Heuristics For Decision And Choice Research Paper

1. Historical Overview

The term ‘heuristic’ (from a Greek root, ‘to discover’) was used for most of the twentieth century to refer to useful, even indispensable, strategies for finding solutions to problems that are difficult to approach by other means. Gestalt psychologists spoke of heuristic reasoning methods such as ‘looking around’ and ‘inspecting the problem’ to guide the search for useful information in the environment, while mathematicians employed heuristics including ‘examining special cases’, ‘exploiting related problems’, ‘decomposing and recombining’, and ‘working backwards’ (Groner 1983). These rather vague strategies were made more precise in computer-based models of human problem solving and reasoning largely based on the means-ends analysis heuristic, which sought some way to reduce the distance between the current partial-solution state and the goal state (e.g., in Newell and Simon’s early General Problem Solver system). Such general purpose or ‘weak’ methods proved insufficient to tackle many problems, so research in AI in the 1970s turned to collecting domain-specific rules of thumb from specialists in a particular field and incorporating these into expert systems. Around the same time, mathematicians working in OR faced new results from computational complexity theory indicating that efficient algorithmic solutions to many classes of challenging combinatorial problems (such as the Traveling Salesman Problem) might not be found; as a consequence, they too turned to the search for problem-specific heuristics, though through invention rather than behavioral observation (Muller-Merbach 1981).

After 1970, though, heuristics gained a different connotation in psychology: fallible cognitive shortcuts that people often use in situations where logic or probability theory should be applied instead. The ‘heuristics-and-biases’ research program launched by Tversky and Kahneman (1974, Kahneman et al. 1982) emphasized how the use of heuristics can lead to systematic errors and lapses of reasoning indicating human irrationality. The heuristics studied (see Sect. 2) were often vaguely defined and broadly applicable to judgments made under uncertainty in any domain, akin to the weak methods explored earlier in AI. This negative view of heuristics and of the people who use them as ‘cognitive misers’ employing little information or cognition to reach biased conclusions has spread to many other social sciences, including economics (Rabin 1998) and law (Hanson and Kysar 1999). More recently, a new appreciation is emerging within psychology that heuristics may be the only available approach to decision making in the many problems where optimal logical solutions are computationally intractable or do not exist (as OR researchers realized), and that domain-specific decision heuristics may be more powerful than domain-general logical approaches in other problems (as AI found). This has led to the study of precisely-specified heuristics matched to particular decision tasks (see Sect. 3), and the ways that learning and evolution can achieve this match in human behavior (e.g., Payne et al. 1993, Gigerenzer et al. 1999). The existence of such evolved adaptive heuristics has already been widely accepted for other animals in research on rules of thumb in behavioral ecology (Gigerenzer et al. 1999).

2. Heuristics For Probability Judgment

Since the rise of the heuristics-and-biases research program in the 1970s, the heuristics most widely studied within the social sciences are those that people use to make judgments or estimates of probabilities or other quantities in situations of uncertainty where they do not have complete, exact information (Tversky and Kahneman 1974). The representativeness heuristic is a means to assess the probability that an object A belongs to a particular category B (e.g., that a person described as meek is a pilot) or that an event A is generated by a particular process B (e.g., that the sequence HHTTTT was generated by throwing a fair coin). This heuristic produces probability judgments according to the extent that A is representative of, or similar to, B (e.g., meekness is not representative of pilots, so a meek person is judged as having a low probability of being a pilot). The availability heuristic can be used to produce assessments of class frequency or event probability based on how easily instances of the class or event can be mentally retrieved (e.g., plane crashes may seem like a frequent cause of death because it is easy to recall examples) or constructed (via the simulation heuristic). The anchoring-and-adjustment heuristic produces estimates of quantities by starting with a particular value (the anchor) and moving away from it (e.g., people asked to quickly estimate either 8×7×6×5×4×3×2×1 or 1×2×3×4×5×6×7×8 give a higher value in the former case, which could come from estimation based on multiplying the first few values together, creating a higher or lower anchor, and then adjusting upwards, which yields a higher overall estimate of the first product).

Most researchers agree that such heuristics are widely used because they usually lead to good choices without much time or mental effort. Despite this, most of the large body of evidence amassed that is consistent with the use of these heuristics comes from showing where they break down and lead to errors (e.g., availability judgments can be manipulated by vivid examples like plane crashes, and anchoring can lead to different choices solely due to the order of information presentation), leading many to conclude that people are poor decision makers after all. These heuristics have also been criticized as being vague redescriptions of observed choice behavior that are therefore difficult to falsify; however, more specific versions (e.g., testing whether availability works in terms of ease of recall or number of items recalled) are also being explored.

3. Heuristics For Choice Between Alternatives

Specifying decision tasks in more detail—for instance, choosing between two (or more) alternatives on the basis of cues or pieces of information known about each—allows still more specific heuristics to be investigated. The typical normative approach to such choice decisions involves weighting and combining (typically adding) all of the available cues, with the weights determined by a process such as multiple regression. Heuristics can shortcut this process in two main ways. First, rather than weighting all cues differentially, positive and negative cues for each alternative can merely be tallied (Dawes 1979). Second, not all cues need be used—instead, a stopping rule can indicate when enough cues have been considered to allow a choice (see below). Despite these simplifications, such heuristics often perform very well when compared to expert human decision makers (Dawes 1979), traditional benchmarks (Payne et al. 1993), or objective measures of correct answers (Gigerenzer et al. 1999).

