Elsevier

Cognition

Volume 157, December 2016, Pages 77-99
Cognition

Original Articles
Strategies for memory-based decision making: Modeling behavioral and neural signatures within a cognitive architecture

https://doi.org/10.1016/j.cognition.2016.08.011Get rights and content

Highlights

  • We investigated strategies for memory-based inferences about real-world objects.

  • The strategies differ in how they use recognition and additional knowledge.

  • We implemented the strategies as computational models in the ACT-R architecture.

  • The models were tested on predictions for response times and neural activation.

  • A strategy that processes recognition and knowledge sequentially performed best.

Abstract

How do people use memories to make inferences about real-world objects? We tested three strategies based on predicted patterns of response times and blood-oxygen-level-dependent (BOLD) responses: one strategy that relies solely on recognition memory, a second that retrieves additional knowledge, and a third, lexicographic (i.e., sequential) strategy, that considers knowledge conditionally on the evidence obtained from recognition memory. We implemented the strategies as computational models within the Adaptive Control of Thought-Rational (ACT-R) cognitive architecture, which allowed us to derive behavioral and neural predictions that we then compared to the results of a functional magnetic resonance imaging (fMRI) study in which participants inferred which of two cities is larger. Overall, versions of the lexicographic strategy, according to which knowledge about many but not all alternatives is searched, provided the best account of the joint patterns of response times and BOLD responses. These results provide insights into the interplay between recognition and additional knowledge in memory, hinting at an adaptive use of these two sources of information in decision making. The results highlight the usefulness of implementing models of decision making within a cognitive architecture to derive predictions on the behavioral and neural level.

Introduction

When judging which alternative scores higher on some feature (i.e., the criterion of interest), people often search for useful information (i.e., cues) in their memories (e.g., Gigerenzer, Hoffrage, & Kleinbölting, 1991). For instance, how do people choose chocolate candies from a wide selection as a gift for a friend? To choose the ones the friend will enjoy most (criterion of interest) people might select candies based solely on whether they recognize their brand name, or they might delve deeper into their memories and retrieve additional information (e.g., how much they themselves enjoyed the candies, whether the company was involved in a recent scandal, etc.).

There has been considerable debate as to which strategies people use when making such decisions from memory (e.g., Hilbig et al., 2011, Marewski et al., 2010, Newell and Shanks, 2004, Oppenheimer, 2003, Pachur et al., 2008, Schwikert and Curran, 2014). One prominent proposal is that the recognition process for decision alternatives provides an important source of information (as illustrated in the example above). We refer to strategies that solely use recognition to make decisions as recognition-based strategies (e.g., Goldstein and Gigerenzer, 2002, Hertwig et al., 2008, Schooler and Hertwig, 2005; for other interpretations see, e.g., Tversky and Kahneman, 1973, Whittlesea, 1993). Another potential source of information is knowledge about the decision alternatives stored in long-term memory. We refer to strategies that search for knowledge beyond recognition as knowledge-based strategies (e.g., Gigerenzer & Goldstein, 1996). Finally, people might use both recognition and knowledge as cues, such that they first consider recognition, and if that does not seem to provide a reliable basis for the decision, they move on to knowledge (cf. Erdfelder et al., 2011, Marewski et al., 2010, Schwikert and Curran, 2014).

Our goal in this article was to test these different decision mechanisms by implementing them as computational models within one common framework—the ACT-R (Adaptive Control of Thought–Rational) cognitive architecture (Anderson, 2007)—and comparing their outcomes to empirical response times and functional magnetic resonance imaging (fMRI) data. We acknowledge that there are alternative frameworks for modeling decisions and response times (e.g., evidence accumulation: Lee and Cummins, 2004, Pleskac and Busemeyer, 2010; and connectionist approaches: Glöckner, Hilbig, & Jeckel, 2014), but rely on ACT-R here because it also allows one to derive specific predictions for fMRI activation.

