These are the voyages of the MIT Artificial Intelligence Laboratory, and these remarks may help to understand the context of this collection, though in many ways the memoranda speak quite clearly for themselves and my comments are not, in any case, to be regarded as history, for I have written them quite hastily, in much the same spirit of the memos themselves, when it was our strategy in those early days to be unscholarly: we tended to assume, for better or for worse, that everything we did was so likely to be new that there was little need for caution or for reviewing literature or for double -checking anything. As luck would have it, that almost always turned out true.

Our group's work in medical decision making has led us to formulate a framework for expert system design, in particular about how the domain knowledge may be decomposed into substructures. We propose that there exist different problem-solving types, i.e., uses of knowledge, and corresponding to each is a separate substructure specializing in that type of problem-solving. Each substructure is in turn further decomposed into a hierarchy of specialist which differ from each other not in the type of problem-solving, but in the conceptual content of their knowledge; e.g.; one of them may specialize in "heart disease," while another may do so in "liver," though both of them are doing the same type of problem solving. Thus ultimately all the knowledge in the system is distributed among problem-solvers which know how to use that knowledge. This is in contrast to the currently dominant expert system paradigm which proposes a common knowledge base accessed by knowledge-free problem-solvers of various kinds. In our framework there is no distinction between knowledge bases and problem-solvers: each knowledge source is a problem-solver. We have so far had occasion to deal with three generic problem-solving types in expert clinical reasoning: diagnosis (classification), data retrieval and organization, and reasoning about consequences of actions. In novice, these expert structures are often incomplete, and other knowledge structures and learning processes are needed to construct and complete them.

Production rules are a popular representation for encoding heuristic knowledge in programs for scientific and medical problem solving. However, experience with one of these programs, mycin, indicates that the representation has serious limitations: people other than the original rule authors find it difficult to modify the rule set, and the rules are unsuitable for use in other settings, such as for application to teaching. These problems are rooted in fundamental limitations in mycin's original rule representation: the view that expert knowledge can be encoded as a uniform, weakly structured set of if/then associations is found to be wanting. To illustrate these problems, this paper examines mycin's rules from the perspective of a teacher trying to justify them and to convey a problem-solving approach. We discover that individual rules play different roles, have different kinds of justifications, and are constructed using different rationales for the ordering and choice of premise clauses.

The AM program, an early attempt to mechanize learning by discovery, has recently been expanded and extended to several other task domains. AM's ultimate failure apparently was due to its inability to discover new, powerful, domain-specific heuristics for the various new fields it uncovered. At that time, it seemed straight-forward to simply add ‘Heuristics’ as one more field in which to let AM explore, observe, define, and develop. That task—learning new heuristics by discovery—turned out to be much more difficult than was realized initially, and we have just now achieved some successes at it. Along the way, it became clearer why AM had succeeded in the first place, and why it was so difficult to use the same paradigm to discover new heuristics.

One of the challenges increasingly facing intelligence analysts, along with professionals in many other fields, is the vast amount of data which needs to be reviewed and converted into meaningful information, and ultimately into rational, wise decisions by policy makers. The advent of the world wide web (WWW) has magnified this challenge. A key hypothesis which has guided us is that threats come from ideas (or ideology), and ideas are almost always put into writing before the threats materialize. While in the past the'writing' might have taken the form of pamphlets or books, today's medium of choice is themore » WWW, precisely because it is a decentralized, flexible, and low-cost method of reaching a wide audience. However, a factor which complicates matters for the analyst is that material published on the WWW may be in any of a large number of languages. In'Identification of Threats Using Linguistics-Based Knowledge Extraction', we have sought to use Latent Semantic Analysis (LSA) and other similar text analysis techniques to map documents from the WWW, in whatever language they were originally written, to a common language-independent vector-based representation.

"We describe the concept of distributed problem solving and define it as the cooperative solution of problems by a decentralized and loosely coupled collection of problem solvers. This approach to problem solving offers the promise of increased performance and provides a useful medium for exploring and developing new problem-solving techniques. We present a framework called the contract net that specifies communication and control in a distributed problem solver. Task distribution is viewed as an interactive process, a discussion carried on between a node with a task to be executed and a group of nodes that may be able to execute the task. We describe the kinds of information that must be passed between nodes during the discussion in order to obtain effective problem-solving behavior. This discussion is the origin of the negotiation metaphor: Task distribution is viewed as a form of contract negotiation. We emphasize that protocols for distributed problem solving should help determine the content of the information transmitted, rather than simply provide a means of sending bits from one node to another. The use of the contract net framework is demonstrated in the solution of a simulated problem in area surveillance, of the sort encountered in ship or air traffic control. We discuss the mode of operation of a distributed sensing system, a network of nodes extending throughout a relatively large geographic area, whose primary aim is the formation of a dynamic map of traffic in the area. From the results of this preliminary study we abstract features of the framework applicable to problem solving in general, examining in particular transfer of control. Comparisons with PLANNER, CONNIVER, HEARSAY-II, AND PUP6 are used to demonstrate that negotiation—the two-way transfer of information—is a natural extension to the transfer of control mechanisms used in earlier problem-solving systems." Artificial Intelligence 20:63-109.

This study presents a process model of very long-term episodic memory. The process presented is a reconstructive process. The first is used to direct search to appropriate conceptual categories in memory. The other two are used to direct search within the chosen conceptual category. A fourth type of strategy, called executive search strategies, guide search for concepts related to the one targeted for retrieval.

A new parameter is introduced to characterize a type of search problem of broad relevance in artificial intelligence, operations research and symbolic logic. This parameter, which we call inter-variable compatibility, is particularly important in that complexity analyses incorporating it are able to capture the dependence of problem complexity on search order used by an algorithm. Thus compatibility-based theories can provide a theoretical basis for the extraction of heuristics for choosing good search orderings—a long-sought goal for such problems, since it can lead to significant savings during search. We carry out expected-complexity analyses for the traditional backtrack algorithm as well as for two more recent algorithms that have been found empirically to be significant improvements: forward checking and word-wise forward checking. We extract compatibility-based ordering-heuristics from the theory for forward checking.

A theory of analogy must describe how the meaning of an analogy is derived from the meanings of its parts. In the structure-mapping theory, the interpretation rules are characterized as implicit rules for mapping knowledge about a base domain into a target domain. Two important features of the theory are (a) the rules depend only on syntactic properties of the knowledge representation, and not on the specific content of the domains; and (b) the theoretical framework allows analogies to be distinguished cleanly from literal similarity statements, applications of abstractions, and other kinds of comparisons. Two mapping principles are described: (a) Relations between objects, rather than attributes of objects, are mapped from base to target; and (b) The particular relations mapped are determined by systematicity, as defined by the existence of higher-order relations.

Almost all game tree search procedures used in artificial intelligence are variants on minimaxing. Until recently, it was almost universally believed that searching deeper on the game tree with such procedures would in general yield a better decision. However, recent investigations have revealed the existence of many game trees and evaluation functions which are ‘pathological’ in the sense that searching deeper consistently degrades the decision. This paper extends these investigations in two ways. First, it is shown that whenever the evaluation function satisfies certain properties, pathology will occur on any game tree of high enough constant branching factor.