Algorithm, heuristic or exemplar: Processes and representation in multiple-cue judgment

Abstract

Algorithm, Heuristic or Exemplar: Process and Representation in Multiple-Cue Judgment Sari Jones (Sari.Jones@psyk.uu.se) Department of Psychology, Uppsala University Box 1225, SE-751 42, Uppsala, Sweden Peter Juslin (Peter.Juslin@psy.umu.se) Department of Psychology, Umea University SE-901 87, Umea, Sweden Henrik Olsson (Henrik.Olsson@psy.umu.se) Department of Psychology, Umea University SE-901 87, Umea, Sweden Anders Winman (Anders.Winman@psyk.uu.se) Department of Psychology, Uppsala University Box 1225, SE-751 42, Uppsala, Sweden Abstract We present an experimental design that allows us to investi- gate the representations and processes used in human multi- ple-cue judgment. We compare three ideal models of how knowledge is stored and applied in a judgment: A linear addi- tive model (LAM), a heuristic model, Take-the-best (TTB) and a generic exemplar-based model (EBM). The results show that people adaptively change processing depending on what information is present in the learning phase and whether or not the learning situation is compatible with the test. Feed- back on a continuous variable provides information sufficient to estimate a LAM that can be used both when learning is and is not compatible with the test. When only dichotomous feed- back is provided, the processes differ depending on the learning-test compatibility. At high compatibility, the proc- essing is best described by EBM, but at low compatibility heuristic processes such as TTB become more frequent alter- natives to LAM. Introduction In the 1950’s and 60’s two new research paradigms emerged in cognitive science, categorization research (e.g., Shepard, Hovland, & Jenkins, 1961) and research on multiple-cue judgment (e.g., Hammond, 1955). While the former has continued to flourish, the Brunswikian inspired judgment research quietly left the arena in the 80’s, although with a re-emergence in studies on realism of confidence (Gigeren- zer, Kleinbolting, & Hoffrage, 1991; Juslin, 1994). The two paradigms have a lot in common, but there is seldom cross- reference between them (but see Kruschke & Johansen, 1999). A major conclusion from research on multiple-cue judgment is that linear models fit judgment data well (Brehmer, 1994) but in regard to knowledge representation and processes, there has been little research. In the categori- zation literature in contrast, a variety of models with explicit representational and process assumptions have been pro- posed (see Medin, 1989). In this article, we bring the two paradigms together by combining multiple-cue learning with theories and methods from research on categorization. We present an experimen- tal design that allows us to investigate knowledge represen- tations and processes in human judgment. As a point of departure we take research that postulates multiple levels of representation (e.g., exemplars, rules) that compete to con- trol the judgments in a specific task (Ashby et al., 1998; Erickson & Kruschke, 1998; Logan, 1988). The idea is that experience with some domain may lead to co-existing repre- sentations at several levels. A general hypothesis is that the process and representation that dominates at the time of judgment is contingent on an interaction between the learn- ing environment and the judgment task to which the knowl- edge is later applied. We offer some preliminary ideas in regard to the principles that determine this interaction. The crucial questions is: In what circumstances will a particular level of representation dominate the judgments? We compare three ideal models of how knowledge is stored and applied in a multiple-cue judgment task. Linear Additive Models (LAM) suggest that we store information in memory about: (a) the weight or cue validity attached to each cue in the form of a linear coefficient, and (b) an alge- braic rule for the combination of the cues, in this case a linear additive rule (Brehmer, 1994). The process at the time of judgment is cue-integration. Recently, a simpler and more heuristic alternative has been proposed in terms of the Take The Best algorithm (TTB; Gigerenzer & Goldstein, 1996; Gigerenzer, Todd, & ABC Group, 1999). TTB sug- gests that the single most valid cue that is applicable is used and that no information is integrated. The knowledge in memory are cue validities and the process amounts to cue- substitution. Finally, Exemplar-Based Models (EBM) from the categorization literature (e.g., Medin & Schaffer, 1978; Nosofsky & Palmeri, 1997) assert that the memory traces of each encountered object are stored in memory and that judgments are based on the similarity between the new ob- ject and the already stored exemplars. In this case, the repre- sentations are exemplars and the process is similarity-based retrieval from memory.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 11 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.