To achieve this goal, we collected behavioral and fMRI data in an inference task in which people judged which of two cities was larger. The city domain is convenient because people are likely to have naturally acquired recognition and knowledge about cities, and both types of information can be good indicators of the decision criterion (e.g., Pachur, Todd, Gigerenzer, Schooler, & Goldstein, 2011); the domain therefore provides a good test bed for investigating how they are used to make inferences in a real-world domain. In addition, using this domain allows for comparability with the many previous studies on memory-based decisions that have also used the city-size task (e.g., Goldstein and Gigerenzer, 2002, Hilbig et al., 2011, Horn et al., in press, Marewski and Schooler, 2011, Pachur et al., 2009, Rosburg et al., 2011, Schwikert and Curran, 2014, Volz et al., 2006). We implemented the decision strategies as computational models within the ACT-R cognitive architecture (Anderson, 2007), which yields predictions at both the behavioral and the neural level for each of the candidate strategies.1

Recognition-based strategies exploit the process of recognizing the decision alternatives to make an inference. Two well-studied instances of recognition-based strategies are applicable in different situations. The first is the recognition heuristic (Goldstein and Gigerenzer, 1999, Goldstein and Gigerenzer, 2002; for reviews see Gigerenzer and Goldstein, 2011, Pachur et al., 2011).2 The recognition heuristic is assumed to be an adaptive cognitive tool that is applied when there is a (strong) positive relationship between recognition and the criterion (e.g., the number of inhabitants of a city in the present study) in the environment. It then predicts that if one of two alternatives (i.e., cities in the present study) is recognized, but not the other, then it is inferred that the recognized alternative has a higher value on the criterion.3 According to the recognition heuristic, recognition is used in a noncompensatory way, which means that recognition cannot be overruled by any other, possibly contradictory cue knowledge.

Second, the fluency heuristic (e.g., Hertwig et al., 2008, Schooler and Hertwig, 2005) predicts that if both alternatives (i.e., cities) are recognized, but one was recognized faster (i.e., more fluently), then it is inferred that the more fluently recognized alternative has a higher value on the criterion. The fluency heuristic exploits the perceived times for successful retrievals of city names from long-term memory. Previous research has shown that exploiting recognition and recognition times can lead to accurate inferences in many real-world domains (e.g., geography, sports, politics, economics; Herzog and Hertwig, 2013, Pachur et al., 2011), among them the city domain (Goldstein & Gigerenzer, 1999). Based on these accounts, we tested a recognition-based strategy that attempts to recognize both alternatives. When only one of the alternatives is recognized, it chooses the recognized one and when both alternatives are recognized, it chooses the one that is recognized faster.

After attempting to recognize the alternatives, knowledge-based strategies retrieve additional knowledge to make a decision. Such strategies can rely on the most important piece of knowledge (e.g., Gigerenzer & Goldstein, 1996) or integrate the available and relevant pieces (e.g., Dawes, 1979, Payne et al., 1993). In the city example, people might retrieve the information that a city has an international airport, significant industry, or a university (cf. Pachur et al., 2008). In the current study we made no specific assumptions about the algorithm by which such individual pieces of knowledge are processed or integrated but instead focused on the general question of whether any additional knowledge beyond recognition was retrieved before the decision. Therefore, we tested a strategy that searches for knowledge about each recognized alternative before making a decision.

A third strategy we tested allows for a lexicographic (i.e., sequential) consideration of recognition (R) and additional knowledge (K). We therefore refer to this strategy as Lex-R-K. Some studies have obtained only limited support for a strict use of the recognition and fluency heuristics (e.g., Hilbig et al., 2010, Hilbig et al., 2011, Schwikert and Curran, 2014) and purely knowledge-based strategies (e.g., Marewski et al., 2010, Pachur and Biele, 2007), so strategies implementing a confluence of different types of memory information might be more appropriate. Lex-R-K draws on proposals that the recognition cue might be evaluated in terms of memory strength (indicating certainty of recognition memory) before it is used as a basis for a decision (Erdfelder et al., 2011, Marewski et al., 2010, Schwikert and Curran, 2014).

