International Encyclopedia of Systems and Cybernetics

1. Any kind of conceptual device designed to think about a problem, to uncover so-called facts, to better understand phenomena and to accelerate the discovery of proofs and laws. (from the Greek "eurisko": I Found; famous through ARQUIMEDES' well known exclamation "Eurêka", I found out)

2. "Step-by-step procedures which in a finite number of steps insure that a satisfactory solution to the problem is reached" (J.van GIGCH, 1978, p.204).

van GIGCH compares algorithms and heuristics: "An algorithm guarantees that the optimum solution is reached. The heuristic leads to a satisfactory solution, not necessarily optimum. Both consist of a finite number of steps… It is important to note the difference between heuristics and rules of thumb. A rule of thumb usually has no analytical foundation and has been developed on the basis of intuition and long-time experience… Heuristics to replace them, will probably start from the rule of thumb and develop on the basis of a solid and rigorous analysis of the problem and the factors involved. Heuristics are methods "to reduce search" (p.205).

M. BODEN's opinion is different: "Most heuristics are pragmatical rules-of-thumb, not surefire methods of proof. Although there is a reasonable chance that they will help you solve your problem, they can sometimes prevent you to doing so" (1990, p.53).

Heuristics are used to explore a defined extension of possibilities within a set of generative rules, i.e. by defining a specific search space and paths of exploration.

M. BODEN gives as an example the rules of chess: "Here the search space consists of all the board states that could be reached by any series of legal moves. Each legal move involves a specific action defined by the rules of chess" (p.77).

However, M. BODEN also observes: "It is one thing to say that people use heuristics to solve problems. It is another thing to say that all the challenges facing creative thinkers can be met by applying ordered heuristics to strictly defined search trees, in the way that traditional A.I. problem-solvers do" (p.111). Heuristics could thus be limitative to very original creative thinkers, whose creativity could be in producing wider embracing heuristics.

We may be led in this way to "… the need for heuristics for changing heuristics" (p.210).

An interesting generalization of heuristics has been proposed by D.B. LENAT in his program EURISKO (or AM). According to him the same exploration process can be used at successive hierarchic levels (p.211).

Heuristics include the discovery of satisfactory rules of inference and the selection of the shortest ones. It follows: "… mainly intuitive… principles and depends on the type of theory" (after the A.I. Dictionary of the "Osterreichische Gesellschaft für Artificial Intelligence", p.15)

According to J.W. SUTHERLAND, "… many of the higher-order paradigms… serve as heuristics, a priori 'masks' we can wear to lend some tentative order to otherwise ill-structured (effectively indeterminate) phenomena. As opposed to a theory, per se, which depends for its utility on its apodictical quality (i.e., its ability to be empirically validated), the heuristic makes no pretense to rectitude or allegorical alignment with some specific real-world entity. Rather, it serves merely as a synthetic construct which is usually a combination of two distinct components:

"A metatheorical component which serves to provide an initial and quite broad boundary for analysis…

"The second component of a proper heuristic is algorithmic in nature: it really involves the specification of the 'rules' we will use in disciplining subsequent analysis, setting in the broadest way the criteria of investigation and truth" (1973, p.182-3).

In a quite different sense, statistical correlations between some variables can be used as a base to search for their significance, as shown by R. MARGALEF (1980, p.118). This could be called "statistical heuristics".

I.G. BLOOR describes the uses of heuristics in decision-making processes. He writes: "Two basic versions exist: one where the method is unknown or unstructured (trial and error) and the objective or end solution is fixed, and the other where the method is proven or known and the goal or solution is unknown… The technique may be used where source data do not easily fit into a mathematical or similar model" (1987, p.18).