In the spring of 1971, I attended a course on statistical inference taught by Arthur Dempster at Harvard. In the fall of that same year Geoffrey Watson suggested I give a talk expositing Dempster's work on upper and lower probabilities to the Department of Statistics at Princeton. This essay is one of the results of the ensuing effort. It offers a reinterpretation of Dempster's work, a reinterpretation that identifies his "lower probabilities" as epistemic probabilities or degrees of belief, takes the rule for combining such degrees of belief as fundamental, and abandons the idea that they arise as lower bounds over classes of Bayesian probabilities.

In Charniak and Wilks, 55-72.

This paper presents a functional overview of the features and capabilities of QLISP, one of the newest of the current generation of very high level languages developed for use in Artificial Intelligence (AI) research.QLISP is both a programming language and an interactive programming environment. It embeds an extended version of QA4, an earlier AI language, in INTERLISP, a widely available version of LISP with a variety of sophisticated programming aids.The language features provided by QLISP include a variety of useful data types, an associative data base for the storage and retrieval of expressions, the ability to associate property lists with arbitrary expressions, a powerful pattern matcher based on a unification algorithm, pattern-directed function invocation, "teams" of pattern invoked functions, a sophisticated mechanism for breaking a data base into contexts, generators for associative data retrieval, and easy extensibility.System features available in QLISP include a very smooth interaction with the underlying INTERLISP language, a facility for aggregating multiple pattern matches, and features for interactive control of programs.A number of applications to which QLISP has been put are briefly discussed, and some directions for future development are presented. SRI Tech.Note 120, AI Center, SRI International, Inc., Menlo Park, Calif.

This text is a description of a computer-based system designed to assist physicians with clinical decision-making. This system, termed MYCIN, utilizes computer techniques derived principally from the subfield of computer science known as artificial intelligence (AI). MYCIN's task is to assist with the decisions involved in the selection of appropriate therapy for patients with infections.

MYCIN contains considerable medical expertise and is also a novel application of computing technology. Thus, this text is addressed both to members of the medical community, who may have limited computer science backgrounds, and to computer scientists with limited knowledge of medical computing and clinical medicine. Some sections of the text may be of greater interest to one community than to the other. A guide to the text follows so that you may select those portions most pertinent to your particular interests and background.

The complete book in a single file.

Reprinted in Readings in Uncertain Reasoning, G. Shafer and J. Pearl, eds., pp. 259-273, San Mateo, CA: Morgan Kaufmann Publishers, Inc., 1990.See also: Stanford Center for Biomedical Informatics Research (BMIR).… quantifying confirmation and then manipulating the numbers as though they were probabilities quickly leads to apparent inconsistencies or paradoxes. Carl Hempel presented an early analysis of confirmation (Hempel, 1965), pointing out as we have that C[h,e] is a very different concept from P(hle ). His famous Paradox of the Ravens was presented early in his discussion of the logic of confirmation. Let hl be the statement that "all ravens are black" and h2 the statement that "all nonblack things are nonravens." Clearly hi is logically equivalent to h,2. If one were to draw an analogy with conditional probability, it might at first seem valid, therefore, to assert that C[hl,e] = C[h2,e] for all e. However, it appears counterintuitive to state that the observation of a green vase supports hi, even though the observation does seem to support h,2. C[h,e] is therefore different from P(hle) for it seems somehow wrong that an observation of a vase could logically support an assertion about ravens. Another characteristic of a quantitative approach to confirmation that distinguishes the concept from probability was well-recognized by Carnap (1950) and discussed by Barker (1957) and Harrd (1970). They note it is counterintuitive to suggest that the confirmation of the negation of a hypothesis is equal to one minus the confirmation of the hypothesis, i.e., C[h,e] is not 1 - C[-qh,e]. The streptococcal decision rule asserted that a gram-positive coccus growing in chains is a Streptococcus with a measure of support specified as 7 out of 10. This translates to C[h,e]=0.7 where h is "the organism is a Streptococcus" and e is the information that "the organism is a gram-positive coccus growing in chains." As discussed above, an expert does not necessarily believe that C[mh,e] = 0.3. The evidence is said to be supportive of the contention that the organism is a Streptococcus and can therefore hardly also support the contention that the organism is not a Streptococcus. Ch.13 of Mycin Book; revised from Math. Biosci. 23:351-379

IEEE Transwactions on Systems, Man and Cybernetics SMC-5, pp. 409-425

 John Holland's pioneering book Adaptation in Natural and Artificial Systems [1975, 2nd ed. 1992] showed how the evolutionary process can be applied to solve a wide variety of problems using a highly parallel technique that is now called the genetic algorithm. The genetic algorithm transforms a population of individual objects, each with an associated fitness value, into a new generation of the population using the Darwinian principle of reproduction and survival of the fittest and naturally occurring genetic operations such as crossover (recombination) and mutation. Each individual in the population represents a possible solution to a given problem. The genetic algorithm attempts to find a very good or best solution to the problem by genetically breeding the population of individuals. In preparing to use the conventional genetic algorithm operating on fixed-length character strings to solve a problem, the user must 1. determine the representation scheme, Ann Arbor: The University of Michigan Press

See also Werbos, Paul J. (1994). The Roots of Backpropagation. From Ordered Derivatives to Neural Networks and Political Forecasting. New York, NY: John Wiley & Sons, Inc. Rumelhart, David E.; Hinton, Geoffrey E., Williams, Ronald J. (8 October 1986). "Learning representations by back-propagating errors". Nature323 (6088): 533–536. doi:10.1038/323533a0.Ph.D. thesis, Harvard University.

In L. W. Gregg (Ed.), Knowledge and cognition. Hillsdale, NJ: Erlbaum

Artificial Intelligence 5:137-148.