Lecture
| The basic approach to solving problems, formed during the first decade of research in the field of artificial intelligence, was a general-purpose search mechanism, with the help of which attempts were made to link in a single chain the elementary stages of reasoning to form complete solutions. Such approaches were called weak methods , since they did not allow to increase the scale of their application to the level of larger or more complex instances of tasks, despite the fact that they were common. An alternative to weak methods has become an approach that involves the use of more substantial knowledge relating to the problem area, which allows you to create longer chains of logical inference steps and makes it easier to cope with the problematic situations that usually arise in specialized areas of knowledge. As it is known, to solve a rather complicated task, it is necessary to know the answer almost completely. One of the first examples of the implementation of this approach was the program Dendral. It was developed at Stanford University by a group of scientists that included Ed Feigenbaum (a former student of Herbert Simon), Bruce Buchanan (a philosopher who changed his specialty and became involved in computer science) and Joshua Lederberg (Nobel Prize winner in the field of genetics). This group was engaged in solving the problem of determining the structure of molecules on the basis of information obtained from a mass spectrometer. The input of this program consisted of chemical formula of the compound (for example, C 6 H 13 NO 2 ) and the mass spectrum, which allows determining the masses of various fragments of a molecule, which was formed when a molecule was bombarded by a stream of electrons. For example, the mass spectrum may contain a peak at the point m = 15, corresponding to the mass of the methyl fragment (CH 3 ). The first, primitive version of this program envisaged the development of all possible structures compatible with this formula, after which it predicted what mass spectrum should be observed for each of these structures, comparing it with the actual spectrum. It can well be expected that such a task with respect to molecules of larger sizes becomes insoluble. Therefore, the developers of the Dendral program consulted with analytical chemists and came to the conclusion that they should try to organize work on the principle of searching for well-known patterns of peak locations in the spectrum, which indicate the presence of common substructures in the molecule. For example, the following rule can be used to recognize ketone subgroups (C = 0) with atomic weights of 28. if there are two peaks at points x 1 and x 2 , such that:
The application of the method involving the recognition that the molecule contains some specific substructures, has greatly reduced the number of possible candidates to be tested. The value of the Dendral program was that it was the first successfully created expert system based on the wide use of knowledge: its ability to cope with the tasks set was due to the use of a large number of special-purpose rules. In later systems, the basic principle of the approach implemented by McCarthy in the Advice Taker program was also widely used - a clear separation of knowledge (in the form of rules) from the component that provides the reasoning. Guided by this experience, Feigenbaum and other specialists from Stanford University began to develop a heuristic programming project (Heuristic Programming Project — NRR), the purpose of which was to investigate to what extent the new methodology of expert systems created by them can be applied in other areas of human intellectual activity. At the next stage, the main efforts were concentrated in the field of medical diagnostics. Feigenbaum, Buchanan and Dr. Edward Shortliff developed the Mycin program for diagnosing infectious diseases of the circulatory system. After about 450 rules were entered into it, the Mycin program acquired the ability to work at the level of some experts, and also showed significantly better results compared to doctors who do not have much experience. She also had two important distinguishing features compared to the Dendral program. First, unlike the Dendral rules, there was no general theoretical model on the basis of which the logical derivation of the Mycin rules could be made. To identify these rules, it was necessary to widely apply the knowledge gained from experts, who, in turn, acquired this knowledge through textbooks, other experts, and direct experience gained through the study of practical cases. Secondly, these rules had to take into account the degree of uncertainty, which is characterized by knowledge in the field of medicine. The Mycin program used uncertainty calculus based on the so-called confidence factors, which (at that time) seemed quite appropriate to how physicians evaluate the effect of objective data on a diagnosis. The importance of using knowledge in the problem area has also become obvious for specialists who have been working on problems of understanding natural language. Although the system of understanding of the natural language Shrdlu, developed by Terry Vinograd, became in its time the subject of universal admiration, its dependence on the results of the syntax analysis caused the appearance of approximately the same problems that were found in early works on machine translation. This system was able to overcome ambiguity and correctly understood references expressed using pronouns, but this was mainly due to the fact that it was specifically designed for only one area - for the world of blocks. Some researchers, including Eugene Charniak, a colleague and a graduate student at Vinograd at the Massachusetts Institute of Technology, have indicated that general knowledge of the world and a common method of using this knowledge will be required to ensure a reliable understanding of the language. Roger Schenk who worked at Yale University, a linguist who became a researcher in the field of artificial intelligence, expressed this idea even more clearly, stating that "there is no such thing as syntax." This statement caused outrage among many linguists, but was the beginning of a useful discussion. Schenk and his students created a number of interesting programs. The task of all these programs was to provide an understanding of the natural language. But they focused less on the language as such and more on the problems of representation and reasoning with the knowledge required to understand the language. The problems addressed included the presentation of stereotypical situations, a description of the organization of human memory, and an understanding of plans and goals. In connection with the wide growth in the number of applications designed to solve real-world problems, the need to create workable knowledge representation schemes has also increased so widely. A large number of different languages were developed to represent knowledge and conduct reasoning. Some of them were based on logic, for example, the Prolog language became widespread in Europe, and the Planner family of languages was widely used in the United States. In other languages based on the frame idea put forward by Minsk, a more structured approach was adopted, providing for the collection of facts about specific types of objects and events, as well as the ordering of these types in the form of a large taxonomic hierarchy similar to biological taxonomy. |
Comments
To leave a comment
Models and research methods
Terms: Models and research methods