We extended these ideas into a mechanism that evaluates whether the recognition information—for one as well as two recognized alternatives—is sufficiently reliable to be used as a basis for a decision; if not, it moves on to additional knowledge (for alternative proposals see Marewski & Mehlhorn, 2011). As a proxy for the reliability of recognition information, the strategy evaluates the speed with which decision alternatives are recognized, which is an indicator of memory strength (for evidence, see, e.g., Pleskac and Busemeyer, 2010, Van Zandt, 2000).

When only one alternative (i.e., city) is recognized, Lex-R-K checks if this alternative was recognized sufficiently quickly—that is, if the recognition time is equal to or falls below a certain threshold (see also Erdfelder et al., 2011)—indicating that recognition provides a reliable cue for the decision. When both alternatives are recognized, the strategy evaluates if one was recognized sufficiently faster than the other—that is, if the difference in their recognition time exceeds a particular threshold (cf. Luan, Schooler, & Gigerenzer, 2014). If one of these cases applies, then recognition information is used to make a decision. Otherwise, a search for additional knowledge is initiated. The existence or nonexistence of retrievable knowledge about an alternative can then serve as a basis for the decision. Thus, Lex-R-K depends on the setting of two thresholds: if only one alternative was recognized, an absolute threshold to determine whether the recognition time was sufficiently fast, and if both alternatives were recognized, a difference threshold to determine whether the recognition time difference was sufficiently large. We tested three versions of Lex-R-K, which assume different values for these thresholds such that Lex-R-K searches for additional knowledge in a small, intermediate, or large number of cases.

To test the strategies, we implemented them as computational models within the ACT-R cognitive architecture (Anderson, 2007), which allowed us to derive predictions for response times and blood-oxygen-level-dependent (BOLD) activation differences between trial types (e.g., whether both, one, or neither city was recognized, and whether additional cue knowledge was available). ACT-R is an integrated theory of cognition and a modeling environment that has been applied to a wide range of human behavior such as perception, learning and memory, language processing, problem solving, decision making, cognitive development, and the design of intelligent tutoring systems (for a list of publications see http://act-r.psy.cmu.edu/publication/). In this architecture each ACT-R model consists of a set of production rules and declarative knowledge that interact with core modules representing cognitive functions, such as vision, long-term memory retrieval, working memory updating, and motor responses.

The activity of ACT-R’s core modules has been mapped to different regions of interest (ROIs) in the brain (Anderson, 2007, Anderson, Fincham, et al., 2008; for a meta-analysis of five different tasks and models, see Borst, Nijboer, Taatgen, van Rijn, & Anderson, 2015). It has been proposed that when a module of the architecture is active, the corresponding brain region shows an increase in BOLD signal. The claim is not that the ROIs are the only parts of the brain that are activated when the modules are operating, but that these regions are good indicators of the modules’ activity. The increase in BOLD activation in an ROI can be calculated by convolving the demand function that indicates when a module is active with a hemodynamic response function (see for details, e.g., Borst & Anderson, 2015). Thus, the predefined mapping of ACT-R modules onto brain regions allows researchers to derive quantitative predictions for the BOLD response in these ROIs, and consequently to constrain models based on fMRI experiments—as we did in the current study.

The models of the three strategies result in different BOLD predictions for two ACT-R modules. First, the activity of the retrieval module reflects the retrieval of information from declarative memory. It was mapped to a region around the inferior frontal gyrus in the prefrontal cortex (PFC; left panel of Fig. 1). Previous studies have found that activation of this region reflects both the amount of retrieved information and the difficulty of retrieving episodic memories (e.g., Anderson, Byrne, et al., 2008, Borst and Anderson, 2013, Danker et al., 2008, Sohn et al., 2003, Sohn et al., 2005). Such factors also impact the retrieval of real-world memories for cities and knowledge facts in our study. Second, activity of a module critical to the functioning of working memory (i.e., ACT-R’s imaginal module, which maintains the problem state of task-relevant information) was mapped to a region close to the intraparietal sulcus in the posterior parietal cortex (PPC; right panel of Fig. 1). This region has been shown to be sensitive to updates to problem representations in working memory (e.g., Anderson et al., 2005, Anderson, Byrne, et al., 2008, Borst and Anderson, 2013, Borst et al., 2010, Borst et al., 2011, Danker et al., 2008, Sohn et al., 2005). In our study, the problem representation contained information about the decision alternatives. This information was updated by retrieving information from declarative memory. For the purpose of our analysis, we refer to these brain areas as the retrieval ROI and the working memory ROI.

