Putting the psychology back into psychological models: Mechanistic versus rational approaches

Abstract

Two basic approaches to explaining the nature of the mind are the rational and the mechanistic approaches. Rational analyses attempt to characterize the environment and the behavioral outcomes that humans seek to optimize. The rational approach holds that people are adaptive and learn (at the individual or species level) to b behave optimally given the nature of the environment (i.e., given available information or statistics). The formal p product of a rational analysis is an abstract mathematical model (often Bayesian) that details the behavioral strategies that optimize some cost function, given the environment. Such models do not have recourse to how people actually process and represent information but are, instead, abstract.Considerations of the environment and optimality also resonate with adherents of the mechanistic program, but unlike for a rational model, the main goal of a mechanistic model is to simulate human behavior by using mechanisms (i.e., analogous processes and representations) that are the same as those that support human behavior. The mechanistic program seeks to reverse engineer the human b brain and peer inside the black box. The issues of primary importance to the mechanistic program are how people p p represent and process information.One common criticism of mechanistic approaches is that they lead to ad hoc explanations that lack the elegance and clarity of models derived from rational analysis. To the extent that two models converge on a common set y of predictions, the more transparent and mathematically motivated model should be favored. Echoing these sentiments, Anderson (1991b) stated, "All mechanistic proposals which implement the same rational prescription are the same," and "a rational theory provides a precise characterization and justification of the behavior the mechanistic theory should achieve." These views are seconded by Chater and Oaksford (1999): "The picture that emerges from this focus on mechanistic explanation is of the cognitive system as an assortment of apparently arbitrary mechanisms, subject to equally capricious limitations, with no apparent rationale or purpose."The upshot of these statements is that mechanisms are subservient to rational accounts of thought. Perhaps in a d moment of candor or euphoria, Anderson (1991b) stated that rational models render mechanistic models unnecessary: "One might take the view (and I have so argued in overenthusiastic moments; Anderson, in press) that we do not need a mechanistic theory, that a rational theory of-f f pp p p y fers a more appropriate explanatory level for behavioral Two basic approaches to explaining the nature of the mind are the rational and the mechanistic approaches. Rational analyses attempt to characterize the environment and the behavioral outcomes that humans seek to optimize, whereas mechanistic models attempt to simulate human behavior using processes and representations analogous to those used by humans. We compared these approaches with regard to their accounts of how humans learn the variability...

Further attempts to clarify the importance of category variability for categorisation

“…Stewart and Chater (2002; Experiment 1) observed higher classification of an NN-halfway test instance into the more variable category if participants were told, during training, that one of the two categories was more variable. Finally, Sakamoto et al (2008) found that a random presentation of the same stimuli in a category, made the category look more variable and so favored it regarding the classification of an NN-halfway critical instance.…”

“…Regarding its other details, this experiment examined a P-halfway critical instance, rather than the nearest exemplars (Cohen et al, 2001;Stewart & Chater, 2002, Experiment 1;Sakamoto et al, 2008). Also, the category boundaries for the low and high variability categories, regardless of how the test instance was classified, were always 18 simple linear ones (contrast with Cohen et al, 2001).…”

“…Thus, we think that the literature on category variability could be usefully extended by a focused study involving Phalfway instances and this is what we aim to achieve with the present research (taking into account a range of relevant methodological manipulations). Notably, some work indicates that the default classification bias is to place an NN-halfway exemplar into the high variability category (Cohen et al, 2001;Sakamoto et al, 2008), others that if category variability is not made salient it will be classified in the low variability category (Stewart & Chater, 2002). There are different ways in which variability can be made salient and this issue will be considered in our experiments.…”

“…A related study is that of Sakamoto, Jones, and Love (2008), where participants saw test items which were NN-halfway; again, these test items were more likely to be classified into the more variable category. In Sakamoto et al's study the stimuli were lines, in a way analogous to that of Cohen et al (2002), and the training procedure was that of a standard supervised categorization experiment (participants went through several stimulus presentation (in their Experiment 1), the probability of classification into the high variability category (without a corresponding hint) was only .25 (it was .37 with the hint).…”

Constraining bridges between levels of analysis: A computational justification for locally Bayesian learning

“…Computational-level models that ignore the process can struggle with making fine-grained predictions (Sakamoto et al, 2008) and algorithmic-level mod-2 els that ignore the computational level risk making incorrect or no predictions for task variants (Griffiths & Tenenbaum, 2009;Sanborn et al, 2013). A classic way to combine computational-and algorithmic-level insights is to begin with an algorithmic-level model developed to fit human behavior and then investigate its computational-level properties (Ashby & Alfonso-Reese, 1995;Gigerenzer & Todd, 1999).…”