This post is based on a paper: “What is adaptive about adaptive decision making? A parallel constraint satisfaction account,” that was written by Andreas Glöckner, Benjamin E. Hilbig, and Marc Jekel and appeared in Cognition 133 (2014) 641–666. The paper is quite similar to that discussed in the post Swiss Army Knife or Adaptive Tool Box. However, it reflects an updated model that they call the PCS-DM model (parallel constraint satisfaction-decision making). From what I can tell this model attempts to address past weaknesses by describing the network structure more fully and does this at least partially by setting up a one-free parameter implementation which can accommodate individual differences and differences between tasks.
The paper provides an account of experiments done to match the battling ideas of fast and frugal heuristics against connectionist models. One switches strategies and the other switches weights to be adaptive. Fast and frugal heuristics assume a repertoire of narrow strategies or tools from which the decision maker picks one that is best suited for the current environment and constraints. Thus, adaptivity lies in choosing between different strategies of information integration. The connectionist models emphasize broad and general models of cognition, typically specifying a mechanism which approximates rational solutions. In this paper, the single-mechanism model is the PCS (parallel constraint satisfaction) theory for decision making (PCS-DM). In this approach, the only aspect changing adaptively is the environmental information concerning cue validities and the sensitivity to differences concerning these cue validities (as captured by the sensitivity parameter (P), whereas the information integration mechanism in itself remains the same.
Since both single-mechanism and multi-strategies make similar behavioral predictions in many situations, Glockner et al rely on process measures (reaction times and confidence ratings–post MM-ML) as well as cross-prediction. In one experiment they additionally used eye-tracking to record information acquisition. In four experiments, they compared the predictions of the single-mechanism PCS-DM model with those from a toolbox consisting of a weighted additive strategy(WADD), an equal weights strategy(EQS) and the lexicographic take-the-best heuristic (TTB).
In all experiments, they manipulated the structure of the environment in terms of cue validities, so as to produce one with a compensatory and one with a non-compensatory structure. The multi-strategy account predicts shifts toward TTB from the compensatory to the non-compensatory environment on all dependent measures. The single-mechanism account, by contrast, predicts that behavior in either environment is best accounted for by PCS-DM and that changes in choice patterns are accompanied by very different gradual changes in process measures such as decision times or confidence patterns.
Across experiments and analyses, the results were relatively clear cut: whereas they fully replicated large changes in choice patterns, there was no evidence for switches in
the underlying information integration mechanism. Rather, the single-mechanism PCS-DM model accounted for data best across all conditions, reinforcing the idea that adaptivity lies in adapting cue weights rather than selecting different strategies. It should be noted that the PCS-DM model used was fully specified and did not comprise any free parameters in one of the implementations. This PCS-DM without fitted parameters can approximate the rational naïve Bayesian solution well. Nonetheless, letting the sensitivity parameter P – which captures subjective differences in scaling of and sensitivity toward cue validities – vary freely provided significant additional fit and further insights. Most decision makers showed a lower sensitivity to cue weights than would be necessary to approximate the Naïve Bayesian solution. Stated differently, differences between cues were only insufficiently taken into account in decision makers’ mental representations of the task. This reduced sensitivity was stronger in non-compensatory environments than in compensatory ones, and learning seemed to lead to better adjustment.
PCS-DM is based on a connectionist mechanism that provides a cognitive and evolutionarily plausible implementation for automatic processes of coherence structuring in decision making (Posts Prediction machine, Explaining away , and Prediction Error Minimization and its Implications). The core advantage of PCS-DM over previous models is that it specifies a general network structure and a flexible transformation function, thus allowing for precise predictions of choice, decision time, and confidence. At the same time, it achieves adaptivity through one free parameter that captures intra and inter-individual differences in the sensitivity to (the distribution of) cue validities. Specifically, individuals may differ in how they translate information about the world into their mental representation of the decision task.
The experiments were done for probabilistic inference tasks only, and it remains for future research to investigate whether fully specified PCS mechanisms for other domains such as risky choice or multiattribute decision making perform similarly well. Glockner et al note they they cannot rule out that multi-strategy accounts perform better in other domains. Recent eye-tracking studies, however, indicate that PCS is a promising account for describing the processes underlying risky choices and expert decisions in sports. Also, for probabilistic inferences involving recognition information, PCS has been shown to be superior to heuristics and PCS mechanisms provide one of the most established accounts for legal decision making. The fact that coherence effects – that are predicted by PCS mechanisms but cannot be explained by any heuristic specified so far – have been observed for many domains of decision making speaks for the generality of the PCS mechanism.
The results provide support for the single-mechanism view that different (subjective) weights are attached to the information under different conditions. However, according to Glockner et al, there is no implication that information search is not driven by different strategies. Whenever decision makers are faced with serial and stepwise information search, when information is costly , or when it must be effortfully retrieved from memory they consider it likely that high cue dispersion, time pressure, or other factors will lead to information search as predicted by TTB. Glockner et al suggest that future research should thus aim for a more integrative and comprehensive understanding of the interplay between information acquisition which may well conform to a multi-strategy account and information integration which, as the current findings imply, is best described by a single-mechanism view.