Philip Sargent

SERC Advanced Research Fellow
Cambridge University Engineering Department
Wednesday, November 6, 1991

This note is an attempt to respond to the often-made comment that Simon’s concept of ‘satisficing’ may imply that a whole theory of design awaits development – by analogy with the techniques of operations research and economics which are based on ‘optimizing’ principles.

I find the idea appealing, but I am not persuaded by it. Therefore I am briefly stating some comments by which I hope to tease the matter out into the open. The discussion is more mathematical than I would like.

Simon introduced the term ‘satisficing’ in 1956 and reached a wider audience with his book The Sciences of the Artificial [Sim81]. He uses it to mean several different, related things. First, that optimal solutions to models of the real world {1} will not really be optimal because they are based on imperfect models, but that the solutions may be good enough. Second, that some search techniques cannot with certainty locate the optimum even of a model problem, but that the solutions they do find may be adequate in practice {2}. Third, that qualitative reasoning is the way that people think, and therefore in practice their solutions will be satisfactory not optimal {3}.

Simon goes on to argue that satisfactory solutions are often more stable than optima in society as a whole and over time. This is naturally true if most objective functions are less steep near the optimum. However if one values ‘stability’ more highly than ‘value’, one merely derives a new objective function with the appropriate behaviour and optimizes that instead: which is what Taguchi methods for ‘robust design’ do.

Simon presents a reasoned discussion of the role of evolution in the economic world and describes why pure market theory (which assumes the existence of optima) is therefore not relevant. Local satisficing solutions are the only good descriptions. This has been much more explicitly described in recent work by Gould [Gou77]. The point is not particularly relevant to design.

Simon’s concept of ‘bounded rationality’ is closely related to that of ‘satisficing’. This too is really several closely-related ideas. First, that with imperfect information one cannot behave objectively fully rationally. Second, even with full information it may be impossible to calculate an answer because the available algorithm would take nearly forever. Third, gathering the full information before taking the decision may also take an unreasonable length of time or resources, thus the search will be terminated as soon as a satisfactory answer is found. Such satisficing decisions are thus strongly affected by the order in which alternatives are examined. (Dawes [Daw88] is clearer than Simon in discussing rationality in decision making.)

Today these matters are not seen as remarkable and we now know that a whole new mathematics is not required to deal with them. Rigourous algorithms inspired by early work in artificial intelligence (itself resulting from Newell and Simon’s work) are effective {4}. ‘Design as resource allocation’, a section of Simon’s book, is now merely accepted as necessary detail in some search algorithms - and genetic algorithms implement some of the ideas directly [DeJ89].

While the concepts of satisficing and bounded rationality were novel and intriguing in the 1960s, and while they may still give overly-traditional operations researchers (OR) pause for thought, modern computer science (including the optimization community) takes them entirely in its stride as a natural way of viewing the world [Har87]. Optimality has subsumed satisfaction. Satisfactory solutions are thus not an ‘alternative view’ to optimal solutions and there is thus no hidden ‘alternative’ design science theory by analogy with OR.

This is not a disappointing result. It means that many design problems are now properly solvable with rigourous mathematical methods, objectively, rationally and ‘scientifically’ {5}.

However there are still problems in the processes of design for which no single mathematical method can ever be appropriate because many design issues require an exploration of the trade-offs between two or more incommensurable viewpoints from different disciplines. Personally I find this represents for me the essence of design [Kon91].