibiting impressive performances like this has only extremely narrow and specialized knowledge. PUFF, for instance, knows nothing about medical complaints apart from conditions of the lung. It may not even be able to answer questions of common knowledge about human anatomy (for example, ``Are my lungs above or below my knees?'').
Now a lot of the knowledge on which you would have to rely if working as a tourist guide would be of a quite unspecialized kind, based upon general knowledge, or common sense. For instance, you would need to know that if it is raining very heavily outside, a scheduled open-air symphony concert is likely to have been cancelled. Or, to take another example, this time concerning our commonsense understanding of the dynamics of conversation, you would need to know that, if an excessively polite inquirer says, ``Would you possibly be able to tell me the way to X?'' this means, ``Please tell me the way ...?'' rather than ``Do you, hypothetically, have the ability ...?'' Such items of knowledge do not fit easily into specialized bodies of expertise -- they are just some of the myriad pieces of life-experience that we pick up. If AI researchers are to do more than build systems which give virtuoso performances within extremely narrow limits, then they will have to find ways of representing large amounts of domain-independent general knowledge.
There are two different sorts of question involved here, although they are intimately linked. The first is, ``How is knowledge represented in the minds (brains) of human subjects, or those of other animals?'' This is a question of psychology, and, as will be made clear many times in this book, the concerns of psychology and of AI overlap to a very large degree. The second is, ``How can we best organize the way knowledge is represented in computer systems, given the constraints on how such systems can be designed, so as to make them useful to us?'' Much of AI is concerned with making systems which actually work (and work effectively); so, in answering the second question, people may not necessarily be too worried if there is no close fit between the methods for representing knowledge in machines and those which might be supposed to occur in humans. However, many AI researchers believe that in order to represent knowledge in flexible and sophisticated forms in machines (especially commonsense knowledge rather than specialized expertise) it is best to build knowledge-structures which closely reproduce those we seem to use ourselves. Some researchers believe that, to parody Alexander Pope, ``The proper study of computerkind is man.'' There is a lively debate within the AI community between those who aim for neat, systematic, logical structures which are intended to suit the demands of the computer as an information-processing device; and those who try to build knowledge representation systems which reflect the rich `messiness' of human experience and ideas. This controversy, between the `neats' and the `scruffs' , as they are often called, has resulted in an interesting divergence of styles of representing knowledge in computational systems.