! 11 ARTIFICIAL INTELLIGENCE: AVRAND PERSPECTIVE Philip Klahr and Donald A. Waterman I. THE EARLY YEARS: THE BEGINNINGS OF ARTIFICIAL INTELLIGENCE In looking at the first published book on Al, the 1963 Computers and Thought anthology of Feigenbaum and Feldman, one realizes that Rand's contribution to the beginnings of Al was substantial. Of the twenty articles included, six had been previously published as Rand research reports [4,18,60,61,68,88]. Much of this early work in Al at Rand was due to the collaboration of two Rand employees, Allen Newell and Cliff Shaw, and a Rand consultant, Herbert Simon of Carnegie Institute of Technology (later to become Carnegie-Mellon University). Beginning in the mid 19505, Newell, Shaw and Simon's research on the Logic Theory Machine, their chess playing program, and GPS (General Problem Solver) defined much of Al-related research during the first decade of Al . Their work encompassed research areas that remain today as prominent subfields of Al : symbolic processing, heuristic search, problem solving, planning, learning, theorem proving, knowledge representation, and cognitive modeling. At Rand they left a legacy of publications that gave Al many of its building blocks and much of its momentum [51-71,79-81]. It is important to note that this surge of Al activity at Rand did not exist in isolation. It occurred at a time and place where a host of fundamental notions about computer science and technology were being generated. Rand in the 1950s was involved in designing and building one of the first stored-program computers, the JOHNNIAC 1 [27]; George Dantzig and his associates were inventing linear programming [11]; Ford and Fulkerson were developing techniques for network flow analysis [24]; Richard Bellman was developing his ideas on dynamic programming [6]; Herman Kahn was advancing techniques for Monti Carlo simulation [35]; Lloyd Shapley was revolutionizing game theory [77]; Stephen Kleene was advancing our understanding of finite automata [42] ; and Alfred Tarski was helping us define a theory of computation [86]. 2 xThe JOHNNIAC, named after John yon Neumann, a Rand consultant in the late 1940s and early 19505, was in operation from 1953-1966. It was used extensively for Newell, Shaw and Simon's work on information processing. 2 Much credit for creating this intellectually stimulating environment goes to John Davis Williams who led Rand's mathematics department in the 19505, and also to the United States Air Force for its generous sponsorship over a broad range of research activities. 12 Within this milieu, Newell, Shaw and Simon were developing methods and directions for Al research. Perhaps equally important was their development of appropriate computational tools for Al programming. Using the notion of linked- list structures to represent symbolic information, Newell and his associates developed some of the first symbol -manipulating and list-processing languages, a series of IPL (Information Processing Language) languages which culminated in IPL-V [57,72]. In their 1963 paper [9], Dan Bobrow and Bertram Raphael (both of MIT at the time, but also Rand consultants) included IPL-V as one of the four earliest and most developed list-processing languages. Because of Rand's unique computing environment and close ties with Carnegie, several Carnegie graduate students were attracted to Rand and several Ph.D. dissertations emerged, including those by Fred Tonge [88,89] and Ed Feigenbaum [17,18]. Feigenbaum, in collaboration with Simon, continued to publish Rand reports describing his experiments with his verbal learning program EPAM during the early 1960s [19-23,82]. Even after leaving Carnegie, Feigenbaum remained a Rand consultant and was most influential in Rand's research on expert systems and expert systems languages that would emerge in the early 19705. Newell and Simon also were Rand consultants during the 1960s and 19705. One of their associates, Don Waterman, joined Rand in the mid 1970s and brought much of their influence in the use of production systems to Rand's first work on expert systems . But Al also had its share of controversy, at Rand as elsewhere. Given its quick rise to popularity and its overly ambitious claims [83], Al soon had its critics, and one of the most prominent, Herbert Dreyfus published his famous critique of Al [15] while consulting at Rand. Also the early promise of automatic machine translation of text from one language to another (most notably Russian to English at Rand) produced only modest systems, and fully automated machine translation, as a goal, was abandoned in the early 19605. The research in machine translation did serve to elucidate the difficult problems that faced automated language understanding and translation. The work turned more toward fundamental and generic issues of linguistic theory. As a result, Rand was engaged in over a decade of activity in computational linguistics. By 1967, Rand researchers had produced a wealth of literature (over 140 articles) on linguistic theory and research methods, computational techniques, the English and Russian