Using ACT-R models of the recognition-based strategy, the knowledge-based strategy, and three versions of Lex-R-K that search for knowledge in a small, intermediate, or large number of cases, we derived quantitative predictions for the behavioral and neural data that we collected. Qualitatively speaking, compared to the recognition-based strategy, the knowledge-based strategy involves (a) more memory retrievals, as it attempts to retrieve knowledge beyond recognition (which increases response time and activity in the retrieval ROI in the PFC); (b) more working memory updates, as it needs to store the retrieved information to have it available for the decision (which increases response time and activity in the working memory ROI in the PPC); and (c) more time to control and coordinate these actions (increasing response time). Depending on whether or not Lex-R-K moves on to consider additional knowledge in a given trial, it makes the same predictions as the knowledge-based or the recognition-based strategy, respectively. Therefore, on the aggregate level predictions of the versions of Lex-R-K fall between those of the recognition-based and knowledge-based models; where exactly they fall depends on the amount of knowledge that is searched and retrieved.

We used empirical response times and fMRI data to evaluate our computational models based on their differential predictions (for more information on this approach see, e.g., Borst et al., 2015, Turner et al., in press). Beyond the response time and BOLD predictions, all models in principle could also be used to derive predictions for choices. However, in the empirical study no information was collected about the cue knowledge that participants might have used for each specific decision, which would have been necessary to predict which alternative is chosen based on knowledge. We nevertheless analyzed the choice data, making the simplifying assumption that models that base their decision on knowledge would chose the alternative for which knowledge was available or—if there was knowledge about both alternatives—would choose the alternative for which the knowledge was retrieved faster.4 As this assumption may often be incorrect, in our primary analyses we focused on response times and BOLD responses and used the choice data only to supplement these analyses.

We next report the empirical study in which participants performed the inference task that asked which of two cities had more inhabitants (the main task) and then a metacognitive judgment task (which is reported separately; Battal, Fechner, Schooler, & Volz, 2012) while positioned in an MRI scanner. Participants then worked outside the scanner on tasks that assessed their recognition of cities and the availability of knowledge about these cities. After reporting the empirical study, we describe the ACT-R models of the three tested strategies in detail and show how we derived predictions for response times and fMRI data. Finally, we compare and evaluate how well these predictions accounted for the observed quantitative and qualitative data patterns of the empirical study.

Section snippets

Participants

Twenty-seven healthy, right-handed volunteers participated in the study (15 women, 12 men; Mage = 26.43 years, SD = 3.98; range 21–37 years). They were recruited via university-wide e-mails at the University of Tübingen. Informed consent was obtained according to the Declaration of Helsinki. The local ethics committee of the University of Tübingen approved the experimental procedure. Participants were paid 12 euros per hour. The duration of the fMRI experiment was about 1 h, and the subsequent

Computational models of the decision strategies in ACT-R

All decision strategies were implemented as computational models in the ACT-R modeling framework (ACT-R 6.0, version 1.5 [r1577]), which consist of chains of production rules that reside in procedural memory (for other examples see, e.g., Marewski & Mehlhorn, 2011; see also Marewski and Schooler, 2011, Schooler and Hertwig, 2005). To simulate cognition, these production rules interact with different cognitive resources (i.e., ACT-R’s modules), such as for vision, retrieval from declarative

Results

On the basis of participants’ responses in the recognition and knowledge tasks, we distinguish between six experimental conditions in the inference task (see Table 1). The conditions differ in terms of whether neither (UU), one (RU), or both (RR) cities in a pair were recognized and whether additional knowledge was present for neither (00), one (10), or both (11) recognized cities. A total of 134 of the 4185 trials were excluded from the analysis: 5 trials because of technical failure in the

General discussion

We investigated how people make decisions based on their real-world memories, using a task in which people judged which of two cities has more inhabitants. We tested a recognition-based strategy, a knowledge-based strategy, and the strategy Lex-R-K, which sequentially considers recognition and—when recognition is deemed an unreliable basis for a decision—moves on to consider additional knowledge. We developed computational models for these strategies in ACT-R that took into account the

Conclusion

We demonstrated how computational models can be used to formalize decision strategies and tested them based on behavioral and neuroimaging data. These computational models incorporate and specify the cognitive processes underlying decision strategies, including the dynamics of memory retrieval and the perception of retrieval times. These models made qualitative and quantitative predictions for behavioral and fMRI data that we used to evaluate the candidate strategies. The results contribute to

Acknowledgments

This work was supported by a grant from the German Academic Exchange Service (DAAD), the Center for Adaptive Behavior and Cognition (ABC) and the Center for Adaptive Rationality (ARC) at the Max Planck Institute for Human Development, the International Max Planck Research Network on Aging (MaxNetAging), and the Werner Reichardt Centre for Integrative Neuroscience (CIN) at the Eberhard Karls University of Tübingen. The CIN is an Excellence Cluster funded by the Deutsche Forschungsgemeinschaft

References (71)

  • J.R. Anderson et al.

    Role of prefrontal and parietal cortices in associative learning

    Cerebral Cortex

    (2008)
  • J.R. Anderson et al.

    Practice and retention: A unifying analysis

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (1999)
  • J.R. Anderson et al.

    Reflections of the environment in memory

    Psychological Science

    (1991)
  • C. Battal et al.

    I guess I know the answer: Neural correlates of feeling of knowing in memory-based inference [abstract]

  • J.P. Borst et al.

    Using model-based functional MRI to locate working memory updates and declarative memory retrievals in the fronto-parietal network

    Proceedings of the National Academy of Sciences of the United States of America

    (2013)
  • J. Borst et al.

    Using the ACT-R cognitive architecture in combination with fMRI data

  • J.P. Borst et al.

    Using data-driven model-brain mappings to constrain formal models of cognition

    PLoS One

    (2015)
  • J.P. Borst et al.

    The neural correlates of problem states: Testing fMRI predictions of a computational model of multitasking

    PLoS One

    (2010)
  • M. Brett et al.

    Region of interest analysis using an SPM toolbox [abstract]

  • A. Bröder et al.

    Recognition ROCs are curvilinear—Or are they? On premature arguments against the two-high-threshold model of recognition

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (2009)
  • J.F. Danker et al.

    A rational account of memory predicts left prefrontal activation during controlled retrieval

    Cerebral Cortex

    (2008)
  • R. Dawes

    The robust beauty of improper linear models in decision making

    American Psychologist

    (1979)
  • M.R. Dougherty et al.

    Psychological plausibility of the theory of probabilistic mental models and the fast and frugal heuristics

    Psychological Review

    (2008)
  • E. Erdfelder et al.

    Threshold models of recognition and the recognition heuristic

    Judgment and Decision Making

    (2011)
  • A. Field et al.

    Discovering statistics using R

    (2012)
  • K.J. Friston et al.

    Statistical parametric mapping. The analysis of functional brain images

    (2007)
  • C.A. Frosch et al.

    Time to decide? Simplicity and congruity in comparative judgment

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (2015)
  • G. Gigerenzer et al.

    Reasoning the fast and frugal way: Models of bounded rationality

    Psychological Review

    (1996)
  • G. Gigerenzer et al.

    The recognition heuristic: A decade of research

    Judgment and Decision Making

    (2011)
  • G. Gigerenzer et al.

    Probabilistic mental models: A Brunswikian theory of confidence

    Psychological Review

    (1991)
  • D.G. Goldstein et al.

    Models of ecological rationality: The recognition heuristic

    Psychological Review

    (2002)
  • D.G. Goldstein et al.

    The recognition heuristic: How ignorance makes us smart

  • R. Hertwig et al.

    Fluency heuristic: A model of how the mind exploits a by-product of information retrieval

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (2008)
  • S. Herzog et al.

    The ecological validity of fluency

  • B.E. Hilbig et al.

    One-reason decision making unveiled: A measurement model of the recognition heuristic

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (2010)
  • Cited by (21)

    • Taste perceptual circuit for the generation of autonomous responses in virtual creatures

      2022, Cognitive Systems Research
      Citation Excerpt :

      So that, they send gustatory information during intake and post-intake. It is the first synapse in the brain of the virtual creature, it serves as an entry port of the gustatory signals coming from the cranial nerves C, IX, and VII, which are sent to the VPmpc (Fechner et al., 2016). These signals can come from the oral cavity as well as from the digestive tract (Kennedy & Patterson, 2012).

    • The influence of visual attention on memory-based preferential choice

      2021, Cognition
      Citation Excerpt :

      As in this example, many of people's every-day decisions rely critically on episodic memory. Accordingly, there is a growing body of research on decisions from memory (Bordalo, Gennaioli, & Shleifer, 2020; Fechner et al., 2016; Gluth, Sommer, Rieskamp, & Büchel, 2015; Hoffmann, von Helversen, & Rieskamp, 2014; Sali, Anderson, & Courtney, 2016; Shadlen & Shohamy, 2016; Weilbächer & Gluth, 2017; Wimmer & Büchel, 2016). In previous studies (Gluth et al., 2015; Kraemer, Fontanesi, Spektor, & Gluth, 2020; Mechera-Ostrovsky & Gluth, 2018; Weilbächer, Kraemer, & Gluth, 2020), we have investigated the role of memory in preferential choice by asking participants to recall pairs of options and to choose between these memorized options (subsequently referred to as remember-and-decide task).

    • Memory-related cognitive load effects in an interrupted learning task: A model-based explanation

      2020, Trends in Neuroscience and Education
      Citation Excerpt :

      But what cognitive mechanisms are responsible for these changes and how do task-inherent and situational features, such as interruptions, affect the underlying dynamics? To investigate these issues, we present a cognitive modeling approach that leverages the cognitive architecture ACT-R [1] to model learner behavior in an interrupted learning task (cf. [2]). The underlying task setting employed a controlled stimulus-response learning paradigm to put the interplay of load-inducing factors under an experimental microscope.

    • Dynamic cognitive models of intertemporal choice

      2018, Cognitive Psychology
      Citation Excerpt :

      Note that although the LGS models assume evidence accumulation within each attribute, in contrast to DFT models they do not assume evidence accumulation across attributes. The major advantage of such models is that one can derive a predicted response time distribution for them, whereas most previous research on lexicographic processing has either ignored response time or derived only qualitative predictions (Bergert & Nosofsky, 2007; Bröder & Gaissmaier, 2007; but see Fechner et al., 2016; Glöckner, Hilbig, & Jekel, 2014). Table 3 summarizes the main parameters of the LGS models.

    • Cognitive costs of decision-making strategies: A resource demand decomposition analysis with a cognitive architecture

      2018, Cognition
      Citation Excerpt :

      We thus focused on what has been called the “local frugality” of strategies, that is, the number of attributes needed for individual decisions (in contrast to “global frugality”, which refers to the total number of attributes needed to resolve all pairwise comparisons in a reference class of objects; see Dougherty & Thomas, 2012, for details). For the simulations, we implemented the decision strategies and concurrent tone counting within the ACT-R cognitive architecture (ACT-R 6.0, version 1.5 [r1577]) as chains of production rules in procedural memory (for similar examples, see, e.g., Borst, Taatgen, & van Rijn, 2010; Fechner et al., 2016; Marewski & Mehlhorn, 2011; Mehlhorn, Taatgen, Lebiere, & Krems, 2011; Nellen, 2003; see also Marewski & Schooler, 2011; Schooler & Hertwig, 2005). These production rules interact with ACT-R’s modules that represent cognitive resources (see Fig. 1), such as vision, audition, working memory updating (i.e., updating the problem state of task-relevant information, held and manipulated by what is referred to as the imaginal module), retrieval from declarative memory, motor responses, and the coordination of mental actions.

    View all citing articles on Scopus
    View full text