Allen Newell

With Herbert Simon, Newell argued that "a physical symbol system has the necessary and sufficient means for general intelligent action" (Newell & Simon, *Computer Science as Empirical Inquiry*, 1976 ACM Turing Award lecture, *Communications of the ACM* 19(3)). Intelligence consists of manipulating symbol structures (expressions) by processes that create, modify, and destroy them. The hypothesis is *empirical*: it is to be tested by building systems and observing whether symbol manipulation produces general intelligent action. For a reviewer, PSSH is the question: *where in your space system are the symbols, what are the symbol-manipulating processes, and is the symbol grounding real or hand-waved?*

In *Human Problem Solving* (Newell & Simon, 1972) and the General Problem Solver (GPS) work (Newell, Shaw & Simon, late 1950s–60s), Newell framed all problem solving as search through a *problem space*: a set of states, a set of operators that transform states, an initial state, and a goal. Because the spaces are combinatorially enormous, intelligence is the use of *heuristics* to prune and guide search rather than enumerate it. The reviewer's question: *what is the state space, what are the operators, what is the goal test, and what heuristic keeps search tractable instead of exhaustive?*

The signature GPS heuristic: detect the *difference* between the current state and the goal state, then select an operator known to reduce that specific kind of difference; if the operator's preconditions are unmet, set up a subgoal to satisfy them, recursively. MEA is the original general-purpose planning control regime and the conceptual ancestor of modern automated planners. Reviewer's question: *does the system reason backward from goals via difference reduction, or does it merely react?*

In his 1987 William James Lectures, published as *Unified Theories of Cognition* (Harvard University Press, 1990), Newell argued against a science of isolated micro-phenomena and for a single architecture that accounts for the full span of cognition. Soar (Newell, Laird, Rosenbloom) is his candidate: a problem-space architecture in which all tasks are formulated as search, *all permanent knowledge is production rules*, *impasses* trigger automatic subgoaling, and *chunking* converts the results of subgoal problem-solving into new rules (learning). Reviewer's question: *is the architecture unified and self-consistent across perception, decision, action, and learning, or a stack of disconnected modules?*

In "The Knowledge Level" (Newell, *Artificial Intelligence* 18(1), 1982), Newell distinguished describing an agent by *what it knows and what goals it pursues* (the knowledge level, governed by the Principle of Rationality: an agent uses knowledge to select actions that attain its goals) from *how that knowledge is symbolically encoded* (the symbol level). This separation is essential for evaluating autonomy claims: a system can be analyzed for rational competence independent of its implementation. Reviewer's question: *at the knowledge level, does this system have the knowledge required to attain its goals, and can you demonstrate it, separately from the code?*

Newell repeatedly warned that "you can't play 20 questions with nature and win" - that piling up isolated empirical effects never yields a theory; you need a mechanism that generates the effects. His standard for a contribution was a *working, operational* mechanism with explanatory reach. Reviewer's question: *is this a mechanism that produces the behavior, or a catalogue of observations dressed as a theory?*