There are several forms of stopping rules that can be used to limit the information considered by choice heuristics. The recognition heuristic (Gigerenzer et al. 1999) only ever considers the single cue of recognition: when one alternative is recognized and the other is not, this heuristic chooses the former (and otherwise guesses). One-reason decision mechanisms (Gigerenzer and Goldstein 1996) compare both alternatives on a single cue-dimension (or reason) at a time and stop at the first dimension that distinguishes between the alternatives (i.e., where their cue values differ, allowing a decision to be made). Accumulator or horse-race models (Busemeyer and Rapoport 1988) tally the cues in favor of each alternative until a specified difference in positive evidence is reached (e.g., until one alternative is three reasons ahead of the other). Elimination heuristics (Tversky 1972, Gigerenzer et al. 1999) begin with a set of alternatives and successively remove subsets that do not match a specified cue value until a single choice remains (e.g., to choose a restaurant, eliminate all those in town over some price threshold, then eliminate all those beyond some distance away, and so on until one restaurant is left). In every case (aside from the recognition heuristic), a search rule must also be specified to determine the order in which cue dimensions are considered; this order can (among other possibilities) be random, or based on recently used cues, or on cue validity or correlation with the choice criterion (which, when combined with a one-reason stopping rule, yields a lexicographic strategy such as the Take The Best heuristic—Gigerenzer and Goldstein 1996).

4. Heuristics For Sequential Search

When choice alternatives are encountered sequentially over time rather than simultaneously (e.g., in comparison shopping by driving from one store to the next), heuristics can be used to decide when to stop searching and select an available alternative. The traditional normative approach is to search until one finds an alternative below a precalculated reservation price balancing expected benefit of further search against its cost; this requires full knowledge of the search costs and distribution of available alternatives. Heuristics that use a simplified version of the reservation price calculation (e.g., replacing an integral with a weighted sum) come very close to normative performance (e.g., at selecting good prices, Moon and Martin 1990). Other heuristics require less knowledge, such as ‘Keep searching until the total search cost exceeds 7.5 percent of the best price found’ (Moon and Martin 1990). Herbert Simon’s bounded rationality principle of satisficing suggests setting an aspiration level equal to an alternative that is good enough (rather than optimal) and searching until that aspiration level is met; exactly how the aspiration level can be set varies with the search setting (e.g., whether it is a one-sided search like shopping, or a two-sided mutual search like finding a mate, Gigerenzer et al. 1999). Finally, another type of search heuristic that people use stops search after a particular pattern of alternatives is encountered, rather than after some threshold is exceeded (despite the fact that pattern should not matter from a normative perspective). For instance, the ‘one-bounce’ and ‘two-bounce’ rules say to keep searching for a low price until prices go up for the last or two last alternatives, respectively (Moon and Martin 1990).

5. Further Directions

Decision heuristics have been studied in different research traditions, in particular one that has focused on when and where verbally described heuristics can break down and yield deviations from classical norms of rationality (Kahneman et al. 1982), and another that has investigated how specific computationally modeled heuristics can exploit structured information to yield fast and accurate decisions (Gigerenzer et al. 1999). Further research should explore not only those aspects of decision environments that lead to different levels of performance, but also what factors lead people to use heuristics at all (e.g., time pressure, Payne et al. 1993), to choose between different possible heuristics (e.g., the influence of domain-specific emotions or social norms), and to learn or develop new heuristics (e.g., through social imitation).

Bibliography:

1. Busemeyer J R, Rapoport A 1988 Psychological models of deferred decision making. Journal of Mathematical Psychology 32: 91–134
2. Dawes R M 1979 The robust beauty of improper linear models in decision making. American Psychologist 34: 571–82
3. Gigerenzer G, Goldstein D G 1996 Reasoning the fast and frugal way: Models of bounded rationality. Psychological Review 103: 650–69
4. Gigerenzer G, Todd P M, ABC Research Group 1999 Simple Heuristics That Make Us Smart. Oxford University Press, New York
5. Groner R (ed.) 1983 Methods of Heuristics. Erlbaum, Hillsdale, NJ
6. Hanson J D, Kysar D A 1999 Taking behavioralism seriously: The problem of market manipulation. New York University Law Review 74: 630–749
7. Kahneman D, Slovic P, Tversky A 1982 Judgment under Uncertainty: Heuristics and Biases. Cambridge University Press, New York
8. Muller-Merbach H 1981 Heuristics and their design: A survey. European Journal of Operational Research 8: 1–23
9. Moon P, Martin A 1990 Better heuristics for economic search. Experimental and simulation evidence. Journal of Behavioral Decision Making 3: 175–93
10. Payne J W, Bettman J R, Johnson E J 1993 The Adaptive Decision Maker. Cambridge University Press, New York
11. Rabin M 1998 Psychology and economics. Journal of Economic Literature 36: 11–46
12. Tversky A 1972 Elimination by aspects: A theory of choice. Psychological Review 79: 281–99
13. Tversky A, Kahneman D 1974 Judgment under uncertainty: Heuristics and biases. Science 185: 1124–31