CADUCEUS [Pople, 1977] (extensions to INTERNIST) embodies more knowledge of internal medicine than any human and can correctly diagnose complex test cases that stymie human experts. In CADUCEUS, each sign and symptom is a starting node in a causal network with end points that are diseases that can account for the sign or symptom. It performs abduction (inference to the best explanation) since in the diagnostic problem, a patient can have more than one disease, which makes the possible number of combinations enormous. (CMU)
CODGER [Shafer et al., 1986] is an architecture for controlling vehicles in outdoor road-following tasks. It uses a blackboard structure to organize incoming perceptual data. It has been used to build a system for driving the experimental NAVLAB vehicle, a commercial van that has been altered for computer control via electric and hydrolic servos. The NAVLAB is completely self-contained, with room for several on-board computers and researchers. (CMU)
DENDRAL [Buchanan and Mitchell, 1977; Lindsay et al., 1980] analyzes organic compounds to determine their structure using mass spectroscopy and nuclear magnetic resonance. Its strategy is to use constraint satisfaction techniques. It has produced analyses that have been published as original research results and supports hundreds of international users daily in chemical structure elucidation. It surpasses all humans at its task and, as a consequence, has caused a redefinition of the roles of humans and machines in chemical research. (STANFORD)
EURISKO [Lenat, 1982] An extension to AM. AM used heuristic search to guide its discovery process in the context of set theory (e.g., it discovered prime numbers). Eurisko treated heuristics, themselves, as concepts that can be created and modified in the same way as task domain knowledge. EURISKO has been applied to number theory as well as design of naval fleets and VLSI design (e.g., it has the ability to induce new kinds of useful slots). (STANFORD)
MACE [Gasser et al., 1987] an example of a message-passing distributed system which provides a general architecture for distributed reasoning systems (much like BB1 provides a general architecture for blackboard systems.) In such a framework, the agents tend to know more about each other than they do in a blackboard system. This knowledge enables them to direct their messages to those agents who are most likey to be able to do what needs to be done. (early intelligent agent material)
MYCIN [Shortliffe, 1976] a medical diagnosis program for infectious blood diseases. It was designed to provide consultive advice on diagnosis and therapy. Uses production rules and probabilistic knowledge (not Bayes' rule, see PROSPECTOR). MYCIN served as the control methodology for EMYCIN, one of the first expert system shells; its inference engine is a backward chaining, exhaustive depth-first search of an AND/OR goal tree. (STANFORD)
PARRY [Colby, 1975] exploited a model of human paranoid behavior to simulate the conversational behavior of a paranoid person. (Meant to test psychological theories of human performance.) An example of programs where simplicity is bought at the price of superficial understanding. (YALE?)
SHRDLU [Winograd, 1972] The main thrust of the work was natural language understanding, for the point was to show how a computer can be made to accept commands and questions expressed in English. For this early work, it was particularly important to confine the English interaction to a narrowly focused domain; hence, the blocksworld. However, it thus represented knowledge in a highly semantically constrained world, and therefore derived simplicity from the fact that it dealt with a simplified world. (MIT)
PIP [Pauker 1976] assists physicians by taking the hostory of the present illness of a patient of edema. The system interweaves the processes of information gathering and diagnosis, alternating between asking questions to gain new information and integrating this new information into developing a picture of the patient. Knowledge in PIP is represented in a network of frames to model cognitive processes of short-term and long-term memory. (MIT)
HEARSAY [Reddy, 1976, Erman, 1980] speech understanding system has been one of the most influential of all AI programs over the years. The first implementation of the system (HEARSAY I) was already based on the idea of cooperating, independent knowledge sources. HEARSAY III is a knowledge engineering language for rule-based representation that also integrates a blackboard architecture consisting of knowledge sources that communicate via a central blackboard or data base. (CMU)
WUMPUS [Yob, 1975] is one of the ICAI systems (Intelligent Computer-Aided Instruction). It uses representations that capture complexity, prerequisite associations, analogy, and generalization relations. What is Wumpus? A game in which the player must track down and slay the vicious Wumpus while avoiding pitfalls that result in certain death. (MIT)
The following are the only AI machines that are available on the network capable of using the system MEDLEY
SOAR [Laird & Newell, 1983] is a general problem-solving architecture for representation of heuristic search-oriented problem solving. The system provides a means for viewing a problem as a search through a problem space, a set of states representing solutions and a set of operators that transform one state into another. The principle characteristics of SOAR include the automatic creation of a hierarchy of subgoals and problem spaces and a parallel rule interpreter. (CMU)
ABEL [Patil, 1981] assists the clinician in diagnosing acid-based and electrolyte disorders in patients by applying knowledge about the diseases and the symptoms they produce. The system uses a causal net of the patient's possible diseases to order queries to the clinician and guide the diagnostic reasoning process. Knowledge is represented within a causal network, a type of semantic net specifying cause-effect relations between diseases and findings. This type of strucutre allows for levels of detail. (MIT)