languages, automatic content analysis, information retrieval, and psycholinquistics [33] . In addition, David Hays produced one of the first textbooks on computational linguistics [32] . During the 19605, Rand provided a center in which natural-language researchers from all over the world could meet, communicate and collaborate. Special seminar programs and summer symposia (e.g., [43]) provided ample opportunity for researchers to exchange ideas and test various theories. Rand work during this period included the development of the MIND system by Martin Kay and his associates which focused on research in morphology [39], semantic networks [40,76], and parsing [36-38]; Jane Robinson was developing new syntactic analyzers [73]; Roger Levien and Bill Maron were developing the Relational Data File [47-49] for information retrieval and question-answering; Larry Kuhns was developing a sophisticated query language for data base inference [44,45]; and, in a somewhat different area, work was beginning on a new theory of "fuzzy sets" [7] . 11. HUMAN-ORIENTED ENVIRONMENTS Beginning with one of the world's first electronic digital computers, the JOHNNIAC in 1953, Rand has continually been involved in developing human-oriented interfaces. Although much of this work does not necessarily fall into the Al framework, the research has provided interactive computing environments that have made Al systems easier to design, implement, debug, and understand. Today, computational environments appropriate for Al systems represent a prominent subfield of Al research. Major milestones in Rand's work on human-oriented environments are JOSS, the Rand Tablet, Videographics, GRAIL and BIOMOD. JOSS [78] can be considered one of the world's first true interactive computing systems. JOSS was executed interpretively, had execution tracing facilities, could be used to solve mathematical problems with considerable precision, and had a clean, easy-to-learn syntax. These characteristics were to influence a number of future programming systems and environments . JOSS has remained a key computing resource at Rand for over 20 years, and is used to this day. The Rand Tablet [12] was the first example of a two-dimensional writing surface for computer communication. Through a capacitative coupling between a stylus and a grid of wires embedded in the tablet, sufficient accuracy was obtained to allow recognition of hand-printed characters of approximately 1/4-inch high. Entire interactive computing systems (e.g., GRAIL and BI0M0D) were based on use of the Rand Tablet -- indeed, the entire GRAIL system was programmed using the Tablet as the sole input device. The Rand Videographics System [90] was one of the first interactive graphics systems using raster-scan technology. High-resolution video displays were driven from a rotating magnetic disk. This rapid-response system allowed many Rand researchers to have simultaneous access to precision computer graphics, at a time when the cost of providing individual self-contained systems would have been prohibitive. 13 14 The GRAIL system [16] and its successor BIOMOD [26] pioneered the use of graphic displays and tablets as I/O devices for programming and modeling. GRAIL allowed the direct execution of programs defined by a combination of flowcharts and coding forms, in which each box on a top- level flowchart could itself be defined in terms of another flowchart, or (at the lowest level) a coding form. BIOMOD applied these interactive concepts to a major biomedical modeling system, in which transfer equations and other forms of equations describing a dynamic system could be input directly on a displayed form. The results of a simulation defined by these equations could then be viewed graphically in terms of time history plots of any of the variables (or combinations of them). Over 15 years ago, these systems were using a sophisticated real-time recognition algorithm for characters hand-printed on a tablet [25]. In addition to these prototype/demonstration systems, a number of studies were also conducted, principally by Barry Boehm, on the changes in the productivity of users depending on the speed and dependability of the response time of interactive computer systems [75]. In 1972, several of the personnel involved in Rand's man-machine interface work, including Keith Uncapher, Tom Ellis, Bob Anderson, and Bob Balzer (who was just beginning to develop some of his ideas on program specification [s]) left Rand to form University of Southern California's Information Sciences Institute. Anderson returned to Rand the following year and started Rand's work in an area that was to become known as expert systems . Rand work on interactive systems in the late 1970s centered on developing UNIX-based tools, such as The Rand Editor [8] and the MH electronic message handler [10], and on research in interactive maps for dynamically displaying geographic information [3]. Map displays continued to play a prominent role in Rand research in the 1980s, particularly in the color graphic displays used to dynamically depict geographic-oriented simulations.