Allen Newell
Allen Newell (1927–1992) was an American computer scientist and cognitive psychologist who, with Herbert Simon and the programmer Cliff Shaw, built the first programs that proved theorems and solved problems by heuristic search, and who spent the rest of his career trying to specify a complete architecture of the human mind. He shared the 1975 Turing Award with Simon for basic contributions to artificial intelligence, the psychology of human cognition and list processing, and he was Simon's closest collaborator for more than three decades.
Newell's path ran through RAND to Carnegie. After studying physics at Stanford he joined the RAND Corporation in 1950, where he worked on simulations of air-defense organizations; a lecture on early pattern recognition by Oliver Selfridge and conversations with Simon, who visited RAND, convinced him that computers could manipulate symbols rather than only numbers. He completed a doctorate at the Carnegie Institute of Technology under Simon, and later helped establish the computer science department at Carnegie Mellon University in the mid-1960s alongside Simon and Alan Perlis. At RAND he and Shaw built the list-structure programming languages called IPL, which made flexible symbolic memory practical and which strongly influenced later list processing languages.
The first line of his work, in the 1950s and 1960s, produced the Logic Theorist and the General Problem Solver, programs designed as theories of human reasoning as much as engineering contributions. Each stored its knowledge about a task separately from a general search method, and Newell and Simon compared the programs' traces with human protocols in their book Human Problem Solving (1972). Newell also contributed to other fronts: he co-authored Computer Structures with Gordon Bell (1971) on computer architecture, and he led the Carnegie Mellon work on speech understanding in the 1970s, of which the Hearsay-II system is the best-known result. In 1976 he and Simon stated the physical symbol system hypothesis in their Turing lecture.
The second line, from the early 1980s, was the quest for a unified theory, culminating in Soar and the book Unified Theories of Cognition (1990, based on his William James Lectures at Harvard). Soar, developed with John Laird and Paul Rosenbloom, casts all behavior as search within problem spaces and learns by chunking: when a subgoal is resolved, the result is stored as a new rule so that the same situation need not be searched again. In an earlier 1973 paper, You can't play 20 questions with nature and win, Newell had argued that psychology kept producing isolated binary experiments that never added up, and that the way out was to commit to a complete model. Soar was his answer, and the book argued that it should be judged against many constraints at once, including reaction times, learning and the capacity to use language. His 1982 essay The Knowledge Level separately proposed a way to describe an agent purely by what it knows and wants, apart from how that knowledge is represented.
There are common misreadings of his work. One is that Newell believed the brain is a serial symbolic computer in all respects; he assigned symbolic behavior to a particular timescale of cognition and expected a neural substrate beneath it. Another is that unified theories failed outright: Soar and its cousin ACT-R, from John Anderson, are still maintained and used, but neither became the single accepted theory he hoped for, and critics point to how many modeling choices each requires. He died in July 1992, shortly after receiving the National Medal of Science.
The legacy is evident in current practice. Cognitive architectures, the idea of chunking and caching search results, and the knowledge-level view of agents all influence planning and agent design, and the notion that capability should be a combination of general mechanism and stored knowledge reappears in modern systems that combine a model with memory, tools and learned skills. His insistence on testing a theory of mind by building it is the working method of the whole field.
Read Allen Newell as the builder who wanted AI to explain the mind: from list-processing languages and the Logic Theorist to Soar, he held that a theory of cognition is a program that runs, and that it should be one program rather than many.