Source
Simon (1973), LIT-1634, §§3–5 (pp. 187–200); read in NOTE-1232.
What was actually shown
The architecture, by description. For house design, following Reitman’s analysis of a composer writing a fugue: the task evokes attributes to specify, requirements, an executive program and generators of alternatives, component by component, and “During any given short period of time, the architect will find himself working on a problem which, perhaps beginning in an ill structured state, soon converts itself through evocation from memory into a well structured problem” (p. 190). The system is “a GPS, which at any given moment finds itself working on some well structured subproblem, with a retrieval system, which continually modifies the problem space” (pp. 191–192; Fig. 1). The same form is found in Murray’s account of battleship design, where the evoking is the consulting of departments and long-term memory is “literally a distributed memory” (p. 194); in a robot, whose problem is “well structured … at each moment … but one that changes from moment to moment” (p. 195); and in a production-based chess program whose effective problem space moves through the large space of its memory (p. 196).
The argument from seriality. A serial solver can bring little information to bear over any short period, so most of a large memory is irrelevant at any moment (p. 192). Assembling everything relevant “once and for all at the outset” gains nothing: “bringing it together where?”; retrieval would still be needed, and the result “would still look like a series of transformations of the problem space” (pp. 192–193).
The thesis. “Any problem solving process will appear ill structured if the problem solver is a machine that has access to a very large long-term memory … and/or access to a very large external memory” (p. 197). “It suggests that there may be nothing other than the size of the knowledge base to distinguish ISPs from WSPs”, and that mechanisms that work on large, apparently well structured domains “should be extendable to ill structured domains without any need for introducing qualitatively new components” (p. 197). Programs of 1958–1973 already did some of this “unknowingly” (NSS chess, GPS planning; pp. 199–200).
What this does not say
- That it has been tested. No system of this design is built or run in the paper. Simon flags his inexperience with large semantic stores (p. 198) and sets aside the two hard sources of new information, natural language instructions and perception (pp. 198–199).
- That revision of a problem space creates distinctions. In the paper, what enters the space is “generated from, or dredged from” memory or a library (p. 188), so the revision it models is mostly selection within a larger fixed store; the robot’s revising its “laws of nature” (p. 195) is the exception, and is not developed. That bears on CLAIM-470’s defeater (a), and does not settle it.
- That the alternation is costless or safe. Simon’s own design account has requirements violated after being satisfied and criteria never evoked (p. 191), and an acquisition process that conserves “existing problem formulations” (p. 199).
- Anything about a parallel solver. The argument from seriality